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

@@ -57,8 +57,15 @@ 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,
 		// CI build/coverage tooling under src/scripts is excluded like the other build scripts:
 		// it is not shipped runtime code and is covered by src/scripts/__tests__/merge-lcov.spec.mjs.
 		excludedPaths: [

scripts/stryker-diff.test.mjs

@@ -254,8 +254,13 @@ 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)
+
 		const types = manifest.packages.find(({ id }) => id === "types")
 		assert.equal(types.discoverRelatedTests, undefined)
 		assert.equal(types.vitestRelated, undefined)

src/api/index.ts

@@ -7,6 +7,7 @@ import {
 	retiredProviderIdentifiers,
 	type ProviderSettings,
 	type ModelInfo,
+	type ReasoningEffortExtended,
 	type ModelRecord,
 } from "@roo-code/types"
 
@@ -117,6 +118,14 @@ 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
 }
 
 /**

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

@@ -34,6 +34,51 @@ export type AnthropicProviderReasoningParams =
 
 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