fix(historian): stop sending temperature to reasoning models

CONFIGURATION.md

@@ -663,7 +663,7 @@ It is useful when starting a new session. It's better to choose a fast and cheap
 |-------|------|-------------|
 | `model` | `string` | Primary model. |
 | `fallback_models` | `string` or `string[]` | Fallback models. |
-| `temperature` | `number` (0–2) | Sampling temperature. |
+| `temperature` | `number` (0–2) | Sampling temperature. Omitted by default: no `temperature` is sent to the provider unless you set one. Reasoning models reject the parameter outright, so only set it for models that accept it. |
 | `variant` | `string` | **OpenCode only.** Agent variant — selects a thinking/reasoning preset. Pi uses `thinking_level` instead. |
 | `thinking_level` | `string` | **Pi only.** Explicit reasoning level (`off`/`low`/`medium`/`high`) passed to Pi for sidekick subagent runs. See `historian.thinking_level`. |
 | `prompt` | `string` | Persistent agent-level system prompt override. Applies to every sidekick run. |

packages/pi-plugin/src/historian-calibration-extension.test.ts

@@ -42,4 +42,56 @@ describe("historian provider calibration", () => {
 			),
 		).toEqual({ inferenceConfig: { temperature: 0.1, maxTokens: 32_000 } });
 	});
+
+	it("never sends temperature when it is not configured", () => {
+		const fixtures = [
+			{ input: { max_tokens: 4096 }, key: "max_tokens" },
+			{ input: { max_completion_tokens: 4096 }, key: "max_completion_tokens" },
+			{ input: { max_output_tokens: 4096 }, key: "max_output_tokens" },
+			{ input: { maxTokens: 4096 }, key: "maxTokens" },
+		] as const;
+		for (const fixture of fixtures) {
+			const result = calibrateHistorianProviderPayload(
+				fixture.input,
+				undefined,
+				32_000,
+			) as Record<string, unknown>;
+			expect("temperature" in result).toBe(false);
+			expect(result[fixture.key]).toBe(32_000);
+		}
+	});
+
+	it("omits temperature from nested provider shapes when unconfigured", () => {
+		expect(
+			calibrateHistorianProviderPayload(
+				{ generationConfig: { topP: 0.9, maxOutputTokens: 4096 } },
+				undefined,
+				32_000,
+			),
+		).toEqual({ generationConfig: { topP: 0.9, maxOutputTokens: 32_000 } });
+		expect(
+			calibrateHistorianProviderPayload(
+				{ inferenceConfig: { maxTokens: 4096 } },
+				undefined,
+				32_000,
+			),
+		).toEqual({ inferenceConfig: { maxTokens: 32_000 } });
+	});
+
+	it("applies temperature alone when no output budget is configured", () => {
+		const result = calibrateHistorianProviderPayload(
+			{ max_tokens: 4096 },
+			0.1,
+			undefined,
+		) as Record<string, unknown>;
+		expect(result.temperature).toBe(0.1);
+		expect(result.max_tokens).toBe(4096);
+	});
+
+	it("returns the payload untouched when neither knob is configured", () => {
+		const payload = { max_tokens: 4096 };
+		expect(
+			calibrateHistorianProviderPayload(payload, undefined, undefined),
+		).toEqual({ max_tokens: 4096 });
+	});
 });

packages/pi-plugin/src/historian-calibration-extension.ts

@@ -10,14 +10,22 @@ function finiteNumber(value: string | undefined): number | undefined {
 	return Number.isFinite(parsed) && parsed >= 0 ? parsed : undefined;
 }
 
-/** Apply the historian calibration to each provider's serialized request shape. */
+/**
+ * Apply the historian calibration to each provider's serialized request shape.
+ *
+ * Both knobs are optional and applied independently: reasoning models reject
+ * `temperature` outright, so an output-token budget must still be applicable
+ * on its own.
+ */
 export function calibrateHistorianProviderPayload(
 	payload: unknown,
-	temperature: number,
-	maxOutputTokens: number,
+	temperature: number | undefined,
+	maxOutputTokens: number | undefined,
 ): unknown {
 	if (typeof payload !== "object" || payload === null || Array.isArray(payload))
 		return payload;
+	if (temperature === undefined && maxOutputTokens === undefined)
+		return payload;
 	const calibrated = { ...(payload as Record<string, unknown>) };
 	const generationConfig = calibrated.generationConfig;
 	if (
@@ -27,8 +35,8 @@ export function calibrateHistorianProviderPayload(
 	) {
 		calibrated.generationConfig = {
 			...(generationConfig as Record<string, unknown>),
-			temperature,
-			maxOutputTokens,
+			...(temperature !== undefined ? { temperature } : {}),
+			...(maxOutputTokens !== undefined ? { maxOutputTokens } : {}),
 		};
 		return calibrated;
 	}
@@ -40,21 +48,25 @@ export function calibrateHistorianProviderPayload(
 	) {
 		calibrated.inferenceConfig = {
 			...(inferenceConfig as Record<string, unknown>),
-			temperature,
-			maxTokens: maxOutputTokens,
+			...(temperature !== undefined ? { temperature } : {}),
+			...(maxOutputTokens !== undefined ? { maxTokens: maxOutputTokens } : {}),
 		};
 		return calibrated;
 	}
 
-	calibrated.temperature = temperature;
-	if ("max_output_tokens" in calibrated) {
-		calibrated.max_output_tokens = maxOutputTokens;
-	} else if ("max_completion_tokens" in calibrated) {
-		calibrated.max_completion_tokens = maxOutputTokens;
-	} else if ("max_tokens" in calibrated) {
-		calibrated.max_tokens = maxOutputTokens;
-	} else if ("maxTokens" in calibrated) {
-		calibrated.maxTokens = maxOutputTokens;
+	if (temperature !== undefined) {
+		calibrated.temperature = temperature;
+	}
+	if (maxOutputTokens !== undefined) {
+		if ("max_output_tokens" in calibrated) {
+			calibrated.max_output_tokens = maxOutputTokens;
+		} else if ("max_completion_tokens" in calibrated) {
+			calibrated.max_completion_tokens = maxOutputTokens;
+		} else if ("max_tokens" in calibrated) {
+			calibrated.max_tokens = maxOutputTokens;
+		} else if ("maxTokens" in calibrated) {
+			calibrated.maxTokens = maxOutputTokens;
+		}
 	}
 	return calibrated;
 }
@@ -64,7 +76,7 @@ export default function historianCalibrationExtension(pi: ExtensionAPI): void {
 	const maxOutputTokens = finiteNumber(
 		process.env[HISTORIAN_MAX_OUTPUT_TOKENS_ENV],
 	);
-	if (temperature === undefined || maxOutputTokens === undefined) return;
+	if (temperature === undefined && maxOutputTokens === undefined) return;
 	pi.on("before_provider_request", (event) =>
 		calibrateHistorianProviderPayload(
 			event.payload,

packages/pi-plugin/src/index.ts

@@ -703,7 +703,7 @@ export function resolveHistorianFromConfig(
 		fallbackModels,
 		historianChunkTokens,
 		timeoutMs: config.historian_timeout_ms,
-		temperature: historian?.temperature ?? 0.1,
+		temperature: historian?.temperature,
 		maxOutputTokens: historian?.maxTokens ?? 32_000,
 		// `historian.two_pass` runs an editor pass after a successful
 		// first pass to clean low-signal U: lines and cross-compartment

packages/pi-plugin/src/pi-historian-runner.ts

@@ -444,7 +444,7 @@ export async function runPiHistorian(deps: PiHistorianDeps): Promise<void> {
 		refreshBoundarySnapshot,
 		currentContextLimit,
 		historianTimeoutMs = DEFAULT_HISTORIAN_TIMEOUT_MS,
-		temperature = 0.1,
+		temperature,
 		maxOutputTokens = 32_000,
 		signal,
 		retryBackoffMs,

packages/pi-plugin/src/subagent-runner.test.ts

@@ -1558,6 +1558,34 @@ describe("PiSubagentRunner spawn lifecycle", () => {
 		});
 	});
 
+	it("surfaces the provider error behind empty assistant text", async () => {
+		const child = createMockChild();
+		const { runner } = runnerWith(child);
+
+		const resultPromise = runner.run(baseOptions);
+		child.writeStdoutLine(
+			agentEnd([
+				{
+					role: "assistant",
+					content: [],
+					stopReason: "error",
+					errorMessage:
+						"OpenAI API error (400): Unsupported parameter: temperature",
+				},
+			]),
+		);
+		child.emitClose(0);
+
+		expect(await resultPromise).toEqual({
+			ok: false,
+			reason: "no_assistant",
+			error:
+				"pi assistant produced empty text (provider error: OpenAI API error (400): Unsupported parameter: temperature)",
+			durationMs: expect.any(Number),
+			meta: { stderr: undefined, sawProtocolOutput: true },
+		});
+	});
+
 	it("returns no_assistant for empty stdout and successful exit", async () => {
 		// Issue #238: an empty-stdout exit-0 primary now fires the one-shot
 		// isolated retry. When the isolated attempt ALSO exits 0 with no output,

packages/pi-plugin/src/subagent-runner.ts

@@ -1378,13 +1378,16 @@ export class PiSubagentRunner implements SubagentRunner {
 						trimmedAssistantText === null ||
 						trimmedAssistantText.length === 0
 					) {
+						const emptyAssistantReason =
+							trimmedAssistantText === null
+								? "pi agent_end did not include an assistant message"
+								: "pi assistant produced empty text";
 						settle({
 							ok: false,
 							reason: "no_assistant",
-							error:
-								trimmedAssistantText === null
-									? "pi agent_end did not include an assistant message"
-									: "pi assistant produced empty text",
+							error: finalErrorMessage
+								? `${emptyAssistantReason} (provider error: ${finalErrorMessage})`
+								: emptyAssistantReason,
 							durationMs: Date.now() - startTime,
 							// Pi machinery worked (agent_end / terminal message_end seen);
 							// the model just returned empty text. Mark protocol output as