diff --git a/docs/users/features/memory.md b/docs/users/features/memory.md
index 077082a2714..cdd7951c7d4 100644
--- a/docs/users/features/memory.md
+++ b/docs/users/features/memory.md
@@ -2,7 +2,7 @@
 
 Every Qwen Code session starts with a fresh context window. Two mechanisms carry knowledge across sessions so you don't have to re-explain yourself every time:
 
-- **QWEN.md** — instructions *you* write once and Qwen reads every session
+- **QWEN.md** — instructions _you_ write once and Qwen reads every session
 - **Auto-memory** — notes Qwen writes itself based on what it learns from you
 
 ---
@@ -24,9 +24,9 @@ Don't include things Qwen can figure out by reading your code. QWEN.md works bes
 
 ### Where to create QWEN.md
 
-| File | Who it applies to |
-|---|---|
-| `~/.qwen/QWEN.md` | You, across all your projects |
+| File                          | Who it applies to                             |
+| ----------------------------- | --------------------------------------------- |
+| `~/.qwen/QWEN.md`             | You, across all your projects                 |
 | `QWEN.md` in the project root | Your whole team (commit it to source control) |
 
 You can have both. Qwen loads all QWEN.md files it finds when you start a session — your personal one plus any in the project.
@@ -45,6 +45,7 @@ You can point QWEN.md at other files so Qwen reads them too:
 See @README.md for project overview.
 
 # Conventions
+
 - Git workflow: @docs/git-workflow.md
 ```
 
@@ -62,12 +63,12 @@ This is different from QWEN.md: you don't write it, Qwen does.
 
 Qwen looks for four kinds of things worth remembering:
 
-| What | Examples |
-|---|---|
-| **About you** | Your role, background, how you like to work |
-| **Your feedback** | Corrections you made, approaches you confirmed |
-| **Project context** | Ongoing work, decisions, goals not obvious from the code |
-| **External references** | Dashboards, ticket trackers, docs links you mentioned |
+| What                    | Examples                                                 |
+| ----------------------- | -------------------------------------------------------- |
+| **About you**           | Your role, background, how you like to work              |
+| **Your feedback**       | Corrections you made, approaches you confirmed           |
+| **Project context**     | Ongoing work, decisions, goals not obvious from the code |
+| **External references** | Dashboards, ticket trackers, docs links you mentioned    |
 
 Qwen doesn't save everything — only things that would actually be useful next time.
 
@@ -150,6 +151,7 @@ Runs the memory cleanup now instead of waiting for the automatic schedule:
 Open `/memory` to see which files are loaded. If your file isn't listed, Qwen can't see it — make sure it's in the project root or `~/.qwen/`.
 
 Instructions work better when they're specific:
+
 - ✓ `Use 2-space indentation for TypeScript files`
 - ✗ `Format code nicely`
 
diff --git a/packages/cli/src/ui/components/MemoryDialog.tsx b/packages/cli/src/ui/components/MemoryDialog.tsx
index c0e781e49c3..8fca1ecb080 100644
--- a/packages/cli/src/ui/components/MemoryDialog.tsx
+++ b/packages/cli/src/ui/components/MemoryDialog.tsx
@@ -393,7 +393,9 @@ export function MemoryDialog({ onClose }: MemoryDialogProps) {
                 {index + 1}. {item.label}
               </Text>
               {item.description ? (
-                <Text color={theme.text.secondary}>{`  ${item.description}`}</Text>
+                <Text
+                  color={theme.text.secondary}
+                >{`  ${item.description}`}</Text>
               ) : null}
             </Box>
           );
diff --git a/packages/cli/src/ui/hooks/useReactToolScheduler.ts b/packages/cli/src/ui/hooks/useReactToolScheduler.ts
index 3991823d55d..3bf3083797b 100644
--- a/packages/cli/src/ui/hooks/useReactToolScheduler.ts
+++ b/packages/cli/src/ui/hooks/useReactToolScheduler.ts
@@ -341,7 +341,9 @@ export function mapToDisplay(
   return {
     type: 'tool_group',
     tools: toolDisplays,
-    memoryWriteCount: toolDisplays.filter((t) => t.isMemoryOp === 'write').length || undefined,
-    memoryReadCount: toolDisplays.filter((t) => t.isMemoryOp === 'read').length || undefined,
+    memoryWriteCount:
+      toolDisplays.filter((t) => t.isMemoryOp === 'write').length || undefined,
+    memoryReadCount:
+      toolDisplays.filter((t) => t.isMemoryOp === 'read').length || undefined,
   };
 }
diff --git a/packages/cli/src/utils/commands.test.ts b/packages/cli/src/utils/commands.test.ts
index 07c5c9fa8f1..fb9410b386f 100644
--- a/packages/cli/src/utils/commands.test.ts
+++ b/packages/cli/src/utils/commands.test.ts
@@ -71,10 +71,7 @@ describe('parseSlashCommand', () => {
   });
 
   it('should parse a subcommand with arguments', () => {
-    const result = parseSlashCommand(
-      '/config set theme dark',
-      mockCommands,
-    );
+    const result = parseSlashCommand('/config set theme dark', mockCommands);
     expect(result.commandToExecute?.name).toBe('set');
     expect(result.args).toBe('theme dark');
     expect(result.canonicalPath).toEqual(['config', 'set']);
diff --git a/packages/core/src/agents/runtime/agent-core.ts b/packages/core/src/agents/runtime/agent-core.ts
index b4b73ee70ef..97627b7e613 100644
--- a/packages/core/src/agents/runtime/agent-core.ts
+++ b/packages/core/src/agents/runtime/agent-core.ts
@@ -110,29 +110,9 @@ export interface CreateChatOptions {
   /**
    * Optional conversation history from a parent session. When provided,
    * this history is prepended to the chat so the agent has prior
-   * conversational context (e.g., from the main session that spawned it).
+   * conversational context (e.g., from AgentInteractive.start()).
    */
   extraHistory?: Content[];
-  /**
-   * When provided, replaces the auto-built generationConfig
-   * (systemInstruction, temperature, etc.) with this exact config.
-   * Used by fork subagents to share the parent conversation's cache
-   * prefix for DashScope prompt caching.
-   */
-  generationConfigOverride?: GenerateContentConfig & {
-    systemInstruction?: string | Content;
-  };
-  /**
-   * When true, skip injecting the env bootstrap messages from
-   * `getInitialChatHistory()`. Set by fork subagents because their
-   * `extraHistory` is the full parent history that already contains
-   * those env messages — re-injecting would duplicate them.
-   *
-   * Other callers (e.g. arena interactive agents) pass an
-   * env-stripped history and DO need fresh env init for their own
-   * working directory, so they must leave this unset.
-   */
-  skipEnvHistory?: boolean;
 }
 
 /**
@@ -243,21 +223,31 @@ export class AgentCore {
     context: ContextState,
     options?: CreateChatOptions,
   ): Promise<GeminiChat | undefined> {
-    if (!this.promptConfig.systemPrompt && !this.promptConfig.initialMessages) {
+    if (
+      !this.promptConfig.systemPrompt &&
+      !this.promptConfig.renderedSystemPrompt &&
+      !this.promptConfig.initialMessages
+    ) {
       throw new Error(
-        'PromptConfig must have either `systemPrompt` or `initialMessages` defined.',
+        'PromptConfig must have `systemPrompt`, `renderedSystemPrompt`, or `initialMessages` defined.',
       );
     }
-    if (this.promptConfig.systemPrompt && this.promptConfig.initialMessages) {
+    if (
+      this.promptConfig.systemPrompt &&
+      this.promptConfig.renderedSystemPrompt
+    ) {
       throw new Error(
-        'PromptConfig cannot have both `systemPrompt` and `initialMessages` defined.',
+        'PromptConfig cannot have both `systemPrompt` and `renderedSystemPrompt` defined.',
       );
     }
 
-    // Skip env bootstrap when the caller (fork) explicitly says its
-    // extraHistory already contains those messages. Other callers that
-    // provide an env-stripped history (e.g. arena) still get fresh env init.
-    const envHistory = options?.skipEnvHistory
+    // When initialMessages is set, the caller owns the full prior history
+    // (including any env bootstrap it wants). Fork relies on this to inherit
+    // the parent conversation verbatim without duplicating env messages.
+    const hasInitialMessages =
+      !!this.promptConfig.initialMessages &&
+      this.promptConfig.initialMessages.length > 0;
+    const envHistory = hasInitialMessages
       ? []
       : await getInitialChatHistory(this.runtimeContext);
 
@@ -267,24 +257,19 @@ export class AgentCore {
       ...(this.promptConfig.initialMessages ?? []),
     ];
 
-    // If an override is provided (fork path), use it directly for cache
-    // sharing. Otherwise, build the config from this agent's promptConfig.
-    // Note: buildChatSystemPrompt is called OUTSIDE the try/catch so template
-    // errors propagate to the caller (not swallowed by reportError).
-    let generationConfig: GenerateContentConfig & {
+    // Build generationConfig. For fork subagents, `renderedSystemPrompt`
+    // carries the parent's exact rendered systemInstruction so the fork
+    // shares a byte-identical cache prefix. Otherwise, template
+    // `systemPrompt` via buildChatSystemPrompt (which may throw — kept
+    // outside the try/catch so template errors surface to the caller).
+    const generationConfig: GenerateContentConfig & {
       systemInstruction?: string | Content;
-    };
-
-    if (options?.generationConfigOverride) {
-      generationConfig = options.generationConfigOverride;
-    } else {
-      const systemInstruction = this.promptConfig.systemPrompt
-        ? this.buildChatSystemPrompt(context, options)
-        : undefined;
-      generationConfig = {
-        temperature: this.modelConfig.temp,
-        topP: this.modelConfig.top_p,
-      };
+    } = {};
+    if (this.promptConfig.renderedSystemPrompt !== undefined) {
+      generationConfig.systemInstruction =
+        this.promptConfig.renderedSystemPrompt;
+    } else if (this.promptConfig.systemPrompt) {
+      const systemInstruction = this.buildChatSystemPrompt(context, options);
       if (systemInstruction) {
         generationConfig.systemInstruction = systemInstruction;
       }
@@ -330,7 +315,10 @@ export class AgentCore {
         (t): t is FunctionDeclaration => typeof t !== 'string',
       );
 
-      if (hasWildcard || asStrings.length === 0) {
+      if (
+        hasWildcard ||
+        (asStrings.length === 0 && onlyInlineDecls.length === 0)
+      ) {
         toolsList.push(
           ...toolRegistry
             .getFunctionDeclarations()
diff --git a/packages/core/src/agents/runtime/agent-headless.test.ts b/packages/core/src/agents/runtime/agent-headless.test.ts
index 01ff1b040e2..89760ee0bb1 100644
--- a/packages/core/src/agents/runtime/agent-headless.test.ts
+++ b/packages/core/src/agents/runtime/agent-headless.test.ts
@@ -231,8 +231,6 @@ describe('subagent.ts', () => {
 
     const defaultModelConfig: ModelConfig = {
       model: 'qwen3-coder-plus',
-      temp: 0.5, // Specific temp to test override
-      top_p: 1,
     };
 
     const defaultRunConfig: RunConfig = {
@@ -439,8 +437,6 @@ describe('subagent.ts', () => {
         // Check Generation Config
         const generationConfig = getGenerationConfigFromMock();
 
-        // Check temperature override
-        expect(generationConfig.temperature).toBe(defaultModelConfig.temp);
         expect(generationConfig.systemInstruction).toContain(
           'Hello Agent, your task is Testing.',
         );
@@ -556,15 +552,20 @@ describe('subagent.ts', () => {
         expect(sysPrompt).not.toContain('---');
       });
 
-      it('should use initialMessages instead of systemPrompt if provided', async () => {
+      it('should replace env history with initialMessages when both initialMessages and systemPrompt are set', async () => {
         const { config } = await createMockConfig();
         vi.mocked(GeminiChat).mockClear();
 
         const initialMessages: Content[] = [
-          { role: 'user', parts: [{ text: 'Hi' }] },
+          { role: 'user', parts: [{ text: 'prior user turn' }] },
+          { role: 'model', parts: [{ text: 'prior model turn' }] },
         ];
-        const promptConfig: PromptConfig = { initialMessages };
+        const promptConfig: PromptConfig = {
+          systemPrompt: 'System ${name}.',
+          initialMessages,
+        };
         const context = new ContextState();
+        context.set('name', 'Agent');
 
         // Model stops immediately
         mockSendMessageStream.mockImplementation(createMockStream(['stop']));
@@ -583,15 +584,44 @@ describe('subagent.ts', () => {
         const generationConfig = getGenerationConfigFromMock();
         const history = callArgs[2];
 
-        expect(generationConfig.systemInstruction).toBeUndefined();
-        expect(history).toEqual([
-          { role: 'user', parts: [{ text: 'Env Context' }] },
-          {
-            role: 'model',
-            parts: [{ text: 'Got it. Thanks for the context!' }],
-          },
-          ...initialMessages,
-        ]);
+        // systemPrompt is templated normally.
+        expect(generationConfig.systemInstruction).toContain('System Agent.');
+        expect(generationConfig.systemInstruction).toContain(
+          'Important Rules:',
+        );
+        // Env bootstrap is skipped; history is exactly initialMessages.
+        expect(history).toEqual(initialMessages);
+      });
+
+      it('should use renderedSystemPrompt verbatim and bypass templating', async () => {
+        const { config } = await createMockConfig();
+        vi.mocked(GeminiChat).mockClear();
+
+        const rendered = 'Verbatim parent system prompt ${name}';
+        const promptConfig: PromptConfig = {
+          renderedSystemPrompt: rendered,
+          initialMessages: [
+            { role: 'user', parts: [{ text: 'hi' }] },
+            { role: 'model', parts: [{ text: 'ok' }] },
+          ],
+        };
+        const context = new ContextState();
+
+        mockSendMessageStream.mockImplementation(createMockStream(['stop']));
+
+        const scope = await AgentHeadless.create(
+          'test-agent',
+          config,
+          promptConfig,
+          defaultModelConfig,
+          defaultRunConfig,
+        );
+
+        await scope.execute(context);
+
+        const generationConfig = getGenerationConfigFromMock();
+        // No ${name} substitution and no non-interactive rules appended.
+        expect(generationConfig.systemInstruction).toBe(rendered);
       });
 
       it('should throw an error if template variables are missing', async () => {
@@ -618,11 +648,11 @@ describe('subagent.ts', () => {
         expect(scope.getTerminateMode()).toBe(AgentTerminateMode.ERROR);
       });
 
-      it('should validate that systemPrompt and initialMessages are mutually exclusive', async () => {
+      it('should validate that systemPrompt and renderedSystemPrompt are mutually exclusive', async () => {
         const { config } = await createMockConfig();
         const promptConfig: PromptConfig = {
           systemPrompt: 'System',
-          initialMessages: [{ role: 'user', parts: [{ text: 'Hi' }] }],
+          renderedSystemPrompt: 'Rendered',
         };
         const context = new ContextState();
 
@@ -635,7 +665,7 @@ describe('subagent.ts', () => {
         );
 
         await expect(agent.execute(context)).rejects.toThrow(
-          'PromptConfig cannot have both `systemPrompt` and `initialMessages` defined.',
+          'PromptConfig cannot have both `systemPrompt` and `renderedSystemPrompt` defined.',
         );
         expect(agent.getTerminateMode()).toBe(AgentTerminateMode.ERROR);
       });
diff --git a/packages/core/src/agents/runtime/agent-headless.ts b/packages/core/src/agents/runtime/agent-headless.ts
index 55fb7ba2412..ac02f80dfba 100644
--- a/packages/core/src/agents/runtime/agent-headless.ts
+++ b/packages/core/src/agents/runtime/agent-headless.ts
@@ -192,21 +192,8 @@ export class AgentHeadless {
   async execute(
     context: ContextState,
     externalSignal?: AbortSignal,
-    options?: {
-      extraHistory?: Array<import('@google/genai').Content>;
-      /** Override generationConfig for cache sharing (fork subagent). */
-      generationConfigOverride?: import('@google/genai').GenerateContentConfig;
-      /** Override tool declarations for cache sharing (fork subagent). */
-      toolsOverride?: Array<import('@google/genai').FunctionDeclaration>;
-      /** Skip env bootstrap injection (fork already inherits parent env). */
-      skipEnvHistory?: boolean;
-    },
   ): Promise<void> {
-    const chat = await this.core.createChat(context, {
-      extraHistory: options?.extraHistory,
-      generationConfigOverride: options?.generationConfigOverride,
-      skipEnvHistory: options?.skipEnvHistory,
-    });
+    const chat = await this.core.createChat(context);
 
     if (!chat) {
       this.terminateMode = AgentTerminateMode.ERROR;
@@ -225,7 +212,7 @@ export class AgentHeadless {
       abortController.abort();
     }
 
-    const toolsList = options?.toolsOverride ?? this.core.prepareTools();
+    const toolsList = this.core.prepareTools();
 
     const initialTaskText = String(
       (context.get('task_prompt') as string) ?? 'Get Started!',
diff --git a/packages/core/src/agents/runtime/agent-types.ts b/packages/core/src/agents/runtime/agent-types.ts
index 1f6f15343c9..e6c4e4e40dc 100644
--- a/packages/core/src/agents/runtime/agent-types.ts
+++ b/packages/core/src/agents/runtime/agent-types.ts
@@ -21,13 +21,23 @@ import type { Content, FunctionDeclaration } from '@google/genai';
 export interface PromptConfig {
   /**
    * A single system prompt string that defines the agent's persona and instructions.
-   * Note: You should use either `systemPrompt` or `initialMessages`, but not both.
+   * Templated via ${var} substitution, optionally suffixed with non-interactive
+   * rules and user memory. Mutually exclusive with `renderedSystemPrompt`.
    */
   systemPrompt?: string;
 
   /**
-   * An array of user/model content pairs to seed the chat history for few-shot prompting.
-   * Note: You should use either `systemPrompt` or `initialMessages`, but not both.
+   * A pre-rendered system instruction consumed verbatim — no templating, no
+   * non-interactive suffix, no user-memory injection. Used by fork subagents
+   * to share the parent conversation's exact cache prefix. Mutually exclusive
+   * with `systemPrompt`.
+   */
+  renderedSystemPrompt?: string | Content;
+
+  /**
+   * Seed chat history. When set, fully replaces the default env bootstrap
+   * (the caller owns the full prior context, e.g. fork inheriting parent
+   * history). Can coexist with `systemPrompt` / `renderedSystemPrompt`.
    */
   initialMessages?: Content[];
 }
@@ -42,10 +52,6 @@ export interface ModelConfig {
    * TODO: In the future, this needs to support 'auto' or some other string to support routing use cases.
    */
   model?: string;
-  /** The temperature for the model's sampling process. */
-  temp?: number;
-  /** The top-p value for nucleus sampling. */
-  top_p?: number;
 }
 
 /**
diff --git a/packages/core/src/config/config.ts b/packages/core/src/config/config.ts
index a86c8e0d71b..5c05935014f 100644
--- a/packages/core/src/config/config.ts
+++ b/packages/core/src/config/config.ts
@@ -58,7 +58,7 @@ import { canUseRipgrep } from '../utils/ripgrepUtils.js';
 import { RipGrepTool } from '../tools/ripGrep.js';
 import { ShellTool } from '../tools/shell.js';
 import { SkillTool } from '../tools/skill.js';
-import { AgentTool } from '../tools/agent.js';
+import { AgentTool } from '../tools/agent/agent.js';
 import { TodoWriteTool } from '../tools/todoWrite.js';
 import { ToolRegistry } from '../tools/tool-registry.js';
 import { WebFetchTool } from '../tools/web-fetch.js';
diff --git a/packages/core/src/core/client.ts b/packages/core/src/core/client.ts
index b8a23b0d339..0a7cb78502c 100644
--- a/packages/core/src/core/client.ts
+++ b/packages/core/src/core/client.ts
@@ -47,7 +47,7 @@ import {
 import { LoopDetectionService } from '../services/loopDetectionService.js';
 
 // Tools
-import { AgentTool } from '../tools/agent.js';
+import { AgentTool } from '../tools/agent/agent.js';
 import type { RelevantAutoMemoryPromptResult } from '../memory/manager.js';
 
 // Telemetry
diff --git a/packages/core/src/index.ts b/packages/core/src/index.ts
index 51593bb9eaf..248dfedffb2 100644
--- a/packages/core/src/index.ts
+++ b/packages/core/src/index.ts
@@ -90,7 +90,7 @@ export * from './tools/ripGrep.js';
 export * from './tools/sdk-control-client-transport.js';
 export * from './tools/shell.js';
 export * from './tools/skill.js';
-export * from './tools/agent.js';
+export * from './tools/agent/agent.js';
 export * from './tools/todoWrite.js';
 export * from './tools/tool-error.js';
 export * from './tools/tool-registry.js';
diff --git a/packages/core/src/memory/const.test.ts b/packages/core/src/memory/const.test.ts
index 3f6f449b88e..d20d7a509fa 100644
--- a/packages/core/src/memory/const.test.ts
+++ b/packages/core/src/memory/const.test.ts
@@ -23,27 +23,26 @@ vi.mock(import('node:fs/promises'), async (importOriginal) => {
 
 vi.mock('os');
 
-
 describe('setGeminiMdFilename', () => {
-    it('should update currentGeminiMdFilename when a valid new name is provided', () => {
-      const newName = 'CUSTOM_CONTEXT.md';
-      setGeminiMdFilename(newName);
-      expect(getCurrentGeminiMdFilename()).toBe(newName);
-    });
+  it('should update currentGeminiMdFilename when a valid new name is provided', () => {
+    const newName = 'CUSTOM_CONTEXT.md';
+    setGeminiMdFilename(newName);
+    expect(getCurrentGeminiMdFilename()).toBe(newName);
+  });
 
-    it('should not update currentGeminiMdFilename if the new name is empty or whitespace', () => {
-      const initialName = getCurrentGeminiMdFilename(); // Get current before trying to change
-      setGeminiMdFilename('  ');
-      expect(getCurrentGeminiMdFilename()).toBe(initialName);
+  it('should not update currentGeminiMdFilename if the new name is empty or whitespace', () => {
+    const initialName = getCurrentGeminiMdFilename(); // Get current before trying to change
+    setGeminiMdFilename('  ');
+    expect(getCurrentGeminiMdFilename()).toBe(initialName);
 
-      setGeminiMdFilename('');
-      expect(getCurrentGeminiMdFilename()).toBe(initialName);
-    });
+    setGeminiMdFilename('');
+    expect(getCurrentGeminiMdFilename()).toBe(initialName);
+  });
 
-    it('should handle an array of filenames', () => {
-      const newNames = ['CUSTOM_CONTEXT.md', 'ANOTHER_CONTEXT.md'];
-      setGeminiMdFilename(newNames);
-      expect(getCurrentGeminiMdFilename()).toBe('CUSTOM_CONTEXT.md');
-      expect(getAllGeminiMdFilenames()).toEqual(newNames);
-    });
+  it('should handle an array of filenames', () => {
+    const newNames = ['CUSTOM_CONTEXT.md', 'ANOTHER_CONTEXT.md'];
+    setGeminiMdFilename(newNames);
+    expect(getCurrentGeminiMdFilename()).toBe('CUSTOM_CONTEXT.md');
+    expect(getAllGeminiMdFilenames()).toEqual(newNames);
   });
+});
diff --git a/packages/core/src/memory/extractionAgentPlanner.ts b/packages/core/src/memory/extractionAgentPlanner.ts
index 058682e0624..6a89127986c 100644
--- a/packages/core/src/memory/extractionAgentPlanner.ts
+++ b/packages/core/src/memory/extractionAgentPlanner.ts
@@ -6,7 +6,7 @@
 
 import type { Config } from '../config/config.js';
 import { runForkedAgent, getCacheSafeParams } from '../utils/forkedAgent.js';
-import { buildFunctionResponseParts } from '../agents/runtime/forkSubagent.js';
+import { buildFunctionResponseParts } from '../tools/agent/fork-subagent.js';
 import type { Content } from '@google/genai';
 import type { PermissionManager } from '../permissions/permission-manager.js';
 import type {
@@ -324,7 +324,6 @@ export async function runAutoMemoryExtractionByAgent(
       ToolNames.EDIT,
     ],
     extraHistory,
-    skipEnvHistory: true,
   });
 
   if (result.status !== 'completed') {
diff --git a/packages/core/src/memory/indexer.test.ts b/packages/core/src/memory/indexer.test.ts
index 707b171cb26..9becaf34695 100644
--- a/packages/core/src/memory/indexer.test.ts
+++ b/packages/core/src/memory/indexer.test.ts
@@ -23,7 +23,10 @@ describe('managed auto-memory indexer', () => {
     tempDir = await fs.mkdtemp(path.join(os.tmpdir(), 'auto-memory-indexer-'));
     projectRoot = path.join(tempDir, 'project');
     await fs.mkdir(projectRoot, { recursive: true });
-    await ensureAutoMemoryScaffold(projectRoot, new Date('2026-04-01T00:00:00.000Z'));
+    await ensureAutoMemoryScaffold(
+      projectRoot,
+      new Date('2026-04-01T00:00:00.000Z'),
+    );
   });
 
   afterEach(async () => {
@@ -49,9 +52,7 @@ describe('managed auto-memory indexer', () => {
       },
     ]);
 
-    expect(content).toBe(
-      '- [User Memory](user/terse.md) — User profile',
-    );
+    expect(content).toBe('- [User Memory](user/terse.md) — User profile');
   });
 
   it('rewrites MEMORY.md from topic file contents', async () => {
@@ -76,7 +77,10 @@ describe('managed auto-memory indexer', () => {
 
     await rebuildManagedAutoMemoryIndex(projectRoot);
 
-    const index = await fs.readFile(getAutoMemoryIndexPath(projectRoot), 'utf-8');
+    const index = await fs.readFile(
+      getAutoMemoryIndexPath(projectRoot),
+      'utf-8',
+    );
     expect(index).toContain('[Project Memory](project/repo-workspaces.md)');
     expect(index).toContain('The repo uses pnpm workspaces.');
   });
diff --git a/packages/core/src/memory/indexer.ts b/packages/core/src/memory/indexer.ts
index 3b7cde10b9b..a4297d2dada 100644
--- a/packages/core/src/memory/indexer.ts
+++ b/packages/core/src/memory/indexer.ts
@@ -6,7 +6,10 @@
 
 import * as fs from 'node:fs/promises';
 import { getAutoMemoryIndexPath, getAutoMemoryMetadataPath } from './paths.js';
-import { scanAutoMemoryTopicDocuments, type ScannedAutoMemoryDocument } from './scan.js';
+import {
+  scanAutoMemoryTopicDocuments,
+  type ScannedAutoMemoryDocument,
+} from './scan.js';
 import type { AutoMemoryMetadata } from './types.js';
 
 const MAX_INDEX_LINE_CHARS = 150;
@@ -22,7 +25,10 @@ function truncateIndexLine(text: string): string {
 
 export function buildManagedAutoMemoryIndex(
   docs: ScannedAutoMemoryDocument[],
-  _metadata?: Pick<AutoMemoryMetadata, 'updatedAt' | 'lastDreamAt' | 'lastDreamSessionId'>,
+  _metadata?: Pick<
+    AutoMemoryMetadata,
+    'updatedAt' | 'lastDreamAt' | 'lastDreamSessionId'
+  >,
 ): string {
   const raw = docs
     .map((doc) =>
@@ -34,7 +40,9 @@ export function buildManagedAutoMemoryIndex(
 
   const lines = raw.split('\n');
   const wasLineTruncated = lines.length > MAX_INDEX_LINES;
-  let truncated = wasLineTruncated ? lines.slice(0, MAX_INDEX_LINES).join('\n') : raw;
+  let truncated = wasLineTruncated
+    ? lines.slice(0, MAX_INDEX_LINES).join('\n')
+    : raw;
 
   if (truncated.length > MAX_INDEX_BYTES) {
     const cutAt = truncated.lastIndexOf('\n', MAX_INDEX_BYTES);
@@ -52,7 +60,10 @@ async function readAutoMemoryMetadata(
   projectRoot: string,
 ): Promise<AutoMemoryMetadata | undefined> {
   try {
-    const content = await fs.readFile(getAutoMemoryMetadataPath(projectRoot), 'utf-8');
+    const content = await fs.readFile(
+      getAutoMemoryMetadataPath(projectRoot),
+      'utf-8',
+    );
     return JSON.parse(content) as AutoMemoryMetadata;
   } catch {
     return undefined;
diff --git a/packages/core/src/memory/prompt.ts b/packages/core/src/memory/prompt.ts
index c584dafed8a..615aa25ddb6 100644
--- a/packages/core/src/memory/prompt.ts
+++ b/packages/core/src/memory/prompt.ts
@@ -143,7 +143,10 @@ function truncateManagedAutoMemoryIndex(indexContent: string): string {
     : trimmed;
 
   if (truncated.length > MAX_MANAGED_AUTO_MEMORY_INDEX_BYTES) {
-    const cutAt = truncated.lastIndexOf('\n', MAX_MANAGED_AUTO_MEMORY_INDEX_BYTES);
+    const cutAt = truncated.lastIndexOf(
+      '\n',
+      MAX_MANAGED_AUTO_MEMORY_INDEX_BYTES,
+    );
     truncated = truncated.slice(
       0,
       cutAt > 0 ? cutAt : MAX_MANAGED_AUTO_MEMORY_INDEX_BYTES,
@@ -231,6 +234,4 @@ export function appendManagedAutoMemoryToUserMemory(
   return `${trimmedUserMemory}\n\n---\n\n${managedPrompt}`;
 }
 
-export {
-  MAX_MANAGED_AUTO_MEMORY_INDEX_LINES,
-};
\ No newline at end of file
+export { MAX_MANAGED_AUTO_MEMORY_INDEX_LINES };
diff --git a/packages/core/src/memory/recall.test.ts b/packages/core/src/memory/recall.test.ts
index 05d0750dc14..e6ff91fe8c2 100644
--- a/packages/core/src/memory/recall.test.ts
+++ b/packages/core/src/memory/recall.test.ts
@@ -129,4 +129,4 @@ describe('auto-memory relevant recall', () => {
       '/tmp/user.md',
     );
   });
-});
\ No newline at end of file
+});
diff --git a/packages/core/src/memory/store.test.ts b/packages/core/src/memory/store.test.ts
index 3838f82cbec..0d02cf865d1 100644
--- a/packages/core/src/memory/store.test.ts
+++ b/packages/core/src/memory/store.test.ts
@@ -67,7 +67,10 @@ describe('auto-memory storage scaffold', () => {
     const now = new Date('2026-04-01T08:00:00.000Z');
     await ensureAutoMemoryScaffold(projectRoot, now);
 
-    const index = await fs.readFile(getAutoMemoryIndexPath(projectRoot), 'utf-8');
+    const index = await fs.readFile(
+      getAutoMemoryIndexPath(projectRoot),
+      'utf-8',
+    );
     expect(index).toBe(createDefaultAutoMemoryIndex());
 
     const metadata = JSON.parse(
@@ -82,20 +85,34 @@ describe('auto-memory storage scaffold', () => {
       updatedAt: '2026-04-01T08:00:00.000Z',
     });
 
-    await expect(fs.stat(getAutoMemoryRoot(projectRoot))).resolves.toBeDefined();
-    await expect(fs.access(getAutoMemoryTopicPath(projectRoot, 'user'))).rejects.toThrow();
+    await expect(
+      fs.stat(getAutoMemoryRoot(projectRoot)),
+    ).resolves.toBeDefined();
+    await expect(
+      fs.access(getAutoMemoryTopicPath(projectRoot, 'user')),
+    ).rejects.toThrow();
   });
 
   it('is idempotent and preserves existing index content', async () => {
-    await ensureAutoMemoryScaffold(projectRoot, new Date('2026-04-01T08:00:00.000Z'));
+    await ensureAutoMemoryScaffold(
+      projectRoot,
+      new Date('2026-04-01T08:00:00.000Z'),
+    );
     const customIndex = '# Existing Index\n\n- keep me\n';
-    await fs.writeFile(getAutoMemoryIndexPath(projectRoot), customIndex, 'utf-8');
-
-    await ensureAutoMemoryScaffold(projectRoot, new Date('2026-04-02T08:00:00.000Z'));
-
-    await expect(fs.readFile(getAutoMemoryIndexPath(projectRoot), 'utf-8')).resolves.toBe(
+    await fs.writeFile(
+      getAutoMemoryIndexPath(projectRoot),
       customIndex,
+      'utf-8',
     );
+
+    await ensureAutoMemoryScaffold(
+      projectRoot,
+      new Date('2026-04-02T08:00:00.000Z'),
+    );
+
+    await expect(
+      fs.readFile(getAutoMemoryIndexPath(projectRoot), 'utf-8'),
+    ).resolves.toBe(customIndex);
   });
 
   it('returns null when the auto-memory index does not exist yet', async () => {
@@ -106,4 +123,4 @@ describe('auto-memory storage scaffold', () => {
     await ensureAutoMemoryScaffold(projectRoot);
     await expect(readAutoMemoryIndex(projectRoot)).resolves.toBe('');
   });
-});
\ No newline at end of file
+});
diff --git a/packages/core/src/memory/store.ts b/packages/core/src/memory/store.ts
index eb966d6744a..9f5924a3ee8 100644
--- a/packages/core/src/memory/store.ts
+++ b/packages/core/src/memory/store.ts
@@ -93,6 +93,4 @@ export async function readAutoMemoryIndex(
   }
 }
 
-export {
-  AUTO_MEMORY_INDEX_FILENAME,
-};
\ No newline at end of file
+export { AUTO_MEMORY_INDEX_FILENAME };
diff --git a/packages/core/src/memory/types.ts b/packages/core/src/memory/types.ts
index 6e05930f6e0..1c46d3af7cd 100644
--- a/packages/core/src/memory/types.ts
+++ b/packages/core/src/memory/types.ts
@@ -40,4 +40,4 @@ export interface AutoMemoryExtractCursor {
   sessionId?: string;
   processedOffset?: number;
   updatedAt: string;
-}
\ No newline at end of file
+}
diff --git a/packages/core/src/telemetry/types.ts b/packages/core/src/telemetry/types.ts
index 1bfd09ced9f..9cef68ce49a 100644
--- a/packages/core/src/telemetry/types.ts
+++ b/packages/core/src/telemetry/types.ts
@@ -20,7 +20,7 @@ export { ToolCallDecision };
 import type { OutputFormat } from '../output/types.js';
 import { ToolNames } from '../tools/tool-names.js';
 import type { SkillTool } from '../tools/skill.js';
-import type { AgentTool } from '../tools/agent.js';
+import type { AgentTool } from '../tools/agent/agent.js';
 
 export interface BaseTelemetryEvent {
   'event.name': string;
diff --git a/packages/core/src/tools/agent.test.ts b/packages/core/src/tools/agent/agent.test.ts
similarity index 91%
rename from packages/core/src/tools/agent.test.ts
rename to packages/core/src/tools/agent/agent.test.ts
index 1a203cebf48..2a1b603f636 100644
--- a/packages/core/src/tools/agent.test.ts
+++ b/packages/core/src/tools/agent/agent.test.ts
@@ -11,26 +11,26 @@ import {
   resolveSubagentApprovalMode,
 } from './agent.js';
 import type { PartListUnion } from '@google/genai';
-import type { ToolResultDisplay, AgentResultDisplay } from './tools.js';
-import { ToolConfirmationOutcome } from './tools.js';
-import { type Config, ApprovalMode } from '../config/config.js';
-import { SubagentManager } from '../subagents/subagent-manager.js';
-import type { SubagentConfig } from '../subagents/types.js';
-import { AgentTerminateMode } from '../agents/runtime/agent-types.js';
+import type { ToolResultDisplay, AgentResultDisplay } from '../tools.js';
+import { ToolConfirmationOutcome } from '../tools.js';
+import { type Config, ApprovalMode } from '../../config/config.js';
+import { SubagentManager } from '../../subagents/subagent-manager.js';
+import type { SubagentConfig } from '../../subagents/types.js';
+import { AgentTerminateMode } from '../../agents/runtime/agent-types.js';
 import {
-  type AgentHeadless,
+  AgentHeadless,
   ContextState,
-} from '../agents/runtime/agent-headless.js';
-import { AgentEventType } from '../agents/runtime/agent-events.js';
+} from '../../agents/runtime/agent-headless.js';
+import { AgentEventType } from '../../agents/runtime/agent-events.js';
 import type {
   AgentToolCallEvent,
   AgentToolResultEvent,
   AgentApprovalRequestEvent,
   AgentEventEmitter,
-} from '../agents/runtime/agent-events.js';
-import { partToString } from '../utils/partUtils.js';
-import type { HookSystem } from '../hooks/hookSystem.js';
-import { PermissionMode } from '../hooks/types.js';
+} from '../../agents/runtime/agent-events.js';
+import { partToString } from '../../utils/partUtils.js';
+import type { HookSystem } from '../../hooks/hookSystem.js';
+import { PermissionMode } from '../../hooks/types.js';
 
 // Type for accessing protected methods in tests
 type AgentToolInvocation = {
@@ -50,8 +50,8 @@ type AgentToolWithProtectedMethods = AgentTool & {
 };
 
 // Mock dependencies
-vi.mock('../subagents/subagent-manager.js');
-vi.mock('../agents/runtime/agent-headless.js');
+vi.mock('../../subagents/subagent-manager.js');
+vi.mock('../../agents/runtime/agent-headless.js');
 
 const MockedSubagentManager = vi.mocked(SubagentManager);
 const MockedContextState = vi.mocked(ContextState);
@@ -404,12 +404,6 @@ describe('AgentTool', () => {
       expect(mockAgent.execute).toHaveBeenCalledWith(
         mockContextState,
         undefined, // signal parameter (undefined when not provided)
-        {
-          extraHistory: undefined,
-          generationConfigOverride: undefined,
-          toolsOverride: undefined,
-          skipEnvHistory: false,
-        },
       );
 
       const llmText = partToString(result.llmContent);
@@ -562,6 +556,99 @@ describe('AgentTool', () => {
     });
   });
 
+  describe('Fork dispatch (subagent_type omitted)', () => {
+    let mockAgent: AgentHeadless;
+    let mockContextState: ContextState;
+
+    beforeEach(() => {
+      mockAgent = {
+        execute: vi.fn().mockResolvedValue(undefined),
+        getFinalText: vi.fn().mockReturnValue(''),
+        getExecutionSummary: vi.fn().mockReturnValue({
+          rounds: 0,
+          totalDurationMs: 0,
+          totalToolCalls: 0,
+          successfulToolCalls: 0,
+          failedToolCalls: 0,
+          successRate: 0,
+          inputTokens: 0,
+          outputTokens: 0,
+          totalTokens: 0,
+          estimatedCost: 0,
+          toolUsage: [],
+        }),
+        getStatistics: vi.fn().mockReturnValue({
+          rounds: 0,
+          totalDurationMs: 0,
+          totalToolCalls: 0,
+          successfulToolCalls: 0,
+          failedToolCalls: 0,
+        }),
+        getTerminateMode: vi.fn().mockReturnValue(AgentTerminateMode.GOAL),
+      } as unknown as AgentHeadless;
+
+      mockContextState = {
+        set: vi.fn(),
+      } as unknown as ContextState;
+
+      MockedContextState.mockImplementation(() => mockContextState);
+
+      // Parent conversation history: empty (first-turn fork — falls back to
+      // the fork agent's own systemPrompt + wildcard tools because no
+      // cache params have been captured yet).
+      vi.mocked(config.getGeminiClient).mockReturnValue({
+        getHistory: vi.fn().mockReturnValue([]),
+        getChat: vi.fn().mockReturnValue({
+          getGenerationConfig: vi.fn().mockReturnValue({}),
+        }),
+      } as unknown as ReturnType<Config['getGeminiClient']>);
+
+      vi.mocked(AgentHeadless.create).mockResolvedValue(mockAgent);
+    });
+
+    it('should call AgentHeadless.create directly and execute without options', async () => {
+      const params: AgentParams = {
+        description: 'fork task',
+        prompt: 'do the thing',
+      };
+
+      const invocation = (
+        agentTool as AgentToolWithProtectedMethods
+      ).createInvocation(params);
+      const result = await invocation.execute();
+
+      // Fork path: AgentHeadless.create invoked directly, bypassing
+      // SubagentManager.createAgentHeadless.
+      expect(AgentHeadless.create).toHaveBeenCalledTimes(1);
+      expect(mockSubagentManager.createAgentHeadless).not.toHaveBeenCalled();
+
+      const createArgs = vi.mocked(AgentHeadless.create).mock.calls[0];
+      expect(createArgs[0]).toBe('fork'); // name
+      // First-turn fork (no cache params): systemPrompt path, no
+      // renderedSystemPrompt. initialMessages is undefined (empty history).
+      const promptConfig = createArgs[2];
+      expect(promptConfig.renderedSystemPrompt).toBeUndefined();
+      expect(promptConfig.systemPrompt).toBeDefined();
+      // ToolConfig inherits wildcard for first-turn fallback.
+      const toolConfig = createArgs[5];
+      expect(toolConfig?.tools).toEqual(['*']);
+
+      // Fork returns the placeholder synchronously.
+      const llmText = partToString(result.llmContent);
+      expect(llmText).toBe('Fork started — processing in background');
+
+      // Drain the background executeSubagent() promise so its assertions
+      // become visible before the test ends.
+      await vi.runAllTimersAsync();
+
+      // execute() called without a third options argument.
+      expect(mockAgent.execute).toHaveBeenCalledWith(
+        mockContextState,
+        undefined,
+      );
+    });
+  });
+
   describe('SubagentStart hook integration', () => {
     let mockAgent: AgentHeadless;
     let mockContextState: ContextState;
diff --git a/packages/core/src/tools/agent.ts b/packages/core/src/tools/agent/agent.ts
similarity index 65%
rename from packages/core/src/tools/agent.ts
rename to packages/core/src/tools/agent/agent.ts
index b7f2f045fc8..2ff98baf395 100644
--- a/packages/core/src/tools/agent.ts
+++ b/packages/core/src/tools/agent/agent.ts
@@ -4,40 +4,56 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import { BaseDeclarativeTool, BaseToolInvocation, Kind } from './tools.js';
-import { ToolNames, ToolDisplayNames } from './tool-names.js';
+import { BaseDeclarativeTool, BaseToolInvocation, Kind } from '../tools.js';
+import { ToolNames, ToolDisplayNames } from '../tool-names.js';
 import type {
   ToolResult,
   ToolResultDisplay,
   AgentResultDisplay,
-} from './tools.js';
-import { ToolConfirmationOutcome } from './tools.js';
+} from '../tools.js';
+import { ToolConfirmationOutcome } from '../tools.js';
 import type {
   ToolCallConfirmationDetails,
   ToolConfirmationPayload,
-} from './tools.js';
-import type { Config } from '../config/config.js';
-import type { SubagentManager } from '../subagents/subagent-manager.js';
-import type { SubagentConfig } from '../subagents/types.js';
-import { AgentTerminateMode } from '../agents/runtime/agent-types.js';
-import { ContextState } from '../agents/runtime/agent-headless.js';
-import { EXCLUDED_TOOLS_FOR_SUBAGENTS } from '../agents/runtime/agent-core.js';
+} from '../tools.js';
+import type { Config } from '../../config/config.js';
+import type { SubagentManager } from '../../subagents/subagent-manager.js';
+import type { SubagentConfig } from '../../subagents/types.js';
+import { AgentTerminateMode } from '../../agents/runtime/agent-types.js';
+import type {
+  PromptConfig,
+  RunConfig,
+  ToolConfig,
+} from '../../agents/runtime/agent-types.js';
+import {
+  AgentHeadless,
+  ContextState,
+} from '../../agents/runtime/agent-headless.js';
+import type { Content, FunctionDeclaration } from '@google/genai';
+import {
+  FORK_AGENT,
+  FORK_PLACEHOLDER_RESULT,
+  buildForkedMessages,
+  buildChildMessage,
+  isInForkExecution,
+  runInForkContext,
+} from './fork-subagent.js';
 import {
   AgentEventEmitter,
   AgentEventType,
-} from '../agents/runtime/agent-events.js';
+} from '../../agents/runtime/agent-events.js';
 import type {
   AgentToolCallEvent,
   AgentToolResultEvent,
   AgentFinishEvent,
   AgentErrorEvent,
   AgentApprovalRequestEvent,
-} from '../agents/runtime/agent-events.js';
-import { BuiltinAgentRegistry } from '../subagents/builtin-agents.js';
-import { createDebugLogger } from '../utils/debugLogger.js';
-import { PermissionMode } from '../hooks/types.js';
-import type { StopHookOutput } from '../hooks/types.js';
-import { ApprovalMode } from '../config/config.js';
+} from '../../agents/runtime/agent-events.js';
+import { BuiltinAgentRegistry } from '../../subagents/builtin-agents.js';
+import { createDebugLogger } from '../../utils/debugLogger.js';
+import { PermissionMode } from '../../hooks/types.js';
+import type { StopHookOutput } from '../../hooks/types.js';
+import { ApprovalMode } from '../../config/config.js';
 
 export interface AgentParams {
   description: string;
@@ -572,141 +588,307 @@ class AgentToolInvocation extends BaseToolInvocation<AgentParams, ToolResult> {
     return this.params.description;
   }
 
-  async execute(
-    signal?: AbortSignal,
-    updateOutput?: (output: ToolResultDisplay) => void,
-  ): Promise<ToolResult> {
+  /**
+   * Creates a fork subagent that inherits the parent's conversation context
+   * and cache-safe generation params.
+   */
+  private async createForkSubagent(agentConfig: Config): Promise<{
+    subagent: AgentHeadless;
+    taskPrompt: string;
+  }> {
+    const geminiClient = this.config.getGeminiClient();
+    const rawHistory = geminiClient ? geminiClient.getHistory(true) : [];
+
+    // Build the history that will seed the fork's chat. Must end with a
+    // model message so agent-headless can send the task_prompt as a user
+    // message without creating consecutive user messages.
+    let initialMessages: Content[] | undefined;
+    let taskPrompt: string | undefined;
+    if (rawHistory.length > 0) {
+      const lastMessage = rawHistory[rawHistory.length - 1];
+      if (lastMessage.role === 'model') {
+        const forkedMessages = buildForkedMessages(
+          this.params.prompt,
+          lastMessage,
+        );
+        if (forkedMessages.length > 0) {
+          // Model had function calls: append tool responses + directive,
+          // then a model ack so history ends with model.
+          initialMessages = [
+            ...rawHistory.slice(0, -1),
+            ...forkedMessages,
+            {
+              role: 'model' as const,
+              parts: [{ text: 'Understood. Executing directive now.' }],
+            },
+          ];
+          // task_prompt is a trigger to start execution
+          taskPrompt = 'Begin.';
+        } else {
+          // Model had no function calls: history ends with model,
+          // directive goes via task_prompt.
+          initialMessages = [...rawHistory];
+        }
+      } else {
+        // History ends with user (unusual) — drop the trailing user
+        // message to avoid consecutive user messages when agent-headless
+        // sends the task_prompt.
+        initialMessages = rawHistory.slice(0, -1);
+      }
+    }
+
+    // Default: directive with fork boilerplate as task_prompt
+    if (!taskPrompt) {
+      taskPrompt = buildChildMessage(this.params.prompt);
+    }
+
+    // Read the parent's live generationConfig (systemInstruction + tool
+    // declarations) so the fork's API requests share the parent's exact
+    // cache prefix for DashScope prompt caching. When the client isn't
+    // available (first turn edge case), fall back to the fork agent's own
+    // system prompt and wildcard tools.
+    let promptConfig: PromptConfig;
+    let toolConfig: ToolConfig;
+
+    const generationConfig = geminiClient?.getChat().getGenerationConfig();
+    if (generationConfig?.systemInstruction) {
+      // Inline FunctionDeclaration[] from the parent — passed verbatim
+      // including `agent` itself so the fork's tool-name set matches the
+      // parent's. prepareTools bypasses the exclusion filter for inline
+      // decls; `isInForkExecution()` (ALS-based) is the sole
+      // recursive-fork block at runtime.
+      const parentToolDecls: FunctionDeclaration[] =
+        (
+          generationConfig.tools as Array<{
+            functionDeclarations?: FunctionDeclaration[];
+          }>
+        )?.flatMap((t) => t.functionDeclarations ?? []) ?? [];
+
+      promptConfig = {
+        renderedSystemPrompt: generationConfig.systemInstruction as
+          | string
+          | Content,
+        initialMessages,
+      };
+      toolConfig = {
+        tools:
+          parentToolDecls.length > 0 ? parentToolDecls : (['*'] as string[]),
+      };
+    } else {
+      promptConfig = {
+        systemPrompt: FORK_AGENT.systemPrompt,
+        initialMessages,
+      };
+      toolConfig = { tools: ['*'] };
+    }
+
+    const subagent = await AgentHeadless.create(
+      FORK_AGENT.name,
+      agentConfig,
+      promptConfig,
+      {},
+      {} as RunConfig,
+      toolConfig,
+      this.eventEmitter,
+    );
+
+    return { subagent, taskPrompt };
+  }
+
+  /**
+   * Runs a subagent with start/stop hook lifecycle, updating the display
+   * as execution progresses.
+   */
+  private async runSubagentWithHooks(
+    subagent: AgentHeadless,
+    contextState: ContextState,
+    opts: {
+      agentId: string;
+      agentType: string;
+      resolvedMode: PermissionMode;
+      signal?: AbortSignal;
+      updateOutput?: (output: ToolResultDisplay) => void;
+    },
+  ): Promise<void> {
+    const { agentId, agentType, resolvedMode, signal, updateOutput } = opts;
+    const hookSystem = this.config.getHookSystem();
+
     try {
-      let subagentConfig: SubagentConfig;
-      let extraHistory: Array<import('@google/genai').Content> | undefined;
-      let forkPlaceholderResult: string | undefined;
-      let forkTaskPrompt: string | undefined;
-      let forkGenerationConfig:
-        | import('@google/genai').GenerateContentConfig
-        | undefined;
-      let forkToolsOverride:
-        | Array<import('@google/genai').FunctionDeclaration>
-        | undefined;
-
-      if (!this.params.subagent_type) {
-        const {
-          FORK_AGENT,
-          FORK_PLACEHOLDER_RESULT,
-          buildForkedMessages,
-          buildChildMessage,
-          isInForkChild,
-        } = await import('../agents/runtime/forkSubagent.js');
-        forkPlaceholderResult = FORK_PLACEHOLDER_RESULT;
-        subagentConfig = FORK_AGENT;
+      if (hookSystem) {
+        try {
+          const startHookOutput = await hookSystem.fireSubagentStartEvent(
+            agentId,
+            agentType,
+            resolvedMode,
+            signal,
+          );
 
-        // Retrieve the parent's cached generationConfig (systemInstruction +
-        // tools) so the fork's API requests share the same prefix for
-        // DashScope prompt cache hits.
-        const { getCacheSafeParams } = await import('../utils/forkedAgent.js');
-        const cacheSafeParams = getCacheSafeParams();
-        if (cacheSafeParams) {
-          forkGenerationConfig = cacheSafeParams.generationConfig;
-          const tools = cacheSafeParams.generationConfig.tools;
-          if (tools && tools.length > 0) {
-            forkToolsOverride = tools
-              .flatMap(
-                (t: import('@google/genai').ToolUnion) =>
-                  (
-                    t as {
-                      functionDeclarations?: Array<
-                        import('@google/genai').FunctionDeclaration
-                      >;
-                    }
-                  ).functionDeclarations ?? [],
-              )
-              .filter(
-                (decl: import('@google/genai').FunctionDeclaration) =>
-                  !(decl.name && EXCLUDED_TOOLS_FOR_SUBAGENTS.has(decl.name)),
-              );
+          // Inject additional context from hook output into subagent context
+          const additionalContext = startHookOutput?.getAdditionalContext();
+          if (additionalContext) {
+            contextState.set('hook_context', additionalContext);
           }
+        } catch (hookError) {
+          debugLogger.warn(
+            `[Agent] SubagentStart hook failed, continuing execution: ${hookError}`,
+          );
         }
+      }
 
-        const geminiClient = this.config.getGeminiClient();
-        if (geminiClient) {
-          const rawHistory = geminiClient.getHistory(true);
+      // Execute the subagent (blocking)
+      await subagent.execute(contextState, signal);
 
-          if (isInForkChild(rawHistory)) {
-            const errorDisplay = {
-              type: 'task_execution' as const,
-              subagentName: FORK_AGENT.name,
-              taskDescription: this.params.description,
-              taskPrompt: this.params.prompt,
-              status: 'failed' as const,
-              terminateReason: 'Recursive forking is not allowed',
-            };
+      // Fire SubagentStop hook after execution and handle block decisions
+      if (hookSystem && !signal?.aborted) {
+        const transcriptPath = this.config.getTranscriptPath();
+        let stopHookActive = false;
+
+        // Loop to handle "block" decisions (prevent subagent from stopping)
+        let continueExecution = true;
+        let iterationCount = 0;
+        const maxIterations = 5; // Prevent infinite loops from hook misconfigurations
+
+        while (continueExecution) {
+          iterationCount++;
 
-            return {
-              llmContent:
-                'Error: Cannot create a fork from within an existing fork child. Please execute tasks directly.',
-              returnDisplay: errorDisplay,
-            };
+          // Safety check to prevent infinite loops
+          if (iterationCount >= maxIterations) {
+            debugLogger.warn(
+              `[TaskTool] SubagentStop hook reached maximum iterations (${maxIterations}), forcing stop to prevent infinite loop`,
+            );
+            continueExecution = false;
+            break;
           }
 
-          // Build extraHistory ensuring it ends with a model message so
-          // agent-headless can send the task_prompt as a user message
-          // without creating consecutive user messages.
-          if (rawHistory.length > 0) {
-            const lastMessage = rawHistory[rawHistory.length - 1];
-            if (lastMessage.role === 'model') {
-              const forkedMessages = buildForkedMessages(
-                this.params.prompt,
-                lastMessage,
-              );
-              if (forkedMessages.length > 0) {
-                // Model had function calls: append tool responses + directive,
-                // then a model ack so history ends with model.
-                extraHistory = [
-                  ...rawHistory.slice(0, -1),
-                  ...forkedMessages,
-                  {
-                    role: 'model' as const,
-                    parts: [{ text: 'Understood. Executing directive now.' }],
-                  },
-                ];
-                // task_prompt is a trigger to start execution
-                forkTaskPrompt = 'Begin.';
-              } else {
-                // Model had no function calls: history ends with model,
-                // directive goes via task_prompt.
-                extraHistory = [...rawHistory];
+          try {
+            const stopHookOutput = await hookSystem.fireSubagentStopEvent(
+              agentId,
+              agentType,
+              transcriptPath,
+              subagent.getFinalText(),
+              stopHookActive,
+              resolvedMode,
+              signal,
+            );
+
+            const typedStopOutput = stopHookOutput as
+              | StopHookOutput
+              | undefined;
+
+            if (
+              typedStopOutput?.isBlockingDecision() ||
+              typedStopOutput?.shouldStopExecution()
+            ) {
+              // Feed the reason back to the subagent and continue execution
+              const continueReason = typedStopOutput.getEffectiveReason();
+              stopHookActive = true;
+
+              const continueContext = new ContextState();
+              continueContext.set('task_prompt', continueReason);
+              await subagent.execute(continueContext, signal);
+
+              if (signal?.aborted) {
+                continueExecution = false;
               }
+              // Loop continues to re-check SubagentStop hook
             } else {
-              // History ends with user (unusual) — drop the trailing user
-              // message to avoid consecutive user messages when agent-headless
-              // sends the task_prompt.
-              extraHistory = rawHistory.slice(0, -1);
+              continueExecution = false;
             }
+          } catch (hookError) {
+            debugLogger.warn(
+              `[TaskTool] SubagentStop hook failed, allowing stop: ${hookError}`,
+            );
+            continueExecution = false;
           }
         }
+      }
+
+      // Get the results
+      const finalText = subagent.getFinalText();
+      const terminateMode = subagent.getTerminateMode();
+      const success = terminateMode === AgentTerminateMode.GOAL;
+      const executionSummary = subagent.getExecutionSummary();
 
-        // Default: directive with fork boilerplate as task_prompt
-        if (!forkTaskPrompt) {
-          forkTaskPrompt = buildChildMessage(this.params.prompt);
+      if (signal?.aborted) {
+        this.updateDisplay(
+          {
+            status: 'cancelled',
+            terminateReason: 'Agent was cancelled by user',
+            executionSummary,
+          },
+          updateOutput,
+        );
+      } else {
+        this.updateDisplay(
+          {
+            status: success ? 'completed' : 'failed',
+            terminateReason: terminateMode,
+            result: finalText,
+            executionSummary,
+          },
+          updateOutput,
+        );
+      }
+    } catch (error) {
+      const errorMessage =
+        error instanceof Error ? error.message : String(error);
+      debugLogger.error(
+        `[AgentTool] Error inside subagent background task: ${errorMessage}`,
+      );
+      this.updateDisplay(
+        {
+          status: 'failed',
+          terminateReason: `Failed to run subagent: ${errorMessage}`,
+        },
+        updateOutput,
+      );
+    }
+  }
+
+  async execute(
+    signal?: AbortSignal,
+    updateOutput?: (output: ToolResultDisplay) => void,
+  ): Promise<ToolResult> {
+    try {
+      const isFork = !this.params.subagent_type;
+      let subagentConfig: SubagentConfig;
+
+      if (isFork) {
+        subagentConfig = FORK_AGENT;
+
+        // Recursive-fork guard. A fork child's reasoning loop runs inside
+        // an AsyncLocalStorage frame set by `runInForkContext`; when its
+        // model calls the `agent` tool, this check fires before any history
+        // or config is touched.
+        if (isInForkExecution()) {
+          return {
+            llmContent:
+              'Error: Cannot create a fork from within an existing fork child. Please execute tasks directly.',
+            returnDisplay: {
+              type: 'task_execution' as const,
+              subagentName: FORK_AGENT.name,
+              taskDescription: this.params.description,
+              taskPrompt: this.params.prompt,
+              status: 'failed' as const,
+              terminateReason: 'Recursive forking is not allowed',
+            },
+          };
         }
       } else {
-        // Load the subagent configuration
         const loadedConfig = await this.subagentManager.loadSubagent(
-          this.params.subagent_type,
+          this.params.subagent_type!,
         );
-
         if (!loadedConfig) {
-          const errorDisplay = {
-            type: 'task_execution' as const,
-            subagentName: this.params.subagent_type,
-            taskDescription: this.params.description,
-            taskPrompt: this.params.prompt,
-            status: 'failed' as const,
-            terminateReason: `Subagent "${this.params.subagent_type}" not found`,
-          };
-
           return {
             llmContent: `Subagent "${this.params.subagent_type}" not found`,
-            returnDisplay: errorDisplay,
+            returnDisplay: {
+              type: 'task_execution' as const,
+              subagentName: this.params.subagent_type!,
+              taskDescription: this.params.description,
+              taskPrompt: this.params.prompt,
+              status: 'failed' as const,
+              terminateReason: `Subagent "${this.params.subagent_type}" not found`,
+            },
           };
         }
         subagentConfig = loadedConfig;
@@ -721,198 +903,64 @@ class AgentToolInvocation extends BaseToolInvocation<AgentParams, ToolResult> {
         status: 'running' as const,
         subagentColor: subagentConfig.color,
       };
-
-      // Set up event listeners for real-time updates
       this.setupEventListeners(updateOutput);
-
-      // Send initial display
       if (updateOutput) {
         updateOutput(this.currentDisplay);
       }
+
       // Resolve the subagent's permission mode before creating it
       const resolvedMode = resolveSubagentApprovalMode(
         this.config.getApprovalMode(),
         subagentConfig.approvalMode,
         this.config.isTrustedFolder(),
       );
-
-      // Create a config override with the resolved approval mode so the
-      // subagent's tool scheduler uses the correct mode for permission checks.
       const resolvedApprovalMode = permissionModeToApprovalMode(resolvedMode);
       const agentConfig =
         resolvedApprovalMode !== this.config.getApprovalMode()
           ? createApprovalModeOverride(this.config, resolvedApprovalMode)
           : this.config;
 
-      const subagent = await this.subagentManager.createAgentHeadless(
-        subagentConfig,
-        agentConfig,
-        { eventEmitter: this.eventEmitter },
-      );
-
-      // Create context state with the task prompt
-      // For fork agents, use the fork directive (with boilerplate) as the task
-      // prompt so it's sent as the first user message by agent-headless.
-      const contextState = new ContextState();
-      contextState.set('task_prompt', forkTaskPrompt || this.params.prompt);
-
-      // Fire SubagentStart hook before execution
-      const hookSystem = this.config.getHookSystem();
-      const agentId = `${subagentConfig.name}-${Date.now()}`;
-      const agentType = this.params.subagent_type || subagentConfig.name;
-
-      const executeSubagent = async () => {
-        try {
-          if (hookSystem) {
-            try {
-              const startHookOutput = await hookSystem.fireSubagentStartEvent(
-                agentId,
-                agentType,
-                resolvedMode,
-                signal,
-              );
+      // Create the subagent. Fork bypasses SubagentManager because its
+      // runtime configs are synthesized from the parent's cache-safe params.
+      let subagent: AgentHeadless;
+      let taskPrompt: string;
 
-              // Inject additional context from hook output into subagent context
-              const additionalContext = startHookOutput?.getAdditionalContext();
-              if (additionalContext) {
-                contextState.set('hook_context', additionalContext);
-              }
-            } catch (hookError) {
-              debugLogger.warn(
-                `[Agent] SubagentStart hook failed, continuing execution: ${hookError}`,
-              );
-            }
-          }
-
-          // Execute the subagent (blocking)
-          await subagent.execute(contextState, signal, {
-            extraHistory,
-            generationConfigOverride: forkGenerationConfig,
-            toolsOverride: forkToolsOverride,
-            skipEnvHistory: !!extraHistory && extraHistory.length > 0,
-          });
-
-          // Fire SubagentStop hook after execution and handle block decisions
-          if (hookSystem && !signal?.aborted) {
-            const transcriptPath = this.config.getTranscriptPath();
-            let stopHookActive = false;
-
-            // Loop to handle "block" decisions (prevent subagent from stopping)
-            let continueExecution = true;
-            let iterationCount = 0;
-            const maxIterations = 5; // Prevent infinite loops from hook misconfigurations
-
-            while (continueExecution) {
-              iterationCount++;
-
-              // Safety check to prevent infinite loops
-              if (iterationCount >= maxIterations) {
-                debugLogger.warn(
-                  `[TaskTool] SubagentStop hook reached maximum iterations (${maxIterations}), forcing stop to prevent infinite loop`,
-                );
-                continueExecution = false;
-                break;
-              }
+      if (isFork) {
+        const fork = await this.createForkSubagent(agentConfig);
+        subagent = fork.subagent;
+        taskPrompt = fork.taskPrompt;
+      } else {
+        subagent = await this.subagentManager.createAgentHeadless(
+          subagentConfig,
+          agentConfig,
+          { eventEmitter: this.eventEmitter },
+        );
+        taskPrompt = this.params.prompt;
+      }
 
-              try {
-                const stopHookOutput = await hookSystem.fireSubagentStopEvent(
-                  agentId,
-                  agentType,
-                  transcriptPath,
-                  subagent.getFinalText(),
-                  stopHookActive,
-                  resolvedMode,
-                  signal,
-                );
-
-                const typedStopOutput = stopHookOutput as
-                  | StopHookOutput
-                  | undefined;
-
-                if (
-                  typedStopOutput?.isBlockingDecision() ||
-                  typedStopOutput?.shouldStopExecution()
-                ) {
-                  // Feed the reason back to the subagent and continue execution
-                  const continueReason = typedStopOutput.getEffectiveReason();
-                  stopHookActive = true;
-
-                  const continueContext = new ContextState();
-                  continueContext.set('task_prompt', continueReason);
-                  await subagent.execute(continueContext, signal, {
-                    extraHistory,
-                    generationConfigOverride: forkGenerationConfig,
-                    toolsOverride: forkToolsOverride,
-                    skipEnvHistory: !!extraHistory && extraHistory.length > 0,
-                  });
-
-                  if (signal?.aborted) {
-                    continueExecution = false;
-                  }
-                  // Loop continues to re-check SubagentStop hook
-                } else {
-                  continueExecution = false;
-                }
-              } catch (hookError) {
-                debugLogger.warn(
-                  `[TaskTool] SubagentStop hook failed, allowing stop: ${hookError}`,
-                );
-                continueExecution = false;
-              }
-            }
-          }
+      const contextState = new ContextState();
+      contextState.set('task_prompt', taskPrompt);
 
-          // Get the results
-          const finalText = subagent.getFinalText();
-          const terminateMode = subagent.getTerminateMode();
-          const success = terminateMode === AgentTerminateMode.GOAL;
-          const executionSummary = subagent.getExecutionSummary();
-
-          if (signal?.aborted) {
-            this.updateDisplay(
-              {
-                status: 'cancelled',
-                terminateReason: 'Agent was cancelled by user',
-                executionSummary,
-              },
-              updateOutput,
-            );
-          } else {
-            this.updateDisplay(
-              {
-                status: success ? 'completed' : 'failed',
-                terminateReason: terminateMode,
-                result: finalText,
-                executionSummary,
-              },
-              updateOutput,
-            );
-          }
-        } catch (error) {
-          const errorMessage =
-            error instanceof Error ? error.message : String(error);
-          debugLogger.error(
-            `[AgentTool] Error inside subagent background task: ${errorMessage}`,
-          );
-          this.updateDisplay(
-            {
-              status: 'failed',
-              terminateReason: `Failed to run subagent: ${errorMessage}`,
-            },
-            updateOutput,
-          );
-        }
+      const hookOpts = {
+        agentId: `${subagentConfig.name}-${Date.now()}`,
+        agentType: this.params.subagent_type || subagentConfig.name,
+        resolvedMode,
+        signal,
+        updateOutput,
       };
 
-      if (!this.params.subagent_type) {
-        // Background fork execution
-        void executeSubagent();
+      if (isFork) {
+        // Background fork execution. Run under an AsyncLocalStorage frame so
+        // nested `agent` tool calls by the fork's model can be detected.
+        void runInForkContext(() =>
+          this.runSubagentWithHooks(subagent, contextState, hookOpts),
+        );
         return {
-          llmContent: [{ text: forkPlaceholderResult! }],
+          llmContent: [{ text: FORK_PLACEHOLDER_RESULT }],
           returnDisplay: this.currentDisplay!,
         };
       } else {
-        await executeSubagent();
+        await this.runSubagentWithHooks(subagent, contextState, hookOpts);
         const finalText = subagent.getFinalText();
         const terminateMode = subagent.getTerminateMode();
         if (terminateMode === AgentTerminateMode.ERROR) {
diff --git a/packages/core/src/agents/runtime/forkSubagent.ts b/packages/core/src/tools/agent/fork-subagent.ts
similarity index 82%
rename from packages/core/src/agents/runtime/forkSubagent.ts
rename to packages/core/src/tools/agent/fork-subagent.ts
index 222d7a6a7d2..e442f727ca8 100644
--- a/packages/core/src/agents/runtime/forkSubagent.ts
+++ b/packages/core/src/tools/agent/fork-subagent.ts
@@ -1,3 +1,4 @@
+import { AsyncLocalStorage } from 'node:async_hooks';
 import type { Content } from '@google/genai';
 
 export const FORK_SUBAGENT_TYPE = 'fork';
@@ -15,13 +16,25 @@ export const FORK_AGENT = {
   level: 'session' as const,
 };
 
-export function isInForkChild(messages: Content[]): boolean {
-  return messages.some((m) => {
-    if (m.role !== 'user') return false;
-    return m.parts?.some(
-      (part) => part.text && part.text.includes(`<${FORK_BOILERPLATE_TAG}>`),
-    );
-  });
+// Recursive-fork guard. A fork child keeps the `agent` tool in its declarations
+// for byte-identical cache parity with the parent, so tool-availability
+// stripping is no longer an option. Instead, mark the async frame as "inside a
+// fork subagent" via AsyncLocalStorage when dispatching; AgentTool.execute()
+// reads the marker and rejects nested fork calls.
+//
+// Why ALS and not a history scan: the nested AgentTool's `this.config` is the
+// main process Config, so `getGeminiClient().getHistory()` returns the parent
+// conversation — not the fork child's chat — and cannot be used to detect
+// nesting. Async context propagation works naturally across the fork's
+// await chain and is scoped per-execution.
+const forkExecutionStorage = new AsyncLocalStorage<{ readonly marker: true }>();
+
+export function runInForkContext<T>(fn: () => Promise<T>): Promise<T> {
+  return forkExecutionStorage.run({ marker: true }, fn);
+}
+
+export function isInForkExecution(): boolean {
+  return forkExecutionStorage.getStore() !== undefined;
 }
 
 export const FORK_PLACEHOLDER_RESULT =
diff --git a/packages/core/src/utils/forkedAgent.ts b/packages/core/src/utils/forkedAgent.ts
index 790f4ca61d2..2f0de662a33 100644
--- a/packages/core/src/utils/forkedAgent.ts
+++ b/packages/core/src/utils/forkedAgent.ts
@@ -226,8 +226,6 @@ export interface AgentPathParams {
   systemPrompt: string;
   /** Model override (defaults to config.getFastModel() ?? config.getModel()). */
   model?: string;
-  /** Sampling temperature (default: 0 for deterministic output). */
-  temp?: number;
   /** Maximum number of agent turns (default: unlimited). */
   maxTurns?: number;
   /** Maximum execution time in minutes (default: unlimited). */
@@ -244,11 +242,6 @@ export interface AgentPathParams {
    * Must end with a `model` role entry; call buildAgentHistory() to enforce this.
    */
   extraHistory?: Content[];
-  /**
-   * Skip env bootstrap injection in createChat() when extraHistory already
-   * contains the env context from the parent conversation.
-   */
-  skipEnvHistory?: boolean;
   /** External cancellation signal. */
   abortSignal?: AbortSignal;
 }
@@ -393,11 +386,13 @@ export async function runForkedAgent(
     }
   });
 
-  const promptConfig: PromptConfig = { systemPrompt: params.systemPrompt };
+  const promptConfig: PromptConfig = {
+    systemPrompt: params.systemPrompt,
+    initialMessages: params.extraHistory,
+  };
   const modelConfig: ModelConfig = {
     model:
       params.model ?? params.config.getFastModel() ?? params.config.getModel(),
-    temp: params.temp ?? 0,
   };
   const runConfig: RunConfig = {
     max_turns: params.maxTurns,
@@ -418,10 +413,7 @@ export async function runForkedAgent(
 
   const context = new ContextState();
   context.set('task_prompt', params.taskPrompt);
-  await headless.execute(context, params.abortSignal, {
-    extraHistory: params.extraHistory,
-    skipEnvHistory: params.skipEnvHistory,
-  });
+  await headless.execute(context, params.abortSignal);
 
   const terminateReason = headless.getTerminateMode();
   const finalText = headless.getFinalText() || undefined;