@oh-my-pi/pi-ai 18.2.0 → 18.2.2

package/src/providers/openai-reasoning-fallback.ts

@@ -187,7 +187,7 @@ function collectMessageParts(error: unknown, captured: CapturedHttpErrorResponse
  * the same rejection with `Supported values are: …`, so value lists count too.
  */
 const REASONING_EFFORT_FIELD_PATTERN =
-	/reasoning[_. ]effort|reasoning value|(?:valid|supported|allowed) (?:levels?|values?)/i;
+	/reasoning[_. ]?effort|reasoning value|(?:valid|supported|allowed) (?:levels?|values?)/i;
 
 function mentionsReasoningEffort(error: unknown, captured: CapturedHttpErrorResponse | undefined): boolean {
 	const param = capturedStringField(captured, "param");
@@ -224,17 +224,17 @@ interface EffortRejectionSignal {
 	rejectedMatches: boolean;
 }
 
-const EFFORT_FIELD_PATTERN = /reasoning[_. ]effort|reasoning value/i;
+const EFFORT_FIELD_PATTERN = /reasoning[_. ]?effort|reasoning value/i;
 const ALLOWED_LEVELS_PATTERN = /(?:valid|supported|allowed) levels?/i;
 
 /** Fielded rejection verdicts in any word order: verdict-first, field-first, or bare mention plus verdict. */
 function messageCarriesEffortVerdict(message: string): boolean {
 	return (
-		/invalid[^\n]*(?:reasoning[_. ]effort|reasoning value)/i.test(message) ||
-		/(?:reasoning[_. ]effort|reasoning value)[^\n]*(?:invalid|unsupported|not supported|not permitted|must be|expected|unknown|unexpected|unrecognized)/i.test(
+		/invalid[^\n]*(?:reasoning[_. ]?effort|reasoning value)/i.test(message) ||
+		/(?:reasoning[_. ]?effort|reasoning value)[^\n]*(?:invalid|unsupported|not supported|not permitted|must be|expected|unknown|unexpected|unrecognized)/i.test(
 			message,
 		) ||
-		/(?:unsupported|not supported|not permitted|unknown|unexpected|unrecognized|extra)[^\n]*(?:reasoning[_. ]effort|reasoning value)/i.test(
+		/(?:unsupported|not supported|not permitted|unknown|unexpected|unrecognized|extra)[^\n]*(?:reasoning[_. ]?effort|reasoning value)/i.test(
 			message,
 		)
 	);
@@ -367,7 +367,7 @@ function nearestEnabledReasoningFallback(currentEffort: string, allowed: Set<str
  * supported`).
  */
 const TEMPLATE_KWARG_EFFORT_PATTERN =
-	/chat_template_kwargs[^\n]{0,120}reasoning[_. ]effort|reasoning[_. ]effort[^\n]{0,120}chat_template_kwargs/i;
+	/chat_template_kwargs[^\n]{0,120}reasoning[_. ]?effort|reasoning[_. ]?effort[^\n]{0,120}chat_template_kwargs/i;
 const FIELD_REJECTION_PATTERN =
 	/invalid|unsupported|not supported|not permitted|unknown|unexpected|unrecognized|rejected|extra input/i;
 

package/src/providers/openai-responses-server.ts

@@ -43,7 +43,7 @@ import {
 	type OpenAIResponsesTool,
 	openaiResponsesRequestSchema,
 } from "./openai-responses-server-schema";
-import { encodeTextSignatureV1, parseTextSignature } from "./openai-shared";
+import { coerceNullMessageContentInPlace, encodeTextSignatureV1, parseTextSignature } from "./openai-shared";
 
 export type { ParsedRequest };
 
@@ -365,6 +365,7 @@ export function parseRequest(body: unknown, headers?: Headers): ParsedRequest {
 	// `resolvePromptCacheKey` call further down.
 
 	rejectUnsupportedExplicitPromptCacheFields(body);
+	coerceNullMessageContentInPlace(isObj(body) ? body.input : undefined);
 	const data = openaiResponsesRequestSchema(body);
 	if (data instanceof type.errors) {
 		throw new AIError.ValidationError(`openai-responses: ${data.summary}`);

package/src/providers/openai-responses.ts

@@ -295,6 +295,33 @@ function resetOpenAIResponsesChainState(state: OpenAIResponsesChainState): void
 	state.lastPromptCacheBreakpointPolicy = undefined;
 }
 
+/**
+ * Drop the account-bound half of every retained `openai-responses` record in
+ * `states`: the stateful `previous_response_id` chain baselines.
+ *
+ * Chaining stores the turn server-side under the account that created it, so a
+ * baseline minted by one credential is dead weight the moment the session is
+ * switched to a sibling account — the next delta request answers
+ * `Previous response not found` and burns a turn re-learning that. Everything
+ * else this record holds describes the *deployment*, not the account
+ * (strict-tools demotion, reasoning-effort fallback, native-history-replay
+ * warmup, the chaining circuit breaker), and is deliberately preserved:
+ * re-learning an endpoint's limits on every credential switch is the cost this
+ * state exists to avoid.
+ */
+export function resetOpenAIResponsesAccountScopedState(states: Map<string, ProviderSessionState>): void {
+	for (const [key, value] of states) {
+		if (!key.startsWith(OPENAI_RESPONSES_PROVIDER_SESSION_STATE_PREFIX)) continue;
+		const state = value as OpenAIResponsesProviderSessionState;
+		for (const chain of state.chains.values()) {
+			resetOpenAIResponsesChainState(chain);
+			// The stale-failure counter tallies the previous account's 404s; a
+			// fresh account must not inherit a tripped circuit breaker.
+			chain.staleFailures = 0;
+		}
+	}
+}
+
 interface OpenAIResponsesChainedParams {
 	params: OpenAIResponsesSamplingParams;
 	/** Set iff the params carry previous_response_id (delta request). */
@@ -422,6 +449,7 @@ const streamOpenAIResponsesOnce = (
 		let rawRequestDump: RawHttpRequestDump | undefined;
 		let chainState: OpenAIResponsesChainState | undefined;
 		let sentPreviousResponseId: string | undefined;
+		let lastSubmittedRequestWasFullReplay: boolean | undefined;
 		const abortTracker = createAbortSourceTracker(options?.signal);
 		const firstEventTimeoutAbortError = new AIError.StreamTimeoutError(OPENAI_RESPONSES_FIRST_EVENT_TIMEOUT_MESSAGE);
 		const { requestAbortController, requestSignal } = abortTracker;
@@ -552,6 +580,7 @@ const streamOpenAIResponsesOnce = (
 					typeof requestParams.model === "string" ? requestParams.model : model.id,
 				);
 				activeRequestParams = requestParams;
+				lastSubmittedRequestWasFullReplay = requestParams.previous_response_id === undefined;
 				let requestTimeout: NodeJS.Timeout | undefined;
 				if (requestTimeoutMs !== undefined) {
 					requestTimeout = setTimeout(
@@ -576,6 +605,9 @@ const streamOpenAIResponsesOnce = (
 							copilotCacheKey,
 							copilotCacheSnapshot,
 						),
+						shouldRetryResponse: (response, bodyText) =>
+							!AIError.isRequestBodyReadTimeout(response.status, bodyText) ||
+							lastSubmittedRequestWasFullReplay !== true,
 						// Transient 408/429/5xx get Retry-After-aware transport
 						// retries; the first-event watchdog aborts `requestSignal`,
 						// so retries cannot extend the caller's deadline.
@@ -878,7 +910,7 @@ const streamOpenAIResponsesOnce = (
 							: activeParams,
 					);
 					chainState.lastPromptCacheBreakpointPolicy = promptCacheBreakpointPolicy;
-					if (output.responseId) {
+					if (output.responseId && replayableResponseItems.length === nativeOutputItems.length) {
 						chainState.lastResponseId = output.responseId;
 						chainState.lastResponseItems = replayableResponseItems;
 						chainState.canAppend = true;
@@ -886,8 +918,12 @@ const streamOpenAIResponsesOnce = (
 						// full-context success must not mask categorical rejection.
 						if (sentPreviousResponseId) chainState.staleFailures = 0;
 					} else {
-						// Without a response id the append baseline cannot be trusted.
+						// No response id, or replay sanitization dropped an item the server
+						// still holds. Sanitization is 1:1-or-fewer, so either case makes the
+						// append baseline untrustworthy; next turn must replay in full.
 						chainState.canAppend = false;
+						chainState.lastResponseId = undefined;
+						chainState.lastResponseItems = undefined;
 					}
 				}
 			} else if (chainState) {
@@ -926,6 +962,9 @@ const streamOpenAIResponsesOnce = (
 			output.errorStatus = result.status;
 			output.errorId = result.id;
 			output.errorMessage = result.message;
+			if (AIError.isRequestBodyReadTimeout(result.status, result.message) && lastSubmittedRequestWasFullReplay) {
+				output.requestBodyReadTimeoutFullReplay = true;
+			}
 			// Some providers via OpenRouter include extra details here.
 			const rawMetadata = (error as { error?: { metadata?: { raw?: string } } })?.error?.metadata?.raw;
 			if (rawMetadata) output.errorMessage += `\n${rawMetadata}`;
@@ -1153,6 +1192,11 @@ export function buildParams(
 	});
 	const strictResponsesPairing = policy.tools.strictResponsesPairing;
 	const shouldReplayNativeHistory = providerSessionState?.nativeHistoryReplayWarmed ?? true;
+	// Filtering native reasoning must not be undone by reconstruction when the
+	// target also rejects synthetic items (Muse on OpenRouter). Unfiltered targets
+	// retain required text/placeholder replay, including DeepSeek's #10690 fallback.
+	const canReconstructReasoningReplay =
+		!policy.reasoning.filterReasoningHistory || policy.reasoning.allowsSyntheticReasoningContentForToolCalls;
 	const messages = buildResponsesInput({
 		model,
 		context,
@@ -1164,9 +1208,13 @@ export function buildParams(
 		},
 		includeThinkingSignatures: shouldReplayNativeHistory && !policy.reasoning.filterReasoningHistory,
 		requiresReasoningReplayForAllTurns:
-			policy.reasoning.enabled && policy.reasoning.requiresReasoningContentForAllAssistantTurns,
+			policy.reasoning.enabled &&
+			policy.reasoning.requiresReasoningContentForAllAssistantTurns &&
+			canReconstructReasoningReplay,
 		requiresReasoningReplayForToolCalls:
-			policy.reasoning.enabled && policy.reasoning.requiresReasoningContentForToolCalls,
+			policy.reasoning.enabled &&
+			policy.reasoning.requiresReasoningContentForToolCalls &&
+			canReconstructReasoningReplay,
 		repairOrphanOutputs: true,
 	});