feat(task): task-local runtime thinking effort state with per-request override (DTE series 2/5)

packages/types/src/__tests__/experiment.test.ts

@@ -0,0 +1,19 @@
+import { experimentIds, experimentIdsSchema, experimentsSchema } from "../experiment.js"
+
+describe("dynamicThinkingEffort experiment", () => {
+	it("is part of the experiment id enum", () => {
+		expect(experimentIds).toContain("dynamicThinkingEffort")
+		expect(experimentIdsSchema.safeParse("dynamicThinkingEffort").success).toBe(true)
+	})
+
+	it("parses enabled and disabled states", () => {
+		expect(experimentsSchema.parse({ dynamicThinkingEffort: true })).toEqual({ dynamicThinkingEffort: true })
+		expect(experimentsSchema.parse({ dynamicThinkingEffort: false })).toEqual({ dynamicThinkingEffort: false })
+		expect(experimentsSchema.parse({})).toEqual({})
+	})
+
+	it("rejects non-boolean values", () => {
+		expect(experimentsSchema.safeParse({ dynamicThinkingEffort: "yes" }).success).toBe(false)
+		expect(experimentIdsSchema.safeParse("dynamic-thinking-effort").success).toBe(false)
+	})
+})

packages/types/src/experiment.ts

@@ -12,6 +12,7 @@ export const experimentIds = [
 	"runSlashCommand",
 	"customTools",
 	"parallelToolExecution",
+	"dynamicThinkingEffort",
 ] as const
 
 export const experimentIdsSchema = z.enum(experimentIds)
@@ -28,6 +29,7 @@ export const experimentsSchema = z.object({
 	runSlashCommand: z.boolean().optional(),
 	customTools: z.boolean().optional(),
 	parallelToolExecution: z.boolean().optional(),
+	dynamicThinkingEffort: z.boolean().optional(),
 })
 
 export type Experiments = z.infer<typeof experimentsSchema>

scripts/stryker-diff.mjs

@@ -51,8 +51,14 @@ export const PACKAGE_CONFIGS = [
 		root: "src",
 		sourceRoot: "src/",
 		vitestConfig: "vitest.config.ts",
-		vitestRelated: false,
-		discoverRelatedTests: true,
+		// Use plugin-side related-test discovery: with an explicit STRYKER_TEST_FILES
+		// list, stryker-js 10.0.0 plans static mutants with runtime activation
+		// (stryker-mutator/stryker-js#6144, #6209), which the vitest plugin can only
+		// apply after top-level code has run, producing false Survived results.
+		// Revert to vitestRelated: false / discoverRelatedTests: true once the
+		// upstream planner fix lands.
+		vitestRelated: true,
+		discoverRelatedTests: false,
 		excludedPaths: ["src/esbuild.mjs", "src/eslint.config.mjs", "src/utils/vitest-verbosity.ts"],
 	},
 ]

scripts/stryker-diff.test.mjs

@@ -129,8 +129,11 @@ describe("buildManifest", () => {
 		assert.equal(webview.runRoot, ".")
 		assert.equal(webview.discoverRelatedTests, true)
 		assert.equal(webview.vitestRelated, false)
-		assert.equal(extension.discoverRelatedTests, true)
-		assert.equal(extension.vitestRelated, false)
+		// The extension entry uses plugin-side related discovery while stryker-js 10.0.0
+		// plans static mutants with runtime activation when given explicit test files
+		// (stryker-mutator/stryker-js#6144, #6209); see the entry comment in stryker-diff.mjs.
+		assert.equal(extension.discoverRelatedTests, false)
+		assert.equal(extension.vitestRelated, true)
 	})
 
 	it("returns no packages for tests, barrels, unsupported packages, and type-only changes", () => {

src/api/index.ts

@@ -7,6 +7,7 @@ import {
 	retiredProviderIdentifiers,
 	type ProviderSettings,
 	type ModelInfo,
+	type ReasoningEffortExtended,
 } from "@roo-code/types"
 
 import { getRouterRemovalMessage } from "../core/config/routerRemoval"
@@ -115,6 +116,14 @@ export interface ApiHandlerCreateMessageMetadata {
 	 * when the user clicks stop, preventing wasted API tokens/compute on the provider side.
 	 */
 	abortSignal?: AbortSignal
+	/**
+	 * Per-request thinking effort override (DTE series 2/5).
+	 * When defined, takes precedence over the settings-derived `reasoningEffort`
+	 * wherever the effective effort is resolved (see `resolveEffectiveReasoningEffort`).
+	 * Task-scoped and transient: it applies to this request only (the next request
+	 * after being set — no mid-stream effect) and is never persisted to settings.
+	 */
+	reasoningEffort?: ReasoningEffortExtended
 }
 
 export interface ApiHandler {

src/api/transform/__tests__/dte-effective-reasoning-effort.spec.ts

@@ -0,0 +1,58 @@
+// npx vitest run src/api/transform/__tests__/dte-effective-reasoning-effort.spec.ts
+
+import { ADAPTIVE_OUTPUT_CONFIG_EFFORTS, resolveEffectiveReasoningEffort } from "../reasoning"
+
+describe("DTE series 2/5 — resolveEffectiveReasoningEffort", () => {
+	const settingsEffort = "high"
+	const modelDefault = "medium"
+
+	it("returns the per-request override when present (strongest precedence)", () => {
+		expect(
+			resolveEffectiveReasoningEffort({
+				override: "xhigh",
+				settingsReasoningEffort: settingsEffort,
+				modelDefaultEffort: modelDefault,
+			}),
+		).toBe("xhigh")
+	})
+
+	it("lets the override win even when it is out-of-range for the adaptive envelope", () => {
+		// "minimal" is a valid override value but outside the adaptive envelope set;
+		// resolution still returns it — envelope gating is the caller's concern.
+		expect(
+			resolveEffectiveReasoningEffort({
+				override: "minimal",
+				settingsReasoningEffort: settingsEffort,
+				modelDefaultEffort: modelDefault,
+			}),
+		).toBe("minimal")
+	})
+
+	it("falls back to the settings value when no override is present", () => {
+		expect(
+			resolveEffectiveReasoningEffort({ settingsReasoningEffort: "low", modelDefaultEffort: modelDefault }),
+		).toBe("low")
+	})
+
+	it("preserves the settings 'disable' sentinel when no override is present", () => {
+		expect(
+			resolveEffectiveReasoningEffort({ settingsReasoningEffort: "disable", modelDefaultEffort: modelDefault }),
+		).toBe("disable")
+	})
+
+	it("an explicit override wins over a settings 'disable' sentinel", () => {
+		expect(resolveEffectiveReasoningEffort({ override: "low", settingsReasoningEffort: "disable" })).toBe("low")
+	})
+
+	it("falls back to the model default when neither override nor settings is set", () => {
+		expect(resolveEffectiveReasoningEffort({ modelDefaultEffort: "low" })).toBe("low")
+	})
+
+	it("returns undefined when nothing is set", () => {
+		expect(resolveEffectiveReasoningEffort({})).toBeUndefined()
+	})
+
+	it("exposes exactly the in-range adaptive envelope efforts", () => {
+		expect([...ADAPTIVE_OUTPUT_CONFIG_EFFORTS]).toEqual(["low", "medium", "high", "xhigh", "max"])
+	})
+})

src/api/transform/reasoning.ts

@@ -22,6 +22,51 @@ export type AnthropicProviderReasoningParams = AnthropicReasoningParams | { type
 
 export type OpenAiReasoningParams = { reasoning_effort: OpenAI.Chat.ChatCompletionCreateParams["reasoning_effort"] }
 
+/**
+ * DTE series 2/5 — effort levels accepted by the Claude 4.7+ adaptive-thinking
+ * `output_config.effort` envelope. Efforts outside this set (e.g. "none",
+ * "minimal", "disable") omit the envelope so the API applies its own default.
+ */
+export const ADAPTIVE_OUTPUT_CONFIG_EFFORTS: readonly ReasoningEffortExtended[] = [
+	"low",
+	"medium",
+	"high",
+	"xhigh",
+	"max",
+]
+
+/**
+ * DTE series 2/5 — resolves the effective thinking effort for a single request.
+ *
+ * Resolution order (strongest first):
+ * 1. `override` — the per-request task-local effort
+ *    (`ApiHandlerCreateMessageMetadata.reasoningEffort`),
+ * 2. `settingsReasoningEffort` — the settings-derived value,
+ * 3. `modelDefaultEffort` — the model's default effort.
+ *
+ * This is the single shared resolution point for the per-request override:
+ * providers that resolve the effective effort through it inherit the override
+ * without duplicating precedence logic. The override is transient (next request
+ * only) and never persisted to settings.
+ */
+export const resolveEffectiveReasoningEffort = ({
+	override,
+	settingsReasoningEffort,
+	modelDefaultEffort,
+}: {
+	override?: ReasoningEffortExtended
+	settingsReasoningEffort?: ReasoningEffortExtended | "disable"
+	modelDefaultEffort?: ReasoningEffortExtended
+}): ReasoningEffortExtended | "disable" | undefined => {
+	if (override !== undefined) {
+		return override
+	}
+	if (settingsReasoningEffort !== undefined) {
+		return settingsReasoningEffort
+	}
+	return modelDefaultEffort
+}
+
 // Valid Gemini thinking levels for effort-based reasoning
 const GEMINI_THINKING_LEVELS = ["minimal", "low", "medium", "high"] as const
 

src/core/task/Task.ts

@@ -22,6 +22,7 @@ import {
 	type TaskMetadata,
 	type TaskEvents,
 	type ProviderSettings,
+	type ReasoningEffortExtended,
 	type TokenUsage,
 	type ToolUsage,
 	type ToolName,
@@ -318,6 +319,13 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	// API
 	apiConfiguration: ProviderSettings
 	api: ApiHandler
+	// DTE series 2/5: task-local thinking effort override. Transient per-task state —
+	// never persisted to settings; cleared on dispose (see dispose()).
+	private runtimeThinkingEffort?: ReasoningEffortExtended
+	private runtimeThinkingEffortSource?: string
+	// Settings-derived effort captured when the override activates, so clearing
+	// (undefined) restores it in the in-memory apiConfiguration copy.
+	private preOverrideReasoningEffort?: ProviderSettings["reasoningEffort"]
 	private rateLimitClock: RateLimitClock
 	private autoApprovalHandler: AutoApprovalHandler
 
@@ -1633,14 +1641,89 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	 * Updates the API configuration and rebuilds the API handler.
 	 * There is no tool-protocol switching or tool parser swapping.
 	 *
+	 * DTE series 2/5: when a task-local thinking effort override is active
+	 * (`setRuntimeThinkingEffort`), the incoming configuration's `reasoningEffort`
+	 * becomes the new restore value and the override is re-applied on top of the
+	 * fresh in-memory copy — clearing the override later restores the NEW profile's
+	 * value, not a stale one.
+	 *
 	 * @param newApiConfiguration - The new API configuration to use
 	 */
 	public updateApiConfiguration(newApiConfiguration: ProviderSettings): void {
 		// Update the configuration and rebuild the API handler
-		this.apiConfiguration = newApiConfiguration
+		if (this.runtimeThinkingEffort !== undefined) {
+			// DTE series 2/5: a task-local override is active, so re-capture the
+			// incoming profile's value as the restore value and re-apply the
+			// override on top of the new in-memory copy — clearing the override
+			// must restore the NEW profile's value, not the stale one.
+			this.preOverrideReasoningEffort = newApiConfiguration.reasoningEffort
+			this.apiConfiguration = { ...newApiConfiguration, reasoningEffort: this.runtimeThinkingEffort }
+		} else {
+			this.apiConfiguration = newApiConfiguration
+		}
 		this.api = buildApiHandler(this.apiConfiguration)
 	}
 
+	/**
+	 * DTE series 2/5: sets — or clears with `undefined` — the task-local thinking
+	 * effort override.
+	 *
+	 * Resolution order for the affected requests (strongest first): this
+	 * task-local override → settings `reasoningEffort` → model default. The
+	 * override applies to the NEXT API request only (no mid-stream effect): it is
+	 * passed per request as `metadata.reasoningEffort` and, while active, is
+	 * merged into the in-memory `apiConfiguration` copy (profile-switch /
+	 * `updateApiConfiguration` precedent) so the rebuilt handler reflects it too.
+	 * `undefined` clears the override and restores the settings-derived value in
+	 * the copy. Nothing is ever written to persisted settings.
+	 *
+	 * @param effort - The task-local effort, or `undefined` to clear.
+	 * @param source - Optional provenance label (UI wiring lands in a later PR).
+	 */
+	public setRuntimeThinkingEffort(effort: ReasoningEffortExtended | undefined, source?: string): void {
+		const wasActive = this.runtimeThinkingEffort !== undefined
+		this.runtimeThinkingEffort = effort
+		this.runtimeThinkingEffortSource = effort === undefined ? undefined : source
+
+		if (effort !== undefined) {
+			// Capture the settings-derived value once so clearing can restore it.
+			if (!wasActive) {
+				this.preOverrideReasoningEffort = this.apiConfiguration.reasoningEffort
+			}
+			// Merge into the in-memory copy (never the persisted settings object).
+			this.apiConfiguration = { ...this.apiConfiguration, reasoningEffort: effort }
+		} else if (wasActive) {
+			// Restore the settings-derived value captured when the override activated.
+			this.apiConfiguration = { ...this.apiConfiguration, reasoningEffort: this.preOverrideReasoningEffort }
+			this.preOverrideReasoningEffort = undefined
+		} else {
+			// Already inactive: nothing to clear.
+			return
+		}
+
+		// Rebuild the handler from the updated copy so the next request uses it.
+		this.api = buildApiHandler(this.apiConfiguration)
+	}
+
+	/**
+	 * DTE series 2/5: reads the current task-local thinking effort override.
+	 */
+	public getRuntimeThinkingEffort(): { effort?: ReasoningEffortExtended; source?: string } {
+		return {
+			effort: this.runtimeThinkingEffort,
+			source: this.runtimeThinkingEffortSource,
+		}
+	}
+
+	/**
+	 * DTE series 2/5: metadata fragment carrying the active task-local effort
+	 * override on a single request. Empty when no override is active, so the
+	 * existing settings resolution applies unchanged.
+	 */
+	private getRuntimeThinkingEffortMetadata(): Pick<ApiHandlerCreateMessageMetadata, "reasoningEffort"> {
+		return this.runtimeThinkingEffort !== undefined ? { reasoningEffort: this.runtimeThinkingEffort } : {}
+	}
+
 	public async submitUserMessage(
 		text: string,
 		images?: string[],
@@ -1757,6 +1840,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 						parallelToolCalls: true,
 					}
 				: {}),
+			// DTE series 2/5: carry the active task-local effort override.
+			...this.getRuntimeThinkingEffortMetadata(),
 		}
 		// Generate environment details to include in the condensed summary
 		const environmentDetails = await getEnvironmentDetails(this, true)
@@ -2525,6 +2610,10 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 		}
 	}
 
+	/**
+	 * Centralized task teardown: releases task resources and resets transient
+	 * task-local state.
+	 */
 	public dispose(): Promise<void> {
 		if (this.disposalPromise) {
 			return this.disposalPromise
@@ -2617,6 +2706,12 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 			console.error("Error reverting diff changes:", error)
 		}
 
+		// Clear the DTE runtime-effort override via the standard clearing call so a
+		// retained disposed task never serves a stale override: the override also
+		// lives in apiConfiguration.reasoningEffort and the built api handler, and
+		// the clearing call restores both before clearing the runtime fields.
+		this.setRuntimeThinkingEffort(undefined)
+
 		await pendingCleanup
 		await this.diffReversionPromise
 	}
@@ -4200,6 +4295,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 						parallelToolCalls: true,
 					}
 				: {}),
+			// DTE series 2/5: carry the active task-local effort override.
+			...this.getRuntimeThinkingEffortMetadata(),
 		}
 
 		try {
@@ -4426,6 +4523,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 							parallelToolCalls: true,
 						}
 					: {}),
+				// DTE series 2/5: carry the active task-local effort override.
+				...this.getRuntimeThinkingEffortMetadata(),
 			}
 
 			// Only generate environment details when context management will actually run.
@@ -4591,6 +4690,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 			taskId: this.taskId,
 			suppressPreviousResponseId: this.skipPrevResponseIdOnce,
 			abortSignal,
+			// DTE series 2/5: carry the active task-local effort override for this request.
+			...this.getRuntimeThinkingEffortMetadata(),
 			// Include tools whenever they are present.
 			...(shouldIncludeTools
 				? {