feat(server): custom Claude models can pick reasoning, thinking, and 1M

apps/server/src/claudeModelOptions.test.ts

@@ -3,6 +3,10 @@ import { describe, expect, it } from "@effect/vitest";
 import { ProviderInstanceId, type ModelSelection } from "@t3tools/contracts";
 
 import { compileClaudeModelSelection } from "./claudeModelOptions.ts";
+import {
+  BUNDLED_CLAUDE_MODEL_CATALOG,
+  scopeClaudeModelCatalog,
+} from "./provider/ClaudeModelCatalog.ts";
 
 const selection = (
   model: string,
@@ -74,4 +78,42 @@ describe("compileClaudeModelSelection", () => {
         .settings,
     ).toEqual({ alwaysThinkingEnabled: false });
   });
+
+  it("defaults custom models to thinking on and 200k, with a reversible 1M choice", () => {
+    const catalog = scopeClaudeModelCatalog(BUNDLED_CLAUDE_MODEL_CATALOG, ["glm-flash"]);
+    for (const options of [[], [{ id: "contextWindow", value: "200k" }]]) {
+      expect(compileClaudeModelSelection(selection("glm-flash", options), catalog)).toMatchObject({
+        apiModelId: "glm-flash",
+        effort: "high",
+        settings: { alwaysThinkingEnabled: true },
+      });
+    }
+    expect(
+      compileClaudeModelSelection(
+        selection("glm-flash", [{ id: "contextWindow", value: "1m" }]),
+        catalog,
+      ),
+    ).toMatchObject({
+      apiModelId: "glm-flash[1m]",
+      settings: { alwaysThinkingEnabled: true },
+    });
+  });
+
+  it("compiles 1M, effort, and thinking for a bare custom model", () => {
+    const catalog = scopeClaudeModelCatalog(BUNDLED_CLAUDE_MODEL_CATALOG, ["glm-flash"]);
+    expect(
+      compileClaudeModelSelection(
+        selection("glm-flash", [
+          { id: "contextWindow", value: "1m" },
+          { id: "effort", value: "max" },
+          { id: "thinking", value: false },
+        ]),
+        catalog,
+      ),
+    ).toMatchObject({
+      apiModelId: "glm-flash[1m]",
+      effort: "max",
+      settings: { alwaysThinkingEnabled: false },
+    });
+  });
 });

apps/server/src/claudeModelOptions.ts

@@ -3,6 +3,7 @@ import {
   getModelSelectionBooleanOptionValue,
   getModelSelectionStringOptionValue,
   getProviderOptionDescriptors,
+  getProviderOptionCurrentValue,
   resolvePromptInjectedEffort,
 } from "@t3tools/shared/model";
 
@@ -39,7 +40,8 @@ export function compileClaudeModelSelection(
     ? (getModelSelectionBooleanOptionValue(selection, "fastMode") ?? false)
     : undefined;
   const thinking = supportsBoolean("thinking")
-    ? getModelSelectionBooleanOptionValue(selection, "thinking")
+    ? (getModelSelectionBooleanOptionValue(selection, "thinking") ??
+      getProviderOptionCurrentValue(descriptors.find((descriptor) => descriptor.id === "thinking")))
     : undefined;
   const settings = {
     ...(typeof thinking === "boolean" ? { alwaysThinkingEnabled: thinking } : {}),

apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts

@@ -88,6 +88,10 @@ import { WorktreeToolkit } from "../../mcp/toolkits/worktree/tools.ts";
 import { ThreadToolkit } from "../../mcp/toolkits/thread/tools.ts";
 import { OrchestratorToolkit } from "../../mcp/toolkits/orchestrator/tools.ts";
 import { ClaudeExecutableFileCheck } from "../../provider/Drivers/ClaudeExecutable.ts";
+import {
+  BUNDLED_CLAUDE_MODEL_CATALOG,
+  scopeClaudeModelCatalog,
+} from "../../provider/ClaudeModelCatalog.ts";
 import * as ProviderEventLoggers from "@t3tools/provider-core/server/ProviderEventLoggers";
 import * as ProviderAdapter from "@t3tools/provider-core/server/ProviderAdapter";
 import type * as ProviderContinuationRequests from "@t3tools/provider-core/server/ProviderContinuationRequests";
@@ -126,6 +130,9 @@ const DEFAULT_CLAUDE_SETTINGS = Schema.decodeSync(ClaudeSettings)({});
 const AUTO_COMPACT_CLAUDE_SETTINGS = Schema.decodeSync(ClaudeSettings)({
   autoCompactWindow: "300000",
 });
+const BARE_CUSTOM_MODEL_CLAUDE_SETTINGS = Schema.decodeSync(ClaudeSettings)({
+  customModels: [{ slug: "gateway-bare", name: "Gateway Bare" }],
+});
 const CUSTOM_MODEL_CLAUDE_SETTINGS = Schema.decodeSync(ClaudeSettings)({
   customModels: [
     "opus",
@@ -1026,6 +1033,38 @@ describe("ClaudeAdapterV2 context usage", () => {
     );
     assert.strictEqual(usage.promptCacheExpiresAt, "2026-08-29T00:05:00.000Z");
   });
+
+  it("defaults a bare custom slug and opens 1M through the turn path", () => {
+    const settings = BARE_CUSTOM_MODEL_CLAUDE_SETTINGS;
+    const bareSelection = {
+      instanceId: ProviderInstanceId.make(ClaudeAdapterV2.CLAUDE_PROVIDER),
+      model: "gateway-bare",
+      options: [
+        { id: "effort", value: "max" },
+        { id: "contextWindow", value: "1m" },
+      ],
+    } satisfies ModelSelection;
+    const options = ClaudeAdapterV2.makeClaudeQueryOptions({
+      modelSelection: bareSelection,
+      nativeThreadId: "bare-model-thread",
+      resume: false,
+      cwd: "/workspace",
+      settings,
+    });
+    assert.equal(options.model, "gateway-bare[1m]");
+    assert.equal(options.effort, "max");
+    assert.equal(
+      ClaudeAdapterV2.claudeProviderTurnTokenUsage(
+        { input_tokens: 1, output_tokens: 1 },
+        bareSelection,
+        "2026-08-29T00:00:00.000Z",
+        "",
+        undefined,
+        scopeClaudeModelCatalog(BUNDLED_CLAUDE_MODEL_CATALOG, settings.customModels),
+      ).maxTokens,
+      1_000_000,
+    );
+  });
 });
 
 describe("ClaudeAdapterV2 session permissions", () => {
@@ -12630,7 +12669,11 @@ describe("ClaudeAdapterV2 background wake turns", () => {
             model: "claude-custom-tuned",
             options: [{ id: "effort", value: "brutal" }],
           },
-          [`task-${SHADOW_TOOL_USE_ID}`]: { instanceId, model: "opus" },
+          [`task-${SHADOW_TOOL_USE_ID}`]: {
+            instanceId,
+            model: "opus",
+            options: [{ id: "effort", value: "high" }],
+          },
         });
       }).pipe(
         Effect.provide(

apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts

@@ -191,10 +191,11 @@ import {
 export const CLAUDE_PROVIDER = ProviderDriverKind.make("claudeAgent");
 export const CLAUDE_AGENT_SDK_QUERY_PROTOCOL = "claude-agent-sdk.query" as const;
 
-function claudeContextWindow(modelSelection: ModelSelection): number | null {
-  return (
-    resolveClaudeCatalogContextWindowTokens(BUNDLED_CLAUDE_MODEL_CATALOG, modelSelection) ?? null
-  );
+function claudeContextWindow(
+  catalog: ClaudeModelCatalog,
+  modelSelection: ModelSelection,
+): number | null {
+  return resolveClaudeCatalogContextWindowTokens(catalog, modelSelection) ?? null;
 }
 
 interface ClaudeCacheCreation {
@@ -243,11 +244,12 @@ export function claudeProviderTurnTokenUsage(
   updatedAt: string,
   autoCompactWindow = "",
   requestedAt?: DateTime.Utc,
+  catalog: ClaudeModelCatalog = BUNDLED_CLAUDE_MODEL_CATALOG,
 ) {
   const inputTokens = claudeMessageInputTokens(usage);
   const outputTokens = usage.output_tokens;
   const promptCacheTtlMs = claudePromptCacheTtlMs(usage.cache_creation);
-  const maxTokens = claudeContextWindow(modelSelection);
+  const maxTokens = claudeContextWindow(catalog, modelSelection);
   // The cache lifetime starts with the request, not when its response ends.
   const promptCacheExpiresAt =
     promptCacheTtlMs === undefined || requestedAt === undefined
@@ -1261,7 +1263,10 @@ export function makeClaudeQueryOptions(input: {
   readonly allowDangerouslySkipPermissions?: boolean;
   readonly appendSystemPromptFileText?: string;
 }): ClaudeAgentSdkQueryOptions {
-  const compiledSelection = compileClaudeModelSelection(input.modelSelection);
+  const compiledSelection = compileClaudeModelSelection(
+    input.modelSelection,
+    scopeClaudeModelCatalog(BUNDLED_CLAUDE_MODEL_CATALOG, input.settings?.customModels ?? []),
+  );
   const {
     "permission-mode": launchArgPermissionMode,
     "dangerously-skip-permissions": launchArgSkipPermissions,
@@ -4215,6 +4220,7 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
             DateTime.formatIso(now),
             adapterOptions.settings.autoCompactWindow,
             context.streamUsage?.requestedAt,
+            subagentModelCatalog,
           );
           if (tokenUsage.inputTokens <= 0) return;
           yield* emitProviderEvent({
@@ -5534,7 +5540,10 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
             existingSubagent === undefined
               ? (input.effort ??
                 input.owner?.effort ??
-                compileClaudeModelSelection(input.context.input.modelSelection).effort)
+                compileClaudeModelSelection(
+                  input.context.input.modelSelection,
+                  subagentModelCatalog,
+                ).effort)
               : existingSubagent.effort;
           const model = input.model ?? priorTask?.model ?? null;
           const task = {
@@ -7691,7 +7700,10 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
                 ? Math.round(postTokens)
                 : undefined;
             if (afterTokenCount !== undefined) {
-              const maxTokens = claudeContextWindow(context.input.modelSelection);
+              const maxTokens = claudeContextWindow(
+                subagentModelCatalog,
+                context.input.modelSelection,
+              );
               yield* emitProviderEvent({
                 type: "provider_turn.updated",
                 driver: CLAUDE_PROVIDER,
@@ -9043,7 +9055,10 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
               : { allowedTools: queryPolicy.allowedTools }),
           });
           const queryPolicyKey = claudeEffectiveQueryPolicyKey(queryPolicy, mcpOverrides);
-          const compiledSelection = compileClaudeModelSelection(turnInput.modelSelection);
+          const compiledSelection = compileClaudeModelSelection(
+            turnInput.modelSelection,
+            subagentModelCatalog,
+          );
           const resumeSessionAt = yield* getNativeConversationHeadId(turnInput.providerThread);
           const existing = yield* Ref.get(queryContext);
           // A continuation prompts nothing: it drains output the live process
@@ -9331,7 +9346,8 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
                 : yield* makeClaudeUserMessageWithAttachments({
                     text: applyClaudePromptEffortPrefix(
                       turnInput.message.text,
-                      compileClaudeModelSelection(turnInput.modelSelection).promptEffort,
+                      compileClaudeModelSelection(turnInput.modelSelection, subagentModelCatalog)
+                        .promptEffort,
                     ),
                     attachments: turnInput.message.attachments,
                     attachmentsDir,
@@ -9761,7 +9777,8 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
             const userMessage = yield* makeClaudeUserMessageWithAttachments({
               text: applyClaudePromptEffortPrefix(
                 turnInput.message.text,
-                compileClaudeModelSelection(currentTurn.input.modelSelection).promptEffort,
+                compileClaudeModelSelection(currentTurn.input.modelSelection, subagentModelCatalog)
+                  .promptEffort,
               ),
               attachments: turnInput.message.attachments,
               priority: "now",
@@ -9896,7 +9913,7 @@ export const makeClaudeAdapterV2 = Effect.fn("makeClaudeAdapterV2")(function* (
           providerSessionId: input.providerSessionId,
           providerSession: session,
           getModelContextWindow: (selection) =>
-            resolveClaudeCatalogContextWindowTokens(BUNDLED_CLAUDE_MODEL_CATALOG, selection),
+            resolveClaudeCatalogContextWindowTokens(subagentModelCatalog, selection),
           events: Stream.fromEffectRepeat(Queue.take(events)),
           backgroundTasksEndWithProcess: true,
           hasPendingBackgroundWork: Effect.gen(function* () {