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

src/api/index.ts

@@ -7,6 +7,7 @@ import {
 	retiredProviderIdentifiers,
 	type ProviderSettings,
 	type ModelInfo,
+	type ReasoningEffortExtended,
 	type ModelRecord,
 } from "@roo-code/types"
 
@@ -127,6 +128,16 @@ export interface ApiHandlerCreateMessageMetadata {
 	 * when the user clicks stop, preventing wasted API tokens/compute on the provider side.
 	 */
 	abortSignal?: AbortSignal
+	/**
+	 * Per-request thinking effort override applied by rebuilding the handler.
+	 * Providers resolve effort from the handler options (each provider reads
+	 * `options.reasoningEffort` through its own model-params step), so passing the
+	 * override in the configuration handed to `buildApiHandler` is what takes effect;
+	 * it outranks the settings-derived `reasoningEffort` for the handler built from
+	 * this object. Task-scoped and transient: it applies to the next request only
+	 * (no mid-stream effect) and is never persisted to settings.
+	 */
+	reasoningEffort?: ReasoningEffortExtended
 }
 
 /**

src/core/condense/__tests__/condense-effort-metadata.spec.ts

@@ -0,0 +1,98 @@
+// npx vitest run src/core/condense/__tests__/condense-effort-metadata.spec.ts
+//
+// The summarizer boundary: Task hands the task-local thinking effort override to
+// to summarizeConversation inside request metadata. This checks the other half of that hand-off:
+// the real summarizer forwards the metadata to the handler's createMessage instead of dropping
+// the effort. The Task-side tests cover Task-to-summarizer propagation only.
+
+import { Anthropic } from "@anthropic-ai/sdk"
+import type { ModelInfo, ReasoningEffortExtended } from "@roo-code/types"
+import { TelemetryService } from "@roo-code/telemetry"
+
+import type { ApiHandlerCreateMessageMetadata } from "../../../api"
+import { BaseProvider } from "../../../api/providers/base-provider"
+import { ApiStream, ApiStreamChunk } from "../../../api/transform/stream"
+import { ApiMessage } from "../../task-persistence/apiMessages"
+import { summarizeConversation } from "../index"
+
+const OVERRIDE_EFFORT: ReasoningEffortExtended = "xhigh"
+const TASK_ID = "dte-condense-task"
+
+class RecordingApiHandler extends BaseProvider {
+	/** The metadata of the last request this handler was asked to make. */
+	public requestedMetadata: ApiHandlerCreateMessageMetadata | undefined
+
+	override createMessage(
+		_systemPrompt: string,
+		_messages: Anthropic.Messages.MessageParam[],
+		metadata?: ApiHandlerCreateMessageMetadata,
+	): ApiStream {
+		this.requestedMetadata = metadata
+		return (async function* (): AsyncGenerator<ApiStreamChunk> {
+			yield { type: "text", text: "summary of the conversation" }
+			yield { type: "usage", inputTokens: 10, outputTokens: 5, totalCost: 0.01 }
+		})()
+	}
+
+	override getModel(): { id: string; info: ModelInfo } {
+		return {
+			id: "dte-condense-model",
+			info: { contextWindow: 100_000, maxTokens: 5_000, supportsPromptCache: false, supportsImages: false },
+		}
+	}
+}
+
+// Three messages with no summary in them: getMessagesSinceLastSummary returns all of them, so
+// summarizeConversation gets past its "not enough messages" guard and issues the request.
+function conversationHistory(): ApiMessage[] {
+	return [
+		{ role: "user", content: [{ type: "text", text: "read the failing test" }], ts: 1 },
+		{ role: "assistant", content: [{ type: "text", text: "the assertion drifted" }], ts: 2 },
+		{ role: "user", content: [{ type: "text", text: "fix it" }], ts: 3 },
+	]
+}
+
+describe("summarizeConversation reasoning-effort forwarding", () => {
+	beforeEach(() => {
+		if (!TelemetryService.hasInstance()) {
+			TelemetryService.createInstance([])
+		}
+	})
+
+	it("forwards an active override to the handler request", async () => {
+		const handler = new RecordingApiHandler()
+
+		const result = await summarizeConversation({
+			messages: conversationHistory(),
+			apiHandler: handler,
+			systemPrompt: "system prompt",
+			taskId: TASK_ID,
+			isAutomaticTrigger: false,
+			metadata: { mode: "code", taskId: TASK_ID, reasoningEffort: OVERRIDE_EFFORT },
+			environmentDetails: "environment details",
+			cwd: "/workspace",
+		})
+
+		expect(result.error).toBeUndefined()
+		expect(result.summary).toBe("summary of the conversation")
+		expect(handler.requestedMetadata?.reasoningEffort).toBe(OVERRIDE_EFFORT)
+	})
+
+	it("omits the key when the caller hands no override", async () => {
+		const handler = new RecordingApiHandler()
+
+		const result = await summarizeConversation({
+			messages: conversationHistory(),
+			apiHandler: handler,
+			systemPrompt: "system prompt",
+			taskId: TASK_ID,
+			isAutomaticTrigger: false,
+			metadata: { mode: "code", taskId: TASK_ID },
+			environmentDetails: "environment details",
+			cwd: "/workspace",
+		})
+
+		expect(result.summary).toBe("summary of the conversation")
+		expect(handler.requestedMetadata).not.toHaveProperty("reasoningEffort")
+	})
+})

src/core/task/Task.ts

@@ -21,6 +21,7 @@ import {
 	type TaskMetadata,
 	type TaskEvents,
 	type ProviderSettings,
+	type ReasoningEffortExtended,
 	type TokenUsage,
 	type ToolUsage,
 	type ToolName,
@@ -429,6 +430,10 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	// API
 	apiConfiguration: ProviderSettings
 	api: ApiHandler
+	// Task-local thinking effort override. Transient per-task state —
+	// never persisted to settings; cleared on dispose (see dispose()).
+	private runtimeThinkingEffort?: ReasoningEffortExtended
+	private runtimeThinkingEffortSource?: string
 	/**
 	 * Units of in-flight API work: a request from setup through its retries and
 	 * background usage drain, or a manual condense. Work can hold any handler it
@@ -2555,16 +2560,39 @@ 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.
 	 *
+	 * 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. Handlers such as
 		// VS Code LM / GitHub Copilot own host listeners, so the old one is
 		// released — immediately when idle, or when the last request, retry,
 		// background drain, or condense that leased it settles.
+		// apiConfiguration stays the profile-derived projection;
+		// an active task-local override is applied to the rebuilt handler only, so
+		// a profile switch keeps the override live and clearing it later restores
+		// the NEW profile's value (it is the configuration itself now).
+		//
+		// Same disposal guard as setRuntimeThinkingEffort: this is the other path that rebuilds the
+		// handler, and once disposal has started its handler set is already collected, so a rebuild
+		// would strand the replacement and re-retire a disposed handler. The task is going away, so
+		// there is nothing left to re-project either.
+		if (this.disposalPromise !== undefined) {
+			return
+		}
+
 		const previousApi = this.api
 		this.apiConfiguration = newApiConfiguration
-		this.api = buildApiHandler(this.apiConfiguration)
+		this.api = buildApiHandler(
+			this.runtimeThinkingEffort !== undefined
+				? { ...newApiConfiguration, reasoningEffort: this.runtimeThinkingEffort }
+				: newApiConfiguration,
+		)
 		this.retireApi(previousApi)
 
 		// Refresh only the soft tool-repetition limit from the new profile so an
@@ -2581,6 +2609,76 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 		this.toolRepetitionDetector.updateLimits(this.toolRepetitionSoftLimit, this.consecutiveMistakeLimit)
 	}
 
+	/**
+	 * 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 applied to the
+	 * rebuilt API handler. `apiConfiguration` itself stays the profile-derived
+	 * projection — the override never enters it — so the webview state (which
+	 * posts the active task's `apiConfiguration`) and any profile save built
+	 * from it can never carry the task-local effort. `undefined` clears the
+	 * override and rebuilds the handler from the projection.
+	 *
+	 * Not inherited: child, handoff, and resumed tasks construct their own Task and
+	 * start with no override — it lives in this task's private fields and is never
+	 * persisted.
+	 * @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 {
+		// Disposal has started (or finished): `disposalPromise` is set once and never cleared, so it
+		// is the liveness signal. The override is dropped rather than applied — the task serves no
+		// further request, and the handler set that dispose collects was already snapshotted, so
+		// rebuilding here would produce a handler nobody owns (still holding its configuration
+		// listeners) and retire the handler disposal has just disposed.
+		if (this.disposalPromise !== undefined) {
+			return
+		}
+
+		const wasActive = this.runtimeThinkingEffort !== undefined
+		this.runtimeThinkingEffort = effort
+		this.runtimeThinkingEffortSource = effort === undefined ? undefined : source
+
+		if (effort === undefined && !wasActive) {
+			// Already inactive: nothing to clear.
+			return
+		}
+
+		// Rebuild the handler with the effective effort applied. The projection
+		// (this.apiConfiguration) is untouched, so clearing needs no restore: the
+		// handler is simply rebuilt from the profile-derived configuration. The
+		// replaced handler is retired like any other swap: handlers with host
+		// listeners must not leak when the effort changes.
+		const previousApi = this.api
+		this.api = buildApiHandler(
+			effort !== undefined ? { ...this.apiConfiguration, reasoningEffort: effort } : this.apiConfiguration,
+		)
+		this.retireApi(previousApi)
+	}
+
+	/**
+	 * Reads the current task-local thinking effort override.
+	 */
+	public getRuntimeThinkingEffort(): { effort?: ReasoningEffortExtended; source?: string } {
+		return {
+			effort: this.runtimeThinkingEffort,
+			source: this.runtimeThinkingEffortSource,
+		}
+	}
+
+	/**
+	 * 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 } : {}
+	}
+
 	/**
 	 * Marks the start of in-flight API work. The returned release is idempotent;
 	 * the last release disposes every handler replaced while work was running.
@@ -2760,6 +2858,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 						parallelToolCalls: true,
 					}
 				: {}),
+			// Carry the active task-local effort override.
+			...this.getRuntimeThinkingEffortMetadata(),
 		}
 		// Generate environment details to include in the condensed summary
 		const environmentDetails = await getEnvironmentDetails(this, true)
@@ -3707,6 +3807,10 @@ 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
@@ -3826,6 +3930,16 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 			console.error("Error reverting diff changes:", error)
 		}
 
+		// Clear the DTE runtime-effort override so a retained disposed task never
+		// reports a stale override. The fields are cleared directly instead of via
+		// setRuntimeThinkingEffort: that setter rebuilds the handler, and during
+		// disposal the rebuild would (a) create a replacement handler this pass has
+		// already collected and disposed, and (b) retire the just-disposed handler a
+		// second time. `apiConfiguration` never carries the override (it stays the
+		// profile-derived projection), so there is nothing to restore.
+		this.runtimeThinkingEffort = undefined
+		this.runtimeThinkingEffortSource = undefined
+
 		await pendingCleanup
 		await this.diffReversionPromise
 	}
@@ -4372,7 +4486,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 										// under the canonical name slip past a name comparison. The first call wins.
 										const idAlreadyUsed = this.assistantMessageContent.some(
 											(entry) =>
-												// Stryker disable next-line ConditionalExpression: equivalent mutant, only tool_use and mcp_tool_use entries carry an id, so the type check cannot change entry.id === event.id
 												(entry.type === "tool_use" || entry.type === "mcp_tool_use") &&
 												(entry as { id?: string }).id === event.id,
 										)
@@ -4425,7 +4538,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 										// keep the deprecated single-entry-per-ID lookup.
 										let streamingKey: string | undefined
 
-										// Stryker disable next-line ConditionalExpression: equivalent mutant, event.name is always a string here, so the legacy nameless branch is unreachable
 										if (event.name !== undefined) {
 											streamingKey = NativeToolCallParser.makeStreamingKey(event.id, event.name)
 										} else {
@@ -4442,7 +4554,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 										}
 
 										const partialToolUse =
-											// Stryker disable next-line ConditionalExpression: equivalent mutant, streamingKey is always a string here (makeStreamingKey), so the null side is unreachable
 											streamingKey === undefined
 												? null
 												: NativeToolCallParser.processStreamingChunk(
@@ -4451,7 +4562,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 														nativeToolCallParserScope,
 													)
 
-										// Stryker disable next-line ConditionalExpression, EqualityOperator: equivalent mutants, streamingKey is always defined here, so both conditions reduce to the partialToolUse check
 										if (partialToolUse && streamingKey !== undefined) {
 											// Reuse the compound key built above: it is exactly the key the start event
 											// registered under, so the delta lookup needs no second encoding.
@@ -4870,7 +4980,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 					if (event.type === "tool_call_end") {
 						// Create compound key for deduplication (same pattern as streaming handler).
 						// End events carry the tool name; fall back to the id for events that don't.
-						// Stryker disable next-line LogicalOperator: equivalent mutant, finalizeRawChunks emits end events with a string name, so the ?? event.id fallback is unreachable
 						let eventName = event.name ?? event.id
 						let dedupKey = NativeToolCallParser.makeStreamingKey(event.id, eventName)
 
@@ -4887,13 +4996,11 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 						// name (provider quirk), the compound-key lookup above misses and both the
 						// parser entry and the dedup tracking would stay stale. Resolve the tracked
 						// entry by id and target the real compound key for finalization and cleanup.
-						// Stryker disable next-line ConditionalExpression, LogicalOperator, EqualityOperator: equivalent mutants, the both-miss case is unreachable via the raw-chunk pipeline
 						if (finalToolUse === null && toolUseIndex === undefined) {
 							const resolved = NativeToolCallParser.getStreamingToolCallById(
 								event.id,
 								nativeToolCallParserScope,
 							)
-							// Stryker disable next-line ConditionalExpression: unreachable defensive region — getStreamingToolCallById can only resolve when the both-miss case above is reachable, which the raw-chunk pipeline cannot produce (see the directive there).
 							if (resolved) {
 								eventName = resolved.name
 								dedupKey = NativeToolCallParser.makeStreamingKey(resolved.id, resolved.name)
@@ -5637,6 +5744,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 						parallelToolCalls: true,
 					}
 				: {}),
+			// Carry the active task-local effort override.
+			...this.getRuntimeThinkingEffortMetadata(),
 		}
 
 		try {
@@ -5888,6 +5997,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 							parallelToolCalls: true,
 						}
 					: {}),
+				// Carry the active task-local effort override.
+				...this.getRuntimeThinkingEffortMetadata(),
 			}
 
 			// Only generate environment details when context management will actually run.
@@ -6016,6 +6127,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 			taskId: this.taskId,
 			suppressPreviousResponseId: this.skipPrevResponseIdOnce,
 			abortSignal,
+			// Carry the active task-local effort override for this request.
+			...this.getRuntimeThinkingEffortMetadata(),
 			// Include tools whenever they are present.
 			...(shouldIncludeTools
 				? {