diff --git a/.changeset/glob-pagination.md b/.changeset/glob-pagination.md
new file mode 100644
index 00000000000..d67f3026a2e
--- /dev/null
+++ b/.changeset/glob-pagination.md
@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Allow file searches to retrieve matches beyond the first 100 results.
diff --git a/apps/kimi-code/src/tui/components/messages/tool-renderers/grep-output.ts b/apps/kimi-code/src/tui/components/messages/tool-renderers/grep-output.ts
index 646a0de41cc..36111c39303 100644
--- a/apps/kimi-code/src/tui/components/messages/tool-renderers/grep-output.ts
+++ b/apps/kimi-code/src/tui/components/messages/tool-renderers/grep-output.ts
@@ -64,6 +64,10 @@ const PAGINATION_TOTAL = /^Results truncated to \d+ lines \(total: (\d+)/m;
 const INCOMPLETE =
   /^(?:\[Output truncated at \d+ bytes|Grep timed out after |Glob timed out after |Glob completed with warnings|\[stdout truncated at |\[Truncated at \d+ matches|Only the first \d+ matches)/m;
 
+const GLOB_PAGE = /^Showing matches (\d+)–(\d+) of (\d+)( collected matches \(partial result set\))?\.$/m;
+const GLOB_CONTINUATION = /^(?:Continue with the same search arguments and offset=\d+\.|(?:To retrieve all collected matches in one search|To remove the match-count limit), omit offset and use head_limit=0\.|Character limit reached; only complete paths are returned\.)$/;
+const GLOB_EMPTY = /^(?:No more matches at offset=\d+ in the (?:current|collected partial) result set \(\d+ matches\)\.|No matches collected; search incomplete\.)$/m;
+
 // `path:line:text`; context lines use `-` separators and are not matches.
 const CONTENT_MATCH = /^(.+?):(\d+):/;
 const COUNT_LINE = /^(.+):(\d+)$/;
@@ -177,7 +181,13 @@ export function parseGrepOutput(toolCall: ToolCallBlockData, output: string): Gr
 }
 
 export function parseGlobOutput(output: string): GlobStats {
-  return { entries: resultLines(output), partial: INCOMPLETE.test(output) };
+  const page = GLOB_PAGE.exec(output);
+  const entries = resultLines(output).filter((line) =>
+    !GLOB_PAGE.test(line) && !GLOB_CONTINUATION.test(line) && !GLOB_EMPTY.test(line),
+  );
+  const partial = INCOMPLETE.test(output) ||
+    (page !== null && (Number(page[2]) < Number(page[3]) || page[4] !== undefined));
+  return { entries, partial };
 }
 
 // Every match was a file the tool excludes as sensitive: the search did find
@@ -195,5 +205,8 @@ export function searchNoticeOnly(toolCall: ToolCallBlockData, output: string): b
     toolCall.name === 'Glob'
       ? parseGlobOutput(output).entries.length === 0
       : parseGrepOutput(toolCall, output).entries.length === 0;
-  return noRows && (INCOMPLETE.test(output) || SENSITIVE_ONLY.test(output));
+  return noRows && (
+    INCOMPLETE.test(output) || SENSITIVE_ONLY.test(output) ||
+    (toolCall.name === 'Glob' && GLOB_EMPTY.test(output))
+  );
 }
diff --git a/apps/kimi-code/test/tui/components/messages/tool-renderers/chip.test.ts b/apps/kimi-code/test/tui/components/messages/tool-renderers/chip.test.ts
index da5b51f2f31..849ce82b7e2 100644
--- a/apps/kimi-code/test/tui/components/messages/tool-renderers/chip.test.ts
+++ b/apps/kimi-code/test/tui/components/messages/tool-renderers/chip.test.ts
@@ -139,6 +139,36 @@ describe('chip registry', () => {
     expect(chipFor('Glob', { pattern: '**/*.ts' }, result('a.ts\nb.ts'))).toBe('2 files');
   });
 
+  it.each([
+    'To retrieve all collected matches in one search, omit offset and use head_limit=0.',
+    'To remove the match-count limit, omit offset and use head_limit=0.',
+  ])('counts only paths on a Glob page with continuation notice: %s', (notice) => {
+    const output = [
+      'Showing matches 1–100 of 347.',
+      'Continue with the same search arguments and offset=100.',
+      notice,
+      ...Array.from({ length: 100 }, (_, i) => `file-${String(i)}.ts`),
+    ].join('\n');
+    expect(chipFor('Glob', {}, result(output))).toBe('100+ files');
+  });
+
+  it.each([
+    'No more matches at offset=347 in the current result set (347 matches).',
+    'No matches collected; search incomplete.',
+  ])('does not count an empty Glob page as a file: %s', (output) => {
+    expect(chipFor('Glob', {}, result(output))).toBe('');
+  });
+
+  it('distinguishes the last Glob page from incomplete search results', () => {
+    expect(chipFor('Glob', {}, result('Showing matches 3–4 of 4.\nc.ts\nd.ts'))).toBe('2 files');
+    expect(chipFor('Glob', {}, result('Showing matches 3–4 of 4 collected matches (partial result set).\nc.ts\nd.ts'))).toBe('2+ files');
+  });
+
+  it('keeps notice-like file names and leaves Grep interpretation unchanged', () => {
+    expect(chipFor('Glob', {}, result('Showing matches.ts\nContinue with.txt\nNo more matches.ts'))).toBe('3 files');
+    expect(chipFor('Grep', {}, result('Showing matches 1–2 of 3.'))).toBe('1 file');
+  });
+
   it('FetchURL chip shows size and is non-empty', () => {
     const out = chipFor('FetchURL', { url: 'https://example.com' }, result('hello world'));
     expect(out).toMatch(/\d+\s*B/);
diff --git a/apps/kimi-code/test/tui/components/messages/tool-renderers/registry.test.ts b/apps/kimi-code/test/tui/components/messages/tool-renderers/registry.test.ts
index 106afe05e4e..81b488c1945 100644
--- a/apps/kimi-code/test/tui/components/messages/tool-renderers/registry.test.ts
+++ b/apps/kimi-code/test/tui/components/messages/tool-renderers/registry.test.ts
@@ -226,6 +226,21 @@ describe('tool-result registry', () => {
     expect(out).toContain('+1 more');
   });
 
+  it('keeps Glob pagination notices out of the collapsed path samples', () => {
+    const output = 'Showing matches 1–2 of 4.\nCharacter limit reached; only complete paths are returned.\nContinue with the same search arguments and offset=2.\na.ts\nb.ts';
+    const renderer = pickResultRenderer('Glob');
+    expect(strip(joinRender(renderer(call('Glob'), result(output), ctx)))).toBe('  a.ts, b.ts');
+    expect(strip(joinRender(renderer(call('Glob'), result(output), expandedCtx)))).toContain('Character limit reached');
+  });
+
+  it.each([
+    'No more matches at offset=347 in the current result set (347 matches).',
+    'No matches collected; search incomplete.',
+  ])('shows a Glob empty-page notice as the outcome: %s', (output) => {
+    const renderer = pickResultRenderer('Glob');
+    expect(strip(joinRender(renderer(call('Glob'), result(output), ctx), 160))).toBe(`  ${output}`);
+  });
+
   it('FetchURL renders no body when collapsed', () => {
     const renderer = pickResultRenderer('FetchURL');
     const out = joinRender(
diff --git a/docs/en/reference/tools.md b/docs/en/reference/tools.md
index 95f74d7b2fc..df56d7cb15a 100644
--- a/docs/en/reference/tools.md
+++ b/docs/en/reference/tools.md
@@ -29,7 +29,9 @@ Tail reads return the newest complete lines in the requested range first. If no
 
 **`Grep`** invokes ripgrep to search file contents, supporting regular expressions (`pattern`), a search path (`path`), file type filtering (`type`, e.g., `ts`, `py`), glob filtering (`glob`), and output mode (`output_mode`: `files_with_matches` / `content` / `count_matches`; defaults to `files_with_matches`). `content` mode supports context lines (`-A`, `-B`, `-C`), case-insensitive matching (`-i`), line numbers (`-n`, default true), and multiline matching (`multiline`). All modes support `offset` + `head_limit` pagination; `head_limit` defaults to 250 and `0` means unlimited. Sensitive files such as `.env` files and private keys are automatically filtered out; set `include_ignored=true` to search files ignored by `.gitignore`, though sensitive files remain filtered.
 
-**`Glob`** matches files in a specified directory (`path`; defaults to the working directory) by glob pattern (`pattern`). Results are sorted by modification time in descending order, with a maximum of 100 entries. It respects `.gitignore`, `.ignore`, and `.rgignore` by default; set `include_ignored=true` to include ignored files such as build outputs, while sensitive files remain filtered. Brace patterns such as `*.{ts,tsx}` are supported, and broad wildcard patterns are allowed but usually truncate at the match cap.
+**`Glob`** matches files in a specified directory (`path`; defaults to the working directory) by glob pattern (`pattern`). Results are sorted by modification time in descending order, returning 100 entries by default. It respects `.gitignore`, `.ignore`, and `.rgignore` by default; set `include_ignored=true` to include ignored files such as build outputs, while sensitive files remain filtered. Brace patterns such as `*.{ts,tsx}` are supported, and broad wildcard patterns are allowed.
+
+Use `offset` (default 0) and `head_limit` (default 100) to page through matching paths; the result provides the next offset when more matches are available. Set `head_limit: 0` to remove the match-count limit. The character limit still applies: pages end at a complete path and provide the next offset when necessary. Large pages are saved to a file that the agent can read with `Read`. Each call searches the current filesystem again, so file changes can shift results between pages. Timeouts, unreadable directories, or the output capture limit can still leave the search incomplete; the result warns about these cases, and increasing the offset cannot recover uncollected paths.
 
 **`ReadMediaFile`** sends an image or video to the model as multimodal content. It accepts `path`, plus optional image-detail controls such as `region` and `full_resolution`; the file size limit is 100 MB. Default image reads are compressed to the configured model limits. If automatic compression cannot meet those limits safely, the tool returns an error without sending the original image and directs the model to create and read a smaller copy. Availability depends on the current model's vision capabilities (`image_in` / `video_in`).
 
diff --git a/docs/zh/reference/tools.md b/docs/zh/reference/tools.md
index 9e6739e4497..f7faf9cda5c 100644
--- a/docs/zh/reference/tools.md
+++ b/docs/zh/reference/tools.md
@@ -29,7 +29,9 @@
 
 **`Grep`** 调用 ripgrep 搜索文件内容，支持正则表达式（`pattern`）、搜索路径（`path`）、文件类型过滤（`type`，如 `ts`、`py`）、glob 过滤（`glob`）和输出模式（`output_mode`：`files_with_matches` / `content` / `count_matches`，默认 `files_with_matches`）。`content` 模式支持上下文行（`-A`、`-B`、`-C`）、忽略大小写（`-i`）、行号（`-n`，默认 true）、跨行匹配（`multiline`）。所有模式支持 `offset` + `head_limit` 分页，`head_limit` 默认 250、传 0 表示不限。`.env`、私钥等敏感文件会被自动过滤；`include_ignored=true` 可搜索被 `.gitignore` 忽略的文件，但敏感文件仍保持过滤。
 
-**`Glob`** 按 glob 模式（`pattern`）在指定目录（`path`，默认工作目录）中匹配文件，结果按修改时间倒序排列，最多返回 100 条。默认尊重 `.gitignore`、`.ignore` 和 `.rgignore`；设置 `include_ignored=true` 可包含构建产物等被忽略的文件，但敏感文件仍会被过滤。支持 `*.{ts,tsx}` 这类花括号模式，也允许宽泛通配符模式，但通常会在匹配上限处截断。
+**`Glob`** 按 glob 模式（`pattern`）在指定目录（`path`，默认工作目录）中匹配文件，结果按修改时间倒序排列，默认返回 100 条。默认尊重 `.gitignore`、`.ignore` 和 `.rgignore`；设置 `include_ignored=true` 可包含构建产物等被忽略的文件，但敏感文件仍会被过滤。支持 `*.{ts,tsx}` 这类花括号模式，也允许宽泛通配符模式。
+
+使用 `offset`（默认 0）和 `head_limit`（默认 100）对匹配路径分页；有更多结果时，工具会给出下一页的 offset。设置 `head_limit: 0` 可取消条数限制，但字符上限仍然有效：达到上限时，页面会在完整路径处结束，并给出下一页的 offset。较大的页面会保存到文件，Agent 可用 `Read` 读取。每次调用都会重新搜索当前文件系统，因此文件变化可能导致跨页结果移动。超时、目录无法读取或输出采集上限仍可能造成搜索不完整；结果会提示这些情况，增加 offset 无法恢复尚未收集的路径。
 
 **`ReadMediaFile`** 将图片或视频以多模态内容发送给模型。它接受 `path`，以及 `region`、`full_resolution` 等可选的图片细节参数；文件大小上限为 100 MB。默认读图会按配置的模型限制压缩；如果自动压缩无法安全满足限制，工具会返回错误且不发送原图，并提示模型先创建更小的副本再读取。是否可用取决于当前模型的视觉能力（`image_in` / `video_in`）。
 
diff --git a/packages/agent-core-v2/src/agent/tools/os/glob/glob.md b/packages/agent-core-v2/src/agent/tools/os/glob/glob.md
index ad299e29afc..6e364bd919c 100644
--- a/packages/agent-core-v2/src/agent/tools/os/glob/glob.md
+++ b/packages/agent-core-v2/src/agent/tools/os/glob/glob.md
@@ -10,7 +10,9 @@ Good patterns:
 - `*.{ts,tsx}` — brace expansion is supported
 - `{src,test}/**/*.ts` — cartesian brace expansion is supported too
 
-Results are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.
+Results default to 100 matching paths. Use `offset` (default 0) and `head_limit` (default 100) to page through results. When more matches are available, the result gives the next offset; keep the other search arguments unchanged. Set `head_limit=0` to remove the match-count limit. Pages still stay within the character retention limit, including notices: when it is reached, only complete paths are returned, with the next offset for continuation. Large pages are saved to a file with a path for Read.
+
+Each call searches the current filesystem again; pagination is not a snapshot, and file changes can shift results between pages. To collect a large list, use `head_limit=0`, read any saved output, and follow continuation offsets if the character limit is reached. Search timeouts, traversal errors, and output capture limits can still produce partial results; the result reports these limits, and pagination cannot recover paths that were never collected. Narrow the search and retry when it is incomplete.
 
 Large-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when `include_ignored` is set:
-- `node_modules/**/*.js`, `.venv/**/*.py`, `__pycache__/**`, `target/**` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like `node_modules/react/src/**/*.js`.
+- `node_modules/**/*.js`, `.venv/**/*.py`, `__pycache__/**`, `target/**` can produce thousands of results and waste search time and context. Prefer specific subpaths like `node_modules/react/src/**/*.js` unless you need a complete listing.
diff --git a/packages/agent-core-v2/src/agent/tools/os/glob/glob.ts b/packages/agent-core-v2/src/agent/tools/os/glob/glob.ts
index 5043a2e68bf..a26fc4320e5 100644
--- a/packages/agent-core-v2/src/agent/tools/os/glob/glob.ts
+++ b/packages/agent-core-v2/src/agent/tools/os/glob/glob.ts
@@ -5,6 +5,22 @@ import { type AgentTool } from '#/tool/toolContract';
 
 export const GlobInputSchema = z.object({
   pattern: z.string().describe('Glob pattern to match files.'),
+  head_limit: z
+    .number()
+    .int()
+    .nonnegative()
+    .optional()
+    .describe(
+      'Maximum number of matching paths to return after offset. Defaults to 100. Pass 0 to remove the match-count limit. The character limit still applies: large pages are saved for Read, and a continuation offset is provided when more paths remain. Search time and output capture limits still apply.',
+    ),
+  offset: z
+    .number()
+    .int()
+    .nonnegative()
+    .optional()
+    .describe(
+      'Number of matching paths to skip. Defaults to 0. Each call searches the current filesystem again; changes can shift results between pages.',
+    ),
   path: z
     .string()
     .optional()
@@ -27,7 +43,7 @@ export const GlobInputSchema = z.object({
 
 export type GlobInput = z.infer<typeof GlobInputSchema>;
 
-export const MAX_MATCHES = 100;
+export const DEFAULT_HEAD_LIMIT = 100;
 
 export const WINDOWS_PATH_HINT =
   '\n\nWindows note: the `path` argument accepts both Windows paths ' +
diff --git a/packages/agent-core-v2/src/agent/tools/os/glob/globTool.ts b/packages/agent-core-v2/src/agent/tools/os/glob/globTool.ts
index 75110f802bb..7be6dda07a6 100644
--- a/packages/agent-core-v2/src/agent/tools/os/glob/globTool.ts
+++ b/packages/agent-core-v2/src/agent/tools/os/glob/globTool.ts
@@ -17,6 +17,7 @@ import { ISessionSkillCatalog } from '#/features/skill/session/skillCatalog';
 import { ISessionWorkspaceContext } from '#/session/workspaceContext/workspaceContext';
 import { ITelemetryService } from '#/app/telemetry/telemetry';
 import {
+  DEFAULT_TOOL_RESULT_MAX_RETAINED_CHARS,
   ToolAccesses,
   type ExecutableToolResult,
   type ToolExecution,
@@ -37,7 +38,7 @@ import {
   type GlobInput,
   GlobInputSchema,
   IGlobTool,
-  MAX_MATCHES,
+  DEFAULT_HEAD_LIMIT,
   WINDOWS_PATH_HINT,
 } from './glob';
 
@@ -238,50 +239,90 @@ export class GlobTool implements IGlobTool {
       }
     }
 
-    const truncated = kept.length > MAX_MATCHES;
-    const limited = truncated ? kept.slice(0, MAX_MATCHES) : kept;
-
-    if (limited.length === 0 && !timedOut) {
-      if (filteredSensitive > 0) {
-        return {
-          output: `No non-sensitive matches found (${String(filteredSensitive)} sensitive file(s) filtered).`,
-        };
-      }
-      return { output: 'No matches found' };
-    }
+    const offset = args.offset ?? 0;
+    const headLimit = args.head_limit ?? DEFAULT_HEAD_LIMIT;
+    const limited = headLimit === 0 ? kept.slice(offset) : kept.slice(offset, offset + headLimit);
+    const partial = bufferTruncated || timedOut || traversalWarning !== undefined;
 
     const pathClass = env.pathClass;
     const shouldRelativize = isWithinDirectory(searchRoot, workspace.workspaceDir, pathClass);
-    const displayLines = limited.map((p) =>
+    const candidates = limited.map((p) =>
       shouldRelativize ? relativizeIfUnder(p, searchRoot, pathClass) : p,
     );
 
-    const lines: string[] = [];
+    const warnings: string[] = [];
     if (timedOut) {
-      lines.push(
+      warnings.push(
         `Glob timed out after ${String(DEFAULT_TIMEOUT_MS / 1000)}s; partial results returned.`,
       );
     }
     if (bufferTruncated) {
-      lines.push(
+      warnings.push(
         `[stdout truncated at ${String(MAX_OUTPUT_BYTES)} bytes; results may be incomplete — use a more specific pattern]`,
       );
     }
     if (traversalWarning !== undefined) {
-      lines.push(traversalWarning);
+      warnings.push(traversalWarning);
     }
-    if (truncated) {
-      lines.push(`[Truncated at ${String(MAX_MATCHES)} matches — use a more specific pattern]`);
-      lines.push(`Only the first ${String(MAX_MATCHES)} matches are returned.`);
-    }
-    lines.push(...displayLines);
-    if (filteredSensitive > 0) {
-      lines.push(`Filtered ${String(filteredSensitive)} sensitive file(s).`);
+    const pageNotices = (count: number, characterLimited: boolean) => {
+      const lines = [...warnings];
+      const footer: string[] = [];
+      const truncated = characterLimited || offset + count < kept.length;
+      if (count === 0) {
+        if (kept.length > 0) {
+          const resultSet = partial ? 'collected partial result set' : 'current result set';
+          lines.push(
+            `No more matches at offset=${String(offset)} in the ${resultSet} (${String(kept.length)} matches).`,
+          );
+        } else if (partial) {
+          lines.push('No matches collected; search incomplete.');
+        } else if (filteredSensitive > 0) {
+          lines.push(
+            `No non-sensitive matches found (${String(filteredSensitive)} sensitive file(s) filtered).`,
+          );
+        } else {
+          lines.push('No matches found');
+        }
+      } else if (truncated || offset > 0 || partial) {
+        const total = partial
+          ? `${String(kept.length)} collected matches (partial result set)`
+          : String(kept.length);
+        lines.push(`Showing matches ${String(offset + 1)}–${String(offset + count)} of ${total}.`);
+      }
+      if (characterLimited) lines.push('Character limit reached; only complete paths are returned.');
+      if (truncated) {
+        lines.push(
+          `Continue with the same search arguments and offset=${String(offset + count)}.`,
+        );
+        if (!characterLimited) lines.push('To remove the match-count limit, omit offset and use head_limit=0.');
+      }
+      if (filteredSensitive > 0 && (kept.length > 0 || partial)) {
+        footer.push(`Filtered ${String(filteredSensitive)} sensitive file(s).`);
+      }
+      if (!truncated && !partial && offset === 0 && headLimit > 0 && count === headLimit) {
+        footer.push(`Found ${String(count)} matches`);
+      }
+      return { lines, footer };
+    };
+    const noticeChars = Math.max(...[false, true].map((characterLimited) => {
+      const { lines, footer } = pageNotices(candidates.length, characterLimited);
+      return [...lines, ...footer].join('\n').length + 2;
+    }));
+    let remaining = DEFAULT_TOOL_RESULT_MAX_RETAINED_CHARS - noticeChars;
+    const displayLines: string[] = [];
+    for (const path of candidates) {
+      if (path.length + 1 > remaining) break;
+      displayLines.push(path);
+      remaining -= path.length + 1;
     }
-    if (!truncated && limited.length === MAX_MATCHES) {
-      lines.push(`Found ${String(limited.length)} matches`);
+    if (candidates.length > 0 && displayLines.length === 0) {
+      return {
+        isError: true,
+        output: 'Glob cannot fit a complete path and its diagnostics within the output limit. Narrow the search path or pattern.',
+      };
     }
-    return { output: lines.join('\n') };
+    const notices = pageNotices(displayLines.length, displayLines.length < candidates.length);
+    return { output: [...notices.lines, ...displayLines, ...notices.footer].join('\n') };
   }
 }
 
diff --git a/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts b/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
index b6d09f9eb25..4d8c49bf233 100644
--- a/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
+++ b/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
@@ -673,7 +673,7 @@ describe('FullCompaction', () => {
       event: 'compaction_finished',
       properties: expect.objectContaining({
         source: 'manual',
-        tokens_before: 18_510,
+        tokens_before: expect.any(Number),
         retry_count: 1,
         trace_id: 'trace-compact-1',
       }),
@@ -1175,7 +1175,7 @@ describe('FullCompaction', () => {
       properties: expect.objectContaining({
         agent_id: 'main',
         source: 'manual',
-        tokens_before: 18_510,
+        tokens_before: expect.any(Number),
         duration_ms: expect.any(Number),
         round: 1,
         retry_count: 0,
@@ -1400,7 +1400,7 @@ describe('FullCompaction', () => {
       event: 'compaction_failed',
       properties: expect.objectContaining({
         source: 'manual',
-        tokens_before: 18_510,
+        tokens_before: expect.any(Number),
         duration_ms: expect.any(Number),
         retry_count: 4,
         error_type: 'APIConnectionError',
@@ -1609,6 +1609,7 @@ describe('FullCompaction', () => {
     const ctx = testAgent();
     ctx.configure({
       provider: CATALOGUED_PROVIDER,
+      tools: SNAPSHOT_VISIBLE_TOOLS,
       modelCapabilities: {
         ...CATALOGUED_MODEL_CAPABILITIES,
         max_context_tokens: maxContextTokens,
diff --git a/packages/agent-core-v2/test/agent/loop/loop.test.ts b/packages/agent-core-v2/test/agent/loop/loop.test.ts
index d62565cb8ec..a2447aa789b 100644
--- a/packages/agent-core-v2/test/agent/loop/loop.test.ts
+++ b/packages/agent-core-v2/test/agent/loop/loop.test.ts
@@ -152,8 +152,8 @@ describe('Agent loop', () => {
       [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-1>", "turnId": "0", "step": 1 }, "time": "<time>" }
       [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "blocked" }
       [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
-      [wire] llm.tools_snapshot          { "agentId": "main", "hash": "0432af3f1410334f38bb79c6965876003ddf8ff088b07654736330b30ffe1791", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
-      [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "0432af3f1410334f38bb79c6965876003ddf8ff088b07654736330b30ffe1791", "messageCount": 1, "turnStep": "0.1", "time": "<time>" }
+      [wire] llm.tools_snapshot          { "agentId": "main", "hash": "6658fc0337183153a13433d81fa64064a006b4466b6051bebe186a78deb044df", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults default to 100 matching paths. Use \`offset\` (default 0) and \`head_limit\` (default 100) to page through results. When more matches are available, the result gives the next offset; keep the other search arguments unchanged. Set \`head_limit=0\` to remove the match-count limit. Pages still stay within the character retention limit, including notices: when it is reached, only complete paths are returned, with the next offset for continuation. Large pages are saved to a file with a path for Read.\\n\\nEach call searches the current filesystem again; pagination is not a snapshot, and file changes can shift results between pages. To collect a large list, use \`head_limit=0\`, read any saved output, and follow continuation offsets if the character limit is reached. Search timeouts, traversal errors, and output capture limits can still produce partial results; the result reports these limits, and pagination cannot recover paths that were never collected. Narrow the search and retry when it is incomplete.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results and waste search time and context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\` unless you need a complete listing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "head_limit": { "description": "Maximum number of matching paths to return after offset. Defaults to 100. Pass 0 to remove the match-count limit. The character limit still applies: large pages are saved for Read, and a continuation offset is provided when more paths remain. Search time and output capture limits still apply.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of matching paths to skip. Defaults to 0. Each call searches the current filesystem again; changes can shift results between pages.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+      [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "6658fc0337183153a13433d81fa64064a006b4466b6051bebe186a78deb044df", "messageCount": 1, "turnStep": "0.1", "time": "<time>" }
       [emit] agent.status.updated        { "time": "<time>", "agentId": "main", "usage": { "byModel": { "mock-model": { "inputOther": 3, "output": 5, "inputCacheRead": 0, "inputCacheCreation": 0 } }, "total": { "inputOther": 3, "output": 5, "inputCacheRead": 0, "inputCacheCreation": 0 }, "currentTurn": { "inputOther": 3, "output": 5, "inputCacheRead": 0, "inputCacheCreation": 0 } } }
       [emit] agent.status.updated        { "time": "<time>", "agentId": "main", "contextTokens": 8 }
       [wire] usage.record                { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 3, "output": 5, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
diff --git a/packages/agent-core-v2/test/agent/toolExecutor/toolExecutor.test.ts b/packages/agent-core-v2/test/agent/toolExecutor/toolExecutor.test.ts
index b0b20e306ee..ad611ba10ee 100644
--- a/packages/agent-core-v2/test/agent/toolExecutor/toolExecutor.test.ts
+++ b/packages/agent-core-v2/test/agent/toolExecutor/toolExecutor.test.ts
@@ -2,6 +2,7 @@ import { readFileSync } from 'node:fs';
 import { mkdtemp, rm, writeFile } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
+import { PassThrough, Readable } from 'node:stream';
 
 import type { ToolCall } from '#human/llm/message';
 import type { ToolInputDisplay } from '#/tool/toolInputDisplay';
@@ -37,9 +38,11 @@ import { parseToolCallArguments } from '#/tool/tool-args-parse';
 import { IAgentToolResultTruncationService } from '#/agent/toolResultTruncation/toolResultTruncation';
 import { ToolResultTruncationService } from '#/agent/toolResultTruncation/toolResultTruncationService';
 import { ReadTool } from '#/agent/tools/os/read/readTool';
+import { GlobTool } from '#/agent/tools/os/glob/globTool';
 import { ReadInputSchema, type ReadInput } from '#/agent/tools/os/read/read';
 import { renderToolResultForModel } from '#/agent/contextMemory/toolResultRender';
 import { HostFileSystem } from '#/os/backends/node-local/hostFsService';
+import { HostProcessService } from '#/os/backends/node-local/hostProcessService';
 import { FakeRuntime } from '#/runtime/fakeRuntime';
 import type { IAgentRuntimeService } from '#/agent/runtimeBinding/agentRuntime';
 import type { ISessionSkillCatalog } from '#/features/skill/session/skillCatalog';
@@ -54,7 +57,7 @@ import { IAgentToolRegistryService } from '#/agent/toolRegistry/toolRegistry';
 import { AgentToolRegistryService } from '#/agent/toolRegistry/toolRegistryService';
 import { IEventBus } from '#/app/event/eventBus';
 import type { LLMRequestTrace } from '#/llm-adapter/contract/request-trace';
-import { ITelemetryService } from '#/app/telemetry/telemetry';
+import { ITelemetryService, noopTelemetryService } from '#/app/telemetry/telemetry';
 import { IBootstrapService } from '#/app/bootstrap/bootstrap';
 import { FileStorageService } from '#/persistence/backends/node-fs/fileStorageService';
 import { IFileSystemStorageService } from '#/persistence/interface/storage';
@@ -1044,6 +1047,7 @@ describe('parseToolCallArguments', () => {
 describe('truncation pipeline', () => {
   let homeDir: string;
   let readConfig: IConfigService;
+  let globProcess: HostProcessService;
 
   beforeEach(async () => {
     homeDir = await mkdtemp(join(tmpdir(), 'tool-executor-truncation-'));
@@ -1069,10 +1073,11 @@ describe('truncation pipeline', () => {
     truncationContainer.set(IConfigService, new SyncDescriptor(ConfigService));
     readConfig = truncationContainer.get(IConfigService);
     await readConfig.ready;
+    globProcess = new HostProcessService();
     const runtime = Object.assign(new FakeRuntime(
       { workspaceId: 'workspace', runtimeId: 'local', generation: 'test' },
-      { capabilities: ['fs'] },
-    ), { fs: new HostFileSystem() });
+      { capabilities: ['fs', 'process'] },
+    ), { fs: new HostFileSystem(), process: globProcess });
     const binding: IAgentRuntimeService = {
       _serviceBrand: undefined,
       onDidChange: () => ({ dispose: () => {} }),
@@ -1087,6 +1092,7 @@ describe('truncation pipeline', () => {
       truncation,
       readConfig,
     ));
+    registry.register(new GlobTool(binding, stubWorkspaceContext(homeDir), noopTelemetryService));
   });
 
   afterEach(async () => {
@@ -1129,6 +1135,83 @@ describe('truncation pipeline', () => {
     expect(readFileSync(outputPath, 'utf8')).toBe(fullOutput);
   });
 
+  it('recovers every Glob match through spill and Read when the match limit is disabled', async () => {
+    const expected = Array.from({ length: 500 }, (_, index) =>
+      `file-${String(index).padStart(3, '0')}-${'x'.repeat(100)}.ts`,
+    );
+    await Promise.all(expected.map((name) => writeFile(join(homeDir, name), '')));
+
+    const [result] = await execute([toolCall('glob_all', 'Glob', { pattern: '*.ts', head_limit: 0 })]);
+
+    expect(result?.isError).not.toBe(true);
+    expect(result?.truncated).toBe(true);
+    if (typeof result?.output !== 'string') throw new Error('expected Glob text');
+    const path = renderedOutputPath(result.output);
+    let args: ReadInput | undefined = { path, max_chars: 8000 };
+    const recovered: string[] = [];
+    let pages = 0;
+    while (args !== undefined && pages < 20) {
+      const [page] = await execute([toolCall(`read_glob_${String(pages++)}`, 'Read', args)]);
+      expect(page?.isError).not.toBe(true);
+      if (typeof page?.output !== 'string') throw new Error('expected Read text');
+      recovered.push(...page.output.replaceAll(/^\d+\t/gm, '').split('\n').filter(Boolean));
+      const next = /Next Read: (\{[^\n]*\})/.exec(page.note ?? '')?.[1];
+      args = next === undefined ? undefined : ReadInputSchema.parse(JSON.parse(next));
+    }
+    expect(args).toBeUndefined();
+    expect(pages).toBeGreaterThan(1);
+    expect(recovered.toSorted()).toEqual(expected);
+  });
+
+  it('recovers an expanded Glob listing beyond spill retention using complete saved pages', async () => {
+    const root = await mkdtemp(join(tmpdir(), 'r'.repeat(180)));
+    try {
+      const names = Array.from({ length: 60_000 }, (_, i) => `file-${String(i).padStart(6, '0')}.ts`);
+      const stdout = names.map((name) => `./${name}`).join('\n') + '\n';
+      vi.spyOn(globProcess, 'spawn').mockImplementation(async () => ({
+        _serviceBrand: undefined,
+        pid: 123,
+        exitCode: 0,
+        stdin: new PassThrough(),
+        stdout: Readable.from([stdout]),
+        stderr: Readable.from([]),
+        wait: async () => 0,
+        kill: async () => {},
+        dispose: () => {},
+      }));
+      const recovered: string[] = [];
+      let offset = 0;
+      let globPages = 0;
+      do {
+        const [page] = await execute([toolCall(`glob_large_${String(globPages++)}`, 'Glob', {
+          pattern: '*.ts', path: root, head_limit: 0, offset,
+        })]);
+        expect(page?.isError).not.toBe(true);
+        if (typeof page?.output !== 'string') throw new Error('expected Glob output');
+        expect(page.output).toContain('the full output was saved to a file');
+        const continuation = /Continue with the same search arguments and offset=(\d+)\./.exec(page.output)?.[1];
+        if (continuation !== undefined) expect(Number(continuation)).toBeGreaterThan(offset);
+        offset = continuation === undefined ? 0 : Number(continuation);
+        let args: ReadInput | undefined = { path: renderedOutputPath(page.output), max_chars: 500_000 };
+        let reads = 0;
+        while (args !== undefined && reads < 40) {
+          const [read] = await execute([toolCall(`read_large_${String(globPages)}_${String(reads++)}`, 'Read', args)]);
+          expect(read?.isError).not.toBe(true);
+          if (typeof read?.output !== 'string') throw new Error('expected Read output');
+          recovered.push(...read.output.replaceAll(/^\d+\t/gm, '').split('\n').filter((line) => line.startsWith(root + '/')));
+          const next = /Next Read: (\{[^\n]*\})/.exec(read.note ?? '')?.[1];
+          args = next === undefined ? undefined : ReadInputSchema.parse(JSON.parse(next));
+        }
+        expect(args).toBeUndefined();
+      } while (offset > 0 && globPages < 5);
+      expect(offset).toBe(0);
+      expect(globPages).toBe(2);
+      expect(recovered).toEqual(names.map((name) => `${root}/${name}`));
+    } finally {
+      await rm(root, { recursive: true, force: true });
+    }
+  });
+
   it('recovers MCP structured records through spill and Read without repeating the MCP call', async () => {
     const structuredContent = {
       rows: Array.from({ length: 1200 }, (_, index) => ({
diff --git a/packages/agent-core-v2/test/os/backends/node-local/tools/glob.test.ts b/packages/agent-core-v2/test/os/backends/node-local/tools/glob.test.ts
index d52364d311f..3ffde881dc9 100644
--- a/packages/agent-core-v2/test/os/backends/node-local/tools/glob.test.ts
+++ b/packages/agent-core-v2/test/os/backends/node-local/tools/glob.test.ts
@@ -13,7 +13,7 @@ import { stubWorkspaceContext } from '../../../../session/workspaceContext/stub-
 import {
   type GlobInput,
   GlobInputSchema,
-  MAX_MATCHES,
+  DEFAULT_HEAD_LIMIT,
   WINDOWS_PATH_HINT,
 } from '#/agent/tools/os/glob/glob';
 import { GlobTool, splitCompletePaths } from '#/agent/tools/os/glob/globTool';
@@ -303,9 +303,9 @@ describe('GlobTool', () => {
     withCwd.toHaveBeenCalledWith('C:/WORKSPACE');
   });
 
-  it('walks pure-wildcard patterns, capping at MAX_MATCHES', async () => {
+  it('walks pure-wildcard patterns and gives a continuation for the default page', async () => {
     const stdout =
-      Array.from({ length: MAX_MATCHES + 5 }, (_, i) => `/workspace/${String(i)}.ts`).join('\n') +
+      Array.from({ length: DEFAULT_HEAD_LIMIT + 5 }, (_, i) => `/workspace/${String(i)}.ts`).join('\n') +
       '\n';
     const exec = execReturning(stdout);
     const { tool, withCwd } = makeTool(workspace, { exec });
@@ -315,7 +315,7 @@ describe('GlobTool', () => {
     expect(result.isError).toBeFalsy();
     withCwd.toHaveBeenCalledWith('/workspace');
     expect(execArgs(exec).at(-1)).toBe('.');
-    expect(result.output).toContain(`[Truncated at ${String(MAX_MATCHES)} matches`);
+    expect(result.output).toContain('Continue with the same search arguments and offset=100.');
   });
 
   it('passes a brace pattern through to a single rg --glob', async () => {
@@ -387,23 +387,23 @@ describe('GlobTool', () => {
     expect(execArgs(exec)).not.toContain('--no-ignore');
   });
 
-  it('caps returned matches and surfaces the truncation header', async () => {
+  it('limits the default page and gives the next offset', async () => {
     const stdout =
-      Array.from({ length: MAX_MATCHES + 1 }, (_, i) => `/workspace/${String(i)}.ts`).join('\n') +
+      Array.from({ length: DEFAULT_HEAD_LIMIT + 1 }, (_, i) => `/workspace/${String(i)}.ts`).join('\n') +
       '\n';
     const exec = execReturning(stdout);
     const { tool } = makeTool(stubWorkspaceContext('/workspace'), { exec });
 
     const result = await execute(tool, { pattern: '*.ts' });
 
-    expect(result.output).toContain(`[Truncated at ${String(MAX_MATCHES)} matches`);
+    expect(result.output).toContain('Continue with the same search arguments and offset=100.');
     expect(result.output).toContain('0.ts');
-    expect(result.output).not.toContain(`${String(MAX_MATCHES)}.ts`);
+    expect(result.output).not.toContain(`${String(DEFAULT_HEAD_LIMIT)}.ts`);
   });
 
-  it('surfaces a "first N matches" header when matches exceed MAX_MATCHES', async () => {
+  it('offers an all-matches query when the default page is incomplete', async () => {
     const stdout =
-      Array.from({ length: MAX_MATCHES + 50 }, (_, i) => `/workspace/file_${String(i)}.txt`).join(
+      Array.from({ length: DEFAULT_HEAD_LIMIT + 50 }, (_, i) => `/workspace/file_${String(i)}.txt`).join(
         '\n',
       ) + '\n';
     const exec = execReturning(stdout);
@@ -411,12 +411,12 @@ describe('GlobTool', () => {
 
     const result = await execute(tool, { pattern: '*.txt' });
 
-    expect(result.output).toContain(`Only the first ${String(MAX_MATCHES)} matches are returned`);
+    expect(result.output).toContain('To remove the match-count limit, omit offset and use head_limit=0.');
   });
 
-  it('returns a "Found N matches" footer at exactly MAX_MATCHES without truncation', async () => {
+  it('returns a "Found N matches" footer at exactly DEFAULT_HEAD_LIMIT without truncation', async () => {
     const stdout =
-      Array.from({ length: MAX_MATCHES }, (_, i) => `/workspace/test_${String(i)}.py`).join('\n') +
+      Array.from({ length: DEFAULT_HEAD_LIMIT }, (_, i) => `/workspace/test_${String(i)}.py`).join('\n') +
       '\n';
     const exec = execReturning(stdout);
     const { tool } = makeTool(stubWorkspaceContext('/workspace'), { exec });
@@ -424,7 +424,95 @@ describe('GlobTool', () => {
     const result = await execute(tool, { pattern: '*.py' });
 
     expect(result.output).not.toContain('Only the first');
-    expect(result.output).toContain(`Found ${String(MAX_MATCHES)} matches`);
+    expect(result.output).toContain(`Found ${String(DEFAULT_HEAD_LIMIT)} matches`);
+  });
+
+  it.each([2, 10])('distinguishes an exhausted page at offset=%s from a search with no matches', async (offset) => {
+    const { tool } = makeTool(workspace, { exec: execReturning('/workspace/a.ts\n/workspace/b.ts\n') });
+    const result = await execute(tool, GlobInputSchema.parse({ pattern: '*.ts', offset }));
+    expect(result.output).toBe(`No more matches at offset=${String(offset)} in the current result set (2 matches).`);
+  });
+
+  it.each([0, 1, 3])('keeps traversal warnings and partial counts on the page at offset=%s', async (offset) => {
+    const { tool } = makeTool(workspace, {
+      exec: execReturning('/workspace/a.ts\n/workspace/b.ts\n/workspace/c.ts\n', 'rg: ./locked: Permission denied', 2),
+    });
+    const result = await execute(tool, GlobInputSchema.parse({ pattern: '*.ts', offset, head_limit: 1 }));
+    expect(result.output).toContain('Permission denied');
+    expect(result.output).toContain('partial result set');
+    expect(result.output).not.toContain('No matches found');
+    if (offset < 3) expect(result.output).toContain('of 3 collected matches (partial result set).');
+    else expect(result.output).toContain('No more matches at offset=3');
+  });
+
+  it.each([
+    { head_limit: 1, expected: ['b.ts'] },
+    { head_limit: 0, expected: ['b.ts', 'c.ts'] },
+  ])('applies head_limit=$head_limit and offset after sensitive filtering', async ({ head_limit, expected }) => {
+    const { tool } = makeTool(workspace, {
+      exec: execReturning('/workspace/.env\n/workspace/a.ts\n/workspace/.aws/credentials\n/workspace/b.ts\n/workspace/c.ts\n'),
+    });
+    const result = await execute(tool, GlobInputSchema.parse({ pattern: '**', offset: 1, head_limit }));
+    const text = toolContentString(result);
+    expect(text.split('\n').filter((line) => line.endsWith('.ts'))).toEqual(expected);
+    expect(text).not.toContain('.env');
+    expect(text).not.toContain('credentials');
+    expect(text).toContain('of 3.');
+  });
+
+  it('does not present a capped capture as a complete search when returning all matches', async () => {
+    const { tool } = makeTool(workspace, {
+      exec: execReturning(`/workspace/${'x'.repeat(200)}.ts\n`.repeat(50_000)),
+    });
+    const result = await execute(tool, GlobInputSchema.parse({ pattern: '*.ts', head_limit: 0 }));
+    expect(result.output).toContain('stdout truncated');
+    expect(result.output).toContain('collected matches (partial result set)');
+    expect(result.output).not.toContain('Continue with');
+  });
+
+  it('keeps the timeout warning when the requested page has no collected matches', async () => {
+    vi.useFakeTimers({ toFake: ['setTimeout', 'clearTimeout'] });
+    try {
+      let finish!: (code: number) => void;
+      const completion = new Promise<number>((resolve) => { finish = resolve; });
+      const proc = fakeProcess('/workspace/a.ts\n');
+      const { tool } = makeTool(workspace, {
+        exec: vi.fn().mockResolvedValue({
+          ...proc,
+          exitCode: null,
+          wait: () => completion,
+          kill: async () => { finish(143); },
+        }),
+      });
+      const pending = execute(tool, GlobInputSchema.parse({ pattern: '*.ts', offset: 1 }));
+      await vi.advanceTimersByTimeAsync(20_000);
+      const result = await pending;
+      expect(result.output).toContain('Glob timed out');
+      expect(result.output).toContain('No more matches at offset=1 in the collected partial result set');
+      expect(result.output).not.toContain('No matches found');
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
+  it.each([0, 60_000])('keeps expanded paths within spill retention and continues with head_limit=%s', async (head_limit) => {
+    const root = `/${'r'.repeat(180)}`;
+    const names = Array.from({ length: 60_000 }, (_, i) => `file-${String(i).padStart(6, '0')}.ts`);
+    const stdout = names.map((name) => `./${name}`).join('\n') + '\n';
+    const { tool } = makeTool(workspace, { exec: vi.fn(async () => fakeProcess(stdout)) });
+    const first = toolContentString(await execute(tool, { pattern: '*.ts', path: root, head_limit }));
+    expect(first.length).toBeLessThanOrEqual(10_000_000);
+    expect(first).toContain('Character limit reached; only complete paths are returned.');
+    expect(first).not.toContain('head_limit=0');
+    const next = /Continue with the same search arguments and offset=(\d+)\./.exec(first)?.[1];
+    expect(next).toBeDefined();
+    const second = toolContentString(await execute(tool, {
+      pattern: '*.ts', path: root, head_limit, offset: Number(next),
+    }));
+    expect(second.length).toBeLessThanOrEqual(10_000_000);
+    expect(second).not.toContain('Continue with');
+    const recovered = [...first.split('\n'), ...second.split('\n')].filter((line) => line.startsWith('/'));
+    expect(recovered).toEqual(names.map((name) => `${root}/${name}`));
   });
 
   it('filters sensitive files from results', async () => {
@@ -867,6 +955,23 @@ describe('GlobTool integration (real ripgrep)', () => {
 
   const ws = () => stubWorkspaceContext(tmpDir!);
 
+  it('continues through every match beyond the default page without duplicates', async () => {
+    const expected = Array.from({ length: 347 }, (_, index) => `file-${String(index).padStart(3, '0')}.ts`);
+    await Promise.all(expected.map((name, index) => touch(name, new Date(1_700_000_000_000 - index * 1000))));
+    const tool = new GlobTool(createRuntime(realFs, realEnv, realProcessService), ws(), noopTelemetryService);
+    const recovered: string[] = [];
+    for (const offset of [0, 100, 200, 300]) {
+      const result = await execute(tool, GlobInputSchema.parse({ pattern: '*.ts', offset }));
+      expect(result.isError).not.toBe(true);
+      const text = toolContentString(result);
+      recovered.push(...text.split('\n').filter((line) => /^file-\d+\.ts$/.test(line)));
+      expect(text).toContain(`Showing matches ${String(offset + 1)}–${String(Math.min(offset + 100, 347))} of 347.`);
+      if (offset < 300) expect(text).toContain(`offset=${String(offset + 100)}`);
+      else expect(text).not.toContain('Continue with');
+    }
+    expect(recovered).toEqual(expected);
+  });
+
   it('returns files newest-first by modification time (--sortr=modified)', async () => {
     await touch('old.ts', new Date('2020-01-01T00:00:00Z'));
     await touch('mid.ts', new Date('2022-01-01T00:00:00Z'));
diff --git a/packages/agent-core-v2/test/tool/tool.test.ts b/packages/agent-core-v2/test/tool/tool.test.ts
index b96fdbf0891..d1668b0f5c6 100644
--- a/packages/agent-core-v2/test/tool/tool.test.ts
+++ b/packages/agent-core-v2/test/tool/tool.test.ts
@@ -4321,10 +4321,10 @@ describe('Agent tools', () => {
         [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-1>", "turnId": "0", "step": 1 }, "time": "<time>" }
         [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "I will look it up." }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
-        [wire] llm.tools_snapshot          { "agentId": "main", "hash": "f8d3fb9a7c26c38032bc3a2beb2fb4ac327a0ecb91388387028fcbe398f8df0a", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Lookup", "description": "Look up a short test value.", "parameters": { "type": "object", "properties": { "query": { "type": "string" } }, "required": [ "query" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+        [wire] llm.tools_snapshot          { "agentId": "main", "hash": "0841d488bc8e0d298f206bfe6d41c4b893d11ac3046cff11b0f1a67a7b67186b", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults default to 100 matching paths. Use \`offset\` (default 0) and \`head_limit\` (default 100) to page through results. When more matches are available, the result gives the next offset; keep the other search arguments unchanged. Set \`head_limit=0\` to remove the match-count limit. Pages still stay within the character retention limit, including notices: when it is reached, only complete paths are returned, with the next offset for continuation. Large pages are saved to a file with a path for Read.\\n\\nEach call searches the current filesystem again; pagination is not a snapshot, and file changes can shift results between pages. To collect a large list, use \`head_limit=0\`, read any saved output, and follow continuation offsets if the character limit is reached. Search timeouts, traversal errors, and output capture limits can still produce partial results; the result reports these limits, and pagination cannot recover paths that were never collected. Narrow the search and retry when it is incomplete.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results and waste search time and context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\` unless you need a complete listing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "head_limit": { "description": "Maximum number of matching paths to return after offset. Defaults to 100. Pass 0 to remove the match-count limit. The character limit still applies: large pages are saved for Read, and a continuation offset is provided when more paths remain. Search time and output capture limits still apply.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of matching paths to skip. Defaults to 0. Each call searches the current filesystem again; changes can shift results between pages.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Lookup", "description": "Look up a short test value.", "parameters": { "type": "object", "properties": { "query": { "type": "string" } }, "required": [ "query" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
         [emit] tool.call.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "toolCallId": "call_lookup", "name": "Lookup", "argumentsPart": "{\\"query\\":\\"moon\\"}" }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "tool_call", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
-        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "f8d3fb9a7c26c38032bc3a2beb2fb4ac327a0ecb91388387028fcbe398f8df0a", "messageCount": 2, "turnStep": "0.1", "time": "<time>" }
+        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "0841d488bc8e0d298f206bfe6d41c4b893d11ac3046cff11b0f1a67a7b67186b", "messageCount": 2, "turnStep": "0.1", "time": "<time>" }
         [emit] agent.status.updated        { "time": "<time>", "agentId": "main", "usage": { "byModel": { "mock-model": { "inputOther": 144, "output": 16, "inputCacheRead": 0, "inputCacheCreation": 0 } }, "total": { "inputOther": 144, "output": 16, "inputCacheRead": 0, "inputCacheCreation": 0 }, "currentTurn": { "inputOther": 144, "output": 16, "inputCacheRead": 0, "inputCacheCreation": 0 } } }
         [emit] agent.status.updated        { "time": "<time>", "agentId": "main", "contextTokens": 160 }
         [wire] usage.record                { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 144, "output": 16, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
@@ -4357,7 +4357,7 @@ describe('Agent tools', () => {
         [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-4>", "turnId": "0", "step": 2 }, "time": "<time>" }
         [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "The lookup result is moon-result." }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 2, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
-        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "f8d3fb9a7c26c38032bc3a2beb2fb4ac327a0ecb91388387028fcbe398f8df0a", "messageCount": 4, "turnStep": "0.2", "time": "<time>" }
+        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "0841d488bc8e0d298f206bfe6d41c4b893d11ac3046cff11b0f1a67a7b67186b", "messageCount": 4, "turnStep": "0.2", "time": "<time>" }
         [wire] usage.record                { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 164, "output": 12, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
         [emit] agent.status.updated        { "time": "<time>", "agentId": "main", "usage": { "byModel": { "mock-model": { "inputOther": 308, "output": 28, "inputCacheRead": 0, "inputCacheCreation": 0 } }, "total": { "inputOther": 308, "output": 28, "inputCacheRead": 0, "inputCacheCreation": 0 }, "currentTurn": { "inputOther": 308, "output": 28, "inputCacheRead": 0, "inputCacheCreation": 0 } } }
         [wire] token_counting.measured     { "agentId": "main", "length": 5, "tokens": 176, "time": "<time>" }
@@ -4400,8 +4400,8 @@ describe('Agent tools', () => {
         [wire] context.append_loop_event      { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-6>", "turnId": "1", "step": 1 }, "time": "<time>" }
         [emit] assistant.delta                { "time": "<time>", "agentId": "main", "turnId": 1, "delta": "No lookup tool is available." }
         [emit] agent.activity.updated         { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 1, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
-        [wire] llm.tools_snapshot             { "agentId": "main", "hash": "0432af3f1410334f38bb79c6965876003ddf8ff088b07654736330b30ffe1791", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
-        [wire] llm.request                    { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "0432af3f1410334f38bb79c6965876003ddf8ff088b07654736330b30ffe1791", "messageCount": 6, "turnStep": "1.1", "time": "<time>" }
+        [wire] llm.tools_snapshot             { "agentId": "main", "hash": "6658fc0337183153a13433d81fa64064a006b4466b6051bebe186a78deb044df", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentSwarm", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentSwarm when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentSwarm launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentSwarm supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentSwarm\` is called, that call must be the only tool call in the response.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole swarm." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n- Run git-mutating commands such as \`git commit\`, \`git push\`, \`git reset\`, and \`git rebase\` only when the user asks for them.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults default to 100 matching paths. Use \`offset\` (default 0) and \`head_limit\` (default 100) to page through results. When more matches are available, the result gives the next offset; keep the other search arguments unchanged. Set \`head_limit=0\` to remove the match-count limit. Pages still stay within the character retention limit, including notices: when it is reached, only complete paths are returned, with the next offset for continuation. Large pages are saved to a file with a path for Read.\\n\\nEach call searches the current filesystem again; pagination is not a snapshot, and file changes can shift results between pages. To collect a large list, use \`head_limit=0\`, read any saved output, and follow continuation offsets if the character limit is reached. Search timeouts, traversal errors, and output capture limits can still produce partial results; the result reports these limits, and pagination cannot recover paths that were never collected. Narrow the search and retry when it is incomplete.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results and waste search time and context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\` unless you need a complete listing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "head_limit": { "description": "Maximum number of matching paths to return after offset. Defaults to 100. Pass 0 to remove the match-count limit. The character limit still applies: large pages are saved for Read, and a continuation offset is provided when more paths remain. Search time and output capture limits still apply.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of matching paths to skip. Defaults to 0. Each call searches the current filesystem again; changes can shift results between pages.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns text within \`max_chars\`, including line numbers and the status block, preferring complete lines. The configured default is 100000 characters; calls can request up to 500000. Characters use JavaScript string length, not UTF-8 bytes or tokens. Read results are not spilled or shortened again by the general tool-output limit.\\n- Omit \`n_lines\` to read toward the end of the file. There is no fixed line-count cap. When the task requires the full text of a large file, request a larger \`max_chars\`, up to 500000, in the first call.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. If the result is incomplete, copy the \`Next Read\` arguments in the status block to continue without gaps or overlaps. Do not answer from a partial page when the task requires the remaining content.\\n- If a single line cannot fit on its own page, Read returns a fragment and reports its column range. Continue on the same line with the supplied \`column_offset\`; do not insert a newline between fragments of one source line. A partial line still counts toward the remaining \`n_lines\` until its ending is returned.\\n- \`column_offset\` is a zero-based position in the first line's displayed text, excluding its line-number prefix. It is supported only for forward reads. Offsets past the line or inside a Unicode surrogate pair return an error. Continuation refers to the current file contents; start a new read if the file changed.\\n- Kimi Code agent event logs (\`wire.jsonl\` under the sessions directory) follow the same character budget; locate a record with Grep, read it with \`n_lines=1\`, and follow \`Next Read\` to retrieve every fragment of a long record.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and checked with strict decoding first. If malformed sequences are found, Read returns readable text with U+FFFD replacements and a lossy-decoding warning on every page; do not treat this view as exact original text. The status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative \`line_offset\` reads from the end of the file (for example, -100 reads the last 100 lines). If the requested tail range exceeds the character budget, the newest complete lines in that range are returned first; \`Next Read\` covers the omitted earlier range. If no complete line fits, Read reports this and supplies forward \`Next Read\` arguments for the entire unread range. Omit \`column_offset\` when using a negative \`line_offset\`.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content. It reports the actual returned range, total lines, effective character budget, whether the requested range is complete, and whether EOF was reached. The block is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file (for example, -100 reads the last 100 lines).", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -9007199254740991, "exclusiveMaximum": 0 } ] }, "column_offset": { "description": "Zero-based character offset within the first line of a forward read, excluding its line-number prefix. Uses JavaScript string length in the displayed text. Copy continuation arguments from the previous result to resume a long line.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "n_lines": { "description": "The number of lines to read. Omit to read toward the end of the file. Results are bounded by max_chars, with continuation arguments when the requested range is incomplete.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "max_chars": { "description": "Maximum characters in the returned text, including line numbers and status. Omit for the configured default; requests above the configured maximum are capped.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+        [wire] llm.request                    { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "6658fc0337183153a13433d81fa64064a006b4466b6051bebe186a78deb044df", "messageCount": 6, "turnStep": "1.1", "time": "<time>" }
         [emit] agent.status.updated           { "time": "<time>", "agentId": "main", "usage": { "byModel": { "mock-model": { "inputOther": 492, "output": 38, "inputCacheRead": 0, "inputCacheCreation": 0 } }, "total": { "inputOther": 492, "output": 38, "inputCacheRead": 0, "inputCacheCreation": 0 }, "currentTurn": { "inputOther": 184, "output": 10, "inputCacheRead": 0, "inputCacheCreation": 0 } } }
         [emit] agent.status.updated           { "time": "<time>", "agentId": "main", "contextTokens": 194 }
         [wire] usage.record                   { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 184, "output": 10, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }