feat(api): per-request thinking effort override and adaptive effort envelope (DTE series 2/5)

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 {
@@ -236,7 +245,7 @@ export function buildApiHandler(configuration: ProviderSettings): ApiHandler {
 		case providerIdentifiers.poe:
 			return new PoeHandler(options)
 		case providerIdentifiers.geminiCli:
-			// Intentionally falls through to the Anthropic handler pending a dedicated Gemini CLI handler implementation.
+		// Intentionally falls through to the Anthropic handler pending a dedicated Gemini CLI handler implementation.
 		default:
 			return new AnthropicHandler(options)
 	}

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