@oh-my-pi/pi-ai 18.14.2 → 18.15.0

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## [Unreleased]
 
+## [18.15.0] - 2026-11-01
+
+### Fixed
+
+- Fixed Codex conversation replay failing with an unknown `internal_chat_message_metadata_passthrough.content_token_count` parameter by omitting backend-internal response metadata from replayed input ([#17590](https://github.com/can1357/oh-my-pi/pull/17590), [#17597](https://github.com/can1357/oh-my-pi/pull/17597) by [@briankeefe](https://github.com/briankeefe)).
+- Fixed `service_tier: "ultrafast"` being dropped for OpenAI-family models on custom providers even when the model's `serviceTiers` lists it ([#17577](https://github.com/can1357/oh-my-pi/pull/17577) by [@jakehemmerle](https://github.com/jakehemmerle))
+- Fixed Google Vertex requests with Workforce Identity Federation ADC failing token refresh against the wrong endpoint ([#17447](https://github.com/can1357/oh-my-pi/issues/17447)).
+- Fixed Anthropic (and other rotating-token) OAuth accounts being signed out automatically when a background usage poll's token refresh got `invalid_grant`, including refreshes a client delegates to an auth broker; usage polling no longer disables credentials, so only a real request's refresh can do that ([#17468](https://github.com/can1357/oh-my-pi/issues/17468))
+
 ## [18.14.2] - 2026-10-31
 
 ### Breaking Changes

package/dist/types/auth/types.d.ts

@@ -279,8 +279,13 @@ export interface CredentialDisabledEvent {
     orgId?: string;
     orgName?: string;
 }
-/** Why a delegated OAuth refresh was requested. */
-export type OAuthRefreshReason = "auth-recovery";
+/**
+ * Why an OAuth refresh was requested.
+ * - `auth-recovery`: a provider rejected the bearer with a 401.
+ * - `usage-probe`: a background usage poll; advisory, so a definitive refresh
+ *   failure never disables the credential (an interactive request decides).
+ */
+export type OAuthRefreshReason = "auth-recovery" | "usage-probe";
 /** Configuration supplied when constructing credential storage. */
 export type AuthStorageOptions = {
     usageProviderResolver?: (provider: Provider) => UsageProvider | undefined;
@@ -979,9 +984,10 @@ export interface OAuthRefreshByIdOptions {
      */
     reuseRecentMint?: boolean;
     /**
-     * Provider-401 recovery intent. Forwarded to a delegated refresh (an auth
-     * broker), which may then return a token it minted in the last five minutes.
-     * Local reuse still requires {@link reuseRecentMint}.
+     * Why the refresh was requested. `auth-recovery` is forwarded to a delegated
+     * refresh (an auth broker), which may then return a token it minted in the
+     * last five minutes; local reuse still requires {@link reuseRecentMint}.
+     * `usage-probe` never disables the row on a definitive refresh failure.
      */
     reason?: OAuthRefreshReason;
     /** The access token a provider 401 rejected; see {@link AuthApiKeyOptions.rejectedAccessToken}. */

package/dist/types/providers/anthropic.d.ts

@@ -17,6 +17,11 @@ export * from "./claude-code-fingerprint.js";
  * pass through untouched, so installing it on every OAuth flow is safe.
  */
 export declare function wrapFetchForCch(base: FetchImpl): FetchImpl;
+/**
+ * Whether a request to `model` sends the tool `name` for the host to run itself, never dispatching the call to the
+ * client: the same decision `encodeAnthropicToolName` makes on the wire.
+ */
+export declare function anthropicHostRunsTool(model: Model<"anthropic-messages">, name: string, headers?: Record<string, string>): boolean;
 export type AnthropicEffort = AnthropicOutputEffort | "adaptive";
 export type AnthropicThinkingDisplay = "summarized" | "omitted";
 export interface AnthropicOptions extends StreamOptions {

package/dist/types/providers/google-auth.d.ts

@@ -3,10 +3,9 @@
  *
  * Replaces `google-auth-library` with a direct WebCrypto + REST implementation.
  * Sources, in priority order:
- *   1. `GOOGLE_APPLICATION_CREDENTIALS` env → file with `type: "service_account"` (RS256 JWT exchange)
- *     or `type: "authorized_user"` (refresh-token exchange).
+ *   1. `GOOGLE_APPLICATION_CREDENTIALS` env → service account, user, or workforce ADC file.
  *   2. gcloud user ADC: `%APPDATA%\gcloud\application_default_credentials.json` on Windows,
- *     `~/.config/gcloud/application_default_credentials.json` elsewhere (authorized_user flow).
+ *     `~/.config/gcloud/application_default_credentials.json` elsewhere.
  *   3. GCE / Cloud Run metadata server (`metadata.google.internal`).
  *
  * Tokens are cached per source key and refreshed `GOOGLE_VERTEX_REFRESH_SKEW_MS` before expiry

package/dist/types/providers/openai-codex/request-transformer.d.ts

@@ -31,6 +31,7 @@ export interface InputItem {
     type?: string | null;
     role?: string;
     content?: unknown;
+    internal_chat_message_metadata_passthrough?: unknown;
     call_id?: string | null;
     name?: string;
     output?: unknown;

package/dist/types/types.d.ts

@@ -144,8 +144,10 @@ export declare function resolveModelServiceTier(tiers: ServiceTierByFamily | nul
  * `service_tier_for_request`); an empty or missing list counts as "not
  * reported" — accounts whose `/models` lists no tiers keep `/fast` — so the
  * provider-level answer stands. `flex` and `default` are never gated.
- * First-party OpenAI takes `ultrafast` as-is. A bare provider string cannot
- * carry the list, so it answers for the provider alone.
+ * First-party OpenAI takes `ultrafast` as-is. An OpenAI-family relay on the
+ * Responses/Completions APIs gets `ultrafast` only when its discovered
+ * `service_tiers` lists it. A bare provider string cannot carry the list, so
+ * it answers for the provider alone.
  */
 export declare function shouldSendServiceTier(serviceTier: ServiceTier | null | undefined, target: Provider | ServiceTierModel | undefined): boolean;
 /**

package/dist/types/utils.d.ts

@@ -34,6 +34,11 @@ export declare function stripOpenAIResponsesOutputOnlyStatusesForReplay<TItem ex
     type?: unknown;
     status?: unknown;
 }>(items: TItem[]): TItem[];
+/** Remove Codex-rejected backend metadata; preserve clean-item identity by default, or copy while keeping/dropping id. */
+export declare function stripOpenAIResponsesInternalMetadataForReplay<TItem extends {
+    id?: unknown;
+    internal_chat_message_metadata_passthrough?: unknown;
+}>(item: TItem, mode?: "preserve" | "copy" | "drop-id"): TItem;
 /** Clamp replayed images without changing item identities or other native replay fields. */
 export declare function clampOpenAIResponsesImageDetailForReplay<TItem extends {
     type?: unknown;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-ai",
-	"version": "18.14.2",
+	"version": "18.15.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"keywords": [
 		"ai",
@@ -155,11 +155,11 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.14.2",
-		"@oh-my-pi/pi-catalog": "18.14.2",
-		"@oh-my-pi/pi-natives": "18.14.2",
-		"@oh-my-pi/pi-utils": "18.14.2",
-		"@oh-my-pi/pi-wire": "18.14.2"
+		"@oh-my-pi/omptype": "18.15.0",
+		"@oh-my-pi/pi-catalog": "18.15.0",
+		"@oh-my-pi/pi-natives": "18.15.0",
+		"@oh-my-pi/pi-utils": "18.15.0",
+		"@oh-my-pi/pi-wire": "18.15.0"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/auth/refresh.ts

@@ -1006,7 +1006,9 @@ export class OAuthRefresher {
 				// `refreshStored` owns the caller budget and the exchange backstop.
 				refresh: (current, refreshSignal) =>
 					this.#requestOAuthCredentialRefresh(provider, current, credentialId, refreshSignal, reason, undefined),
-				isDefinitiveFailure: error => AIError.isDefinitiveOAuthFailure(String(error)),
+				// A usage probe is advisory: leave a dead grant for an interactive request to disable.
+				isDefinitiveFailure:
+					reason === "usage-probe" ? undefined : error => AIError.isDefinitiveOAuthFailure(String(error)),
 				disabledCause: error => `oauth refresh failed: ${String(error)}`,
 			});
 			if (result.credential) {
@@ -1199,7 +1201,8 @@ export class OAuthRefresher {
 				// A definitively-dead grant tears the row down here, where the
 				// attempted credential is known. CAS on the persisted credential so a
 				// peer/login rotation in flight leaves the freshly-rotated row intact.
-				if (AIError.isDefinitiveOAuthFailure(String(error))) {
+				// A usage probe is advisory: leave a dead grant for an interactive request to disable.
+				if (reason !== "usage-probe" && AIError.isDefinitiveOAuthFailure(String(error))) {
 					// CAS-loss (false) means a peer/login rotated the row mid-refresh, so
 					// our #data copy is stale — reload so the next caller serves the
 					// freshly-rotated credential rather than the dead token we attempted.

package/src/auth/types.ts

@@ -329,8 +329,13 @@ export interface CredentialDisabledEvent {
 	orgName?: string;
 }
 
-/** Why a delegated OAuth refresh was requested. */
-export type OAuthRefreshReason = "auth-recovery";
+/**
+ * Why an OAuth refresh was requested.
+ * - `auth-recovery`: a provider rejected the bearer with a 401.
+ * - `usage-probe`: a background usage poll; advisory, so a definitive refresh
+ *   failure never disables the credential (an interactive request decides).
+ */
+export type OAuthRefreshReason = "auth-recovery" | "usage-probe";
 
 /** Configuration supplied when constructing credential storage. */
 export type AuthStorageOptions = {
@@ -1064,9 +1069,10 @@ export interface OAuthRefreshByIdOptions {
 	 */
 	reuseRecentMint?: boolean;
 	/**
-	 * Provider-401 recovery intent. Forwarded to a delegated refresh (an auth
-	 * broker), which may then return a token it minted in the last five minutes.
-	 * Local reuse still requires {@link reuseRecentMint}.
+	 * Why the refresh was requested. `auth-recovery` is forwarded to a delegated
+	 * refresh (an auth broker), which may then return a token it minted in the
+	 * last five minutes; local reuse still requires {@link reuseRecentMint}.
+	 * `usage-probe` never disables the row on a definitive refresh failure.
 	 */
 	reason?: OAuthRefreshReason;
 	/** The access token a provider 401 rejected; see {@link AuthApiKeyOptions.rejectedAccessToken}. */

package/src/auth/usage.ts

@@ -232,6 +232,7 @@ export class UsageService implements UsageApi {
 						refreshableCredential,
 						refreshableCredentialId,
 						timeoutSignal,
+						"usage-probe",
 					);
 					const refreshedCredential = mergeRefreshedUsageCredential(request.credential, refreshed);
 					this.persistRefreshedCredential(

package/src/auth-broker/client.ts

@@ -327,7 +327,7 @@ export class AuthBrokerClient {
 		signal?: AbortSignal,
 		reason?: OAuthRefreshReason,
 	): Promise<CredentialRefreshResponse> {
-		const suffix = reason === "auth-recovery" ? "?reason=auth-recovery" : "";
+		const suffix = reason ? `?reason=${reason}` : "";
 		return this.#request<CredentialRefreshResponse>("POST", `/v1/credential/${id}/refresh${suffix}`, {
 			schema: "credentialRefreshResponseSchema",
 			signal,

package/src/auth-broker/server.ts

@@ -53,6 +53,7 @@ import {
 
 const DEFAULT_EXTERNAL_CHANGE_POLL_MS = 250;
 const AUTH_RECOVERY_REFRESH_OPTIONS = { reuseRecentMint: true } as const;
+const USAGE_PROBE_REFRESH_OPTIONS = { reason: "usage-probe" } as const;
 const TEXT_ENCODER = new TextEncoder();
 const SSE_KEEPALIVE = TEXT_ENCODER.encode(": keepalive\n\n");
 
@@ -935,12 +936,16 @@ export function startAuthBroker(opts: AuthBrokerServerOptions): AuthBrokerServer
 				const refreshMatch = req.method === "POST" ? pathname.match(REFRESH_ROUTE) : null;
 				if (refreshMatch) {
 					const id = Number.parseInt(refreshMatch[1], 10);
-					const authRecovery = url.searchParams.get("reason") === "auth-recovery";
+					const reason = url.searchParams.get("reason");
 					try {
 						const entry = await opts.storage.oauth.refresh(
 							id,
 							req.signal,
-							authRecovery ? AUTH_RECOVERY_REFRESH_OPTIONS : undefined,
+							reason === "auth-recovery"
+								? AUTH_RECOVERY_REFRESH_OPTIONS
+								: reason === "usage-probe"
+									? USAGE_PROBE_REFRESH_OPTIONS
+									: undefined,
 						);
 						const body: CredentialRefreshResponse = { entry };
 						logger.info("auth-broker credential refreshed", {

package/src/providers/anthropic.ts

@@ -652,6 +652,19 @@ function getUmansWebSearchHeader(
 	return provider ? { [UMANS_WEBSEARCH_PROVIDER_HEADER]: provider } : undefined;
 }
 
+/**
+ * Whether a request to `model` sends the tool `name` for the host to run itself, never dispatching the call to the
+ * client: the same decision `encodeAnthropicToolName` makes on the wire.
+ */
+export function anthropicHostRunsTool(
+	model: Model<"anthropic-messages">,
+	name: string,
+	headers?: Record<string, string>,
+): boolean {
+	const gatewayWebSearch = getUmansWebSearchHeader(model, mergeHeaders(model.headers, headers)) !== undefined;
+	return shouldUseUmansGatewayWebSearch(name, gatewayWebSearch);
+}
+
 function shouldUseUmansGatewayWebSearch(name: string, enabled: boolean): boolean {
 	return enabled && name.toLowerCase() === UMANS_WEBSEARCH_TOOL_NAME;
 }

package/src/providers/google-auth.ts

@@ -3,10 +3,9 @@
  *
  * Replaces `google-auth-library` with a direct WebCrypto + REST implementation.
  * Sources, in priority order:
- *   1. `GOOGLE_APPLICATION_CREDENTIALS` env → file with `type: "service_account"` (RS256 JWT exchange)
- *     or `type: "authorized_user"` (refresh-token exchange).
+ *   1. `GOOGLE_APPLICATION_CREDENTIALS` env → service account, user, or workforce ADC file.
  *   2. gcloud user ADC: `%APPDATA%\gcloud\application_default_credentials.json` on Windows,
- *     `~/.config/gcloud/application_default_credentials.json` elsewhere (authorized_user flow).
+ *     `~/.config/gcloud/application_default_credentials.json` elsewhere.
  *   3. GCE / Cloud Run metadata server (`metadata.google.internal`).
  *
  * Tokens are cached per source key and refreshed `GOOGLE_VERTEX_REFRESH_SKEW_MS` before expiry
@@ -45,6 +44,14 @@ interface AuthorizedUserCredentials {
 	refresh_token: string;
 }
 
+interface ExternalAccountAuthorizedUserCredentials {
+	type: "external_account_authorized_user";
+	client_id: string;
+	client_secret: string;
+	refresh_token: string;
+	token_url: string;
+}
+
 interface ImpersonatedServiceAccountCredentials {
 	type: "impersonated_service_account";
 	service_account_impersonation_url: string;
@@ -52,16 +59,28 @@ interface ImpersonatedServiceAccountCredentials {
 	delegates?: string[];
 }
 
-type AdcFileCredentials = ServiceAccountCredentials | AuthorizedUserCredentials | ImpersonatedServiceAccountCredentials;
+type AdcFileCredentials =
+	| ServiceAccountCredentials
+	| AuthorizedUserCredentials
+	| ExternalAccountAuthorizedUserCredentials
+	| ImpersonatedServiceAccountCredentials;
 
 interface TokenResponse {
 	access_token: string;
 	expires_in: number;
 	token_type?: string;
+	/** Replacement refresh token; STS may rotate workforce refresh tokens on exchange. */
+	refresh_token?: string;
 }
 
 const tokenCache = new Map<string, CachedToken>();
 const inflight = new Map<string, Promise<string>>();
+/**
+ * Latest workforce refresh token, keyed by the token stored in the ADC file. The file is never
+ * rewritten here when STS rotates the token, so later refreshes must use the replacement;
+ * a re-login writes a new file token and bypasses the stale entry.
+ */
+const rotatedRefreshTokens = new Map<string, string>();
 
 function getRefreshSkewMs(): number {
 	return $envpos("GOOGLE_VERTEX_REFRESH_SKEW_MS", 60_000);
@@ -168,6 +187,19 @@ async function exchangeRefreshToken(
 	return postForToken(OAUTH_TOKEN_URL, body, signal, fetchImpl);
 }
 
+async function exchangeWorkforceRefreshToken(
+	creds: ExternalAccountAuthorizedUserCredentials,
+	signal: AbortSignal | undefined,
+	fetchImpl: FetchImpl,
+): Promise<TokenResponse> {
+	const refreshToken = rotatedRefreshTokens.get(creds.refresh_token) ?? creds.refresh_token;
+	const body = new URLSearchParams({ grant_type: "refresh_token", refresh_token: refreshToken });
+	const authorization = `Basic ${Buffer.from(`${creds.client_id}:${creds.client_secret}`, "utf8").toString("base64")}`;
+	const token = await postForToken(creds.token_url, body, signal, fetchImpl, { Authorization: authorization });
+	if (token.refresh_token) rotatedRefreshTokens.set(creds.refresh_token, token.refresh_token);
+	return token;
+}
+
 async function fetchMetadataToken(
 	signal: AbortSignal | undefined,
 	fetchImpl: FetchImpl,
@@ -192,10 +224,11 @@ async function postForToken(
 	body: URLSearchParams,
 	signal: AbortSignal | undefined,
 	fetchImpl: FetchImpl,
+	extraHeaders?: Record<string, string>,
 ): Promise<TokenResponse> {
 	const response = await fetchImpl(url, {
 		method: "POST",
-		headers: { "Content-Type": "application/x-www-form-urlencoded" },
+		headers: { "Content-Type": "application/x-www-form-urlencoded", ...extraHeaders },
 		body: body.toString(),
 		signal,
 	});
@@ -217,6 +250,7 @@ async function resolveAccessTokenUncached(
 	const adc = await loadAdcCredentials();
 	if (adc) {
 		const creds = adc.creds;
+		const credentialType: string = creds.type;
 		let token: TokenResponse;
 
 		if (creds.type === "impersonated_service_account") {
@@ -228,10 +262,16 @@ async function resolveAccessTokenUncached(
 				throw new RangeError(`Cannot extract target principal from ${creds.service_account_impersonation_url}`);
 			}
 
-			const sourceToken =
-				creds.source_credentials.type === "service_account"
-					? await exchangeJwtForToken(creds.source_credentials, signal, fetchImpl)
-					: await exchangeRefreshToken(creds.source_credentials, signal, fetchImpl);
+			const sourceCredentials = creds.source_credentials;
+			const sourceType: string = sourceCredentials.type;
+			let sourceToken: TokenResponse;
+			if (sourceCredentials.type === "service_account") {
+				sourceToken = await exchangeJwtForToken(sourceCredentials, signal, fetchImpl);
+			} else if (sourceCredentials.type === "authorized_user") {
+				sourceToken = await exchangeRefreshToken(sourceCredentials, signal, fetchImpl);
+			} else {
+				throw new AIError.ConfigurationError(`Unsupported ADC credential type: ${sourceType}`);
+			}
 
 			const response = await fetchImpl(
 				`https://iamcredentials.googleapis.com/v1/projects/-/serviceAccounts/${targetPrincipal}:generateAccessToken`,
@@ -260,11 +300,14 @@ async function resolveAccessTokenUncached(
 			const data = (await response.json()) as { accessToken: string; expireTime: string };
 			const expiresIn = Math.max(0, Math.floor((new Date(data.expireTime).getTime() - Date.now()) / 1000));
 			token = { access_token: data.accessToken, expires_in: expiresIn, token_type: "Bearer" };
+		} else if (creds.type === "service_account") {
+			token = await exchangeJwtForToken(creds, signal, fetchImpl);
+		} else if (creds.type === "authorized_user") {
+			token = await exchangeRefreshToken(creds, signal, fetchImpl);
+		} else if (creds.type === "external_account_authorized_user") {
+			token = await exchangeWorkforceRefreshToken(creds, signal, fetchImpl);
 		} else {
-			token =
-				creds.type === "service_account"
-					? await exchangeJwtForToken(creds, signal, fetchImpl)
-					: await exchangeRefreshToken(creds, signal, fetchImpl);
+			throw new AIError.ConfigurationError(`Unsupported ADC credential type: ${credentialType}`);
 		}
 		return { source: adc.source, token };
 	}
@@ -334,4 +377,5 @@ export async function getVertexAccessToken(options?: { signal?: AbortSignal; fet
 export function __resetVertexTokenCache(): void {
 	tokenCache.clear();
 	inflight.clear();
+	rotatedRefreshTokens.clear();
 }

package/src/providers/openai-codex/request-transformer.ts

@@ -2,6 +2,7 @@ import { Effort } from "@oh-my-pi/pi-catalog/effort";
 import { requireSupportedEffort } from "@oh-my-pi/pi-catalog/model-thinking";
 import { $env } from "@oh-my-pi/pi-utils";
 import type { Model } from "../../types";
+import { stripOpenAIResponsesInternalMetadataForReplay } from "../../utils";
 import { mapOpenAIReasoningEffort } from "../openai-shared";
 
 /** Reasoning replay scope for the Codex Responses API (`reasoning.context`). */
@@ -50,6 +51,7 @@ export interface InputItem {
 	type?: string | null;
 	role?: string;
 	content?: unknown;
+	internal_chat_message_metadata_passthrough?: unknown;
 	call_id?: string | null;
 	name?: string;
 	output?: unknown;
@@ -179,14 +181,12 @@ function filterInput(input: InputItem[] | undefined): InputItem[] | undefined {
 
 	return input
 		.filter(item => item.type !== "item_reference")
-		.map(item => {
-			if (item.type === "computer_call") return item;
-			if (item.id != null) {
-				const { id: _id, ...rest } = item;
-				return rest as InputItem;
-			}
-			return item;
-		});
+		.map(item =>
+			stripOpenAIResponsesInternalMetadataForReplay(
+				item,
+				item.type !== "computer_call" && item.id != null ? "drop-id" : "preserve",
+			),
+		);
 }
 
 const CODEX_ORPHAN_OUTPUT_LIMIT = 16_000;

package/src/types.ts

@@ -240,8 +240,10 @@ export function resolveModelServiceTier(
  * `service_tier_for_request`); an empty or missing list counts as "not
  * reported" — accounts whose `/models` lists no tiers keep `/fast` — so the
  * provider-level answer stands. `flex` and `default` are never gated.
- * First-party OpenAI takes `ultrafast` as-is. A bare provider string cannot
- * carry the list, so it answers for the provider alone.
+ * First-party OpenAI takes `ultrafast` as-is. An OpenAI-family relay on the
+ * Responses/Completions APIs gets `ultrafast` only when its discovered
+ * `service_tiers` lists it. A bare provider string cannot carry the list, so
+ * it answers for the provider alone.
  */
 export function shouldSendServiceTier(
 	serviceTier: ServiceTier | null | undefined,
@@ -260,7 +262,11 @@ export function shouldSendServiceTier(
 		if (advertised !== undefined && advertised.length > 0) return advertised.includes(serviceTier);
 	}
 	if (serviceTier === "ultrafast") {
-		return provider === "openai" || (typeof target === "string" && provider === "openai-codex");
+		if (typeof target === "string") return provider === "openai" || provider === "openai-codex";
+		if (provider === "openai") return true;
+		return (
+			target !== undefined && isOpenAIServiceTierModel(target) && target.serviceTiers?.includes(serviceTier) === true
+		);
 	}
 	if (provider === "openai" || provider === "openai-codex") return true;
 	if (provider === "openrouter") {

package/src/utils.ts

@@ -111,6 +111,17 @@ export function stripOpenAIResponsesOutputOnlyStatusesForReplay<TItem extends {
 	return sanitized ?? items;
 }
 
+/** Remove Codex-rejected backend metadata; preserve clean-item identity by default, or copy while keeping/dropping id. */
+export function stripOpenAIResponsesInternalMetadataForReplay<
+	TItem extends { id?: unknown; internal_chat_message_metadata_passthrough?: unknown },
+>(item: TItem, mode: "preserve" | "copy" | "drop-id" = "preserve"): TItem {
+	if (mode === "preserve" && !Object.hasOwn(item, "internal_chat_message_metadata_passthrough")) return item;
+	const sanitized = { ...item };
+	delete sanitized.internal_chat_message_metadata_passthrough;
+	if (mode === "drop-id") delete sanitized.id;
+	return sanitized;
+}
+
 /**
  * Clamp `detail: "original"` only where Responses input_image parts live —
  * top-level items and `message.content[]`. Avoids a deep tree walk/clone of
@@ -614,8 +625,10 @@ function sanitizeOpenAIResponsesHistoryItemForReplay(
 	if (item.type === "reasoning") {
 		return sanitizeOpenAIResponsesReasoningItemForReplay(item, preserveReasoningItemIds);
 	}
-	const { id: _id, ...sanitizedItem } = item;
-	if (item.type === "computer_call" && typeof item.id === "string") sanitizedItem.id = item.id;
+	const sanitizedItem = stripOpenAIResponsesInternalMetadataForReplay(
+		item,
+		item.type === "computer_call" && typeof item.id === "string" ? "copy" : "drop-id",
+	);
 	if (item.type === "function_call") sanitizedItem.arguments = replayFunctionCallArguments(item);
 
 	return clampReplayItemImageDetail(