diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 6f9ef27f5..ef8be8105 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -1,12 +1,15 @@
 # Changelog
 
 ## [Unreleased]
+
 ### Added
 
 - Added `CheckCredentialsOptions.completionProbe` (and `completionTimeoutMs`) so `AuthStorage.checkCredentials` can additionally exercise each credential against the provider's chat-completion endpoint after refresh-on-expiry. Result lands on `CredentialHealthResult.completion` ({ok, reason?, modelId?, latencyMs?}) without disturbing the usage `ok` field. Public types: `CompletionProbe`, `CompletionProbeInput`, `CompletionProbeCredential`, `CredentialCompletionResult`. The probe is invoked even when no `UsageProvider` is registered for the row, and is skipped when OAuth refresh fails (the stale bytes would only mask the upstream failure).
 
 ### Changed
 
+- Changed auth-gateway credential resolution to use per-conversation `promptCacheKey`/`sessionId` when calling `AuthStorage.getApiKey`, so repeated turns can keep the same credential until it becomes unavailable
+- Changed auth-gateway and pi-native request handling to align `sessionId` with prompt/context identity before credential lookup
 - Changed Anthropic prompt preparation to downscale image blocks over 2000px when a request includes 20+ images, reducing oversized payloads automatically
 - Changed OpenAI chat request parsing to accept `name` on `tool` messages and fall back to the matching assistant `tool_calls` name, so parsed tool results now carry a proper tool name when the wire omits it
 - Changed `checkCredentials` to skip running `completionProbe` when OAuth refresh fails, so stale bearer tokens are never probed and the refresh failure remains the returned `reason`
@@ -15,6 +18,9 @@
 
 ### Fixed
 
+- Fixed auth-gateway to classify usage-limit messages such as `usage_limit_reached`, `resource_exhausted`, and Codex-style `Try again in ~X min` text as 429 `rate_limit_error` responses
+- Fixed auth-gateway usage-limit handling to honor parsed retry hints and switch to a sibling credential via `markUsageLimitReached` instead of invalidating the rate-limited credential
+- Fixed `streamSimple` to retry on usage-limit errors (including message-only error events) before any content is emitted, so `onAuthError` can rotate credentials automatically
 - Fixed auth-gateway error classification to extract embedded status codes and use word-boundary matching, so `GenerateContentRequest` and similar messages are no longer misreported as rate-limit errors
 - Fixed `checkCredentials` to handle `completionProbe` exceptions by recording the failure in `CredentialHealthResult.completion.reason` while still returning the usage probe result
 
diff --git a/packages/ai/src/auth-gateway/server.ts b/packages/ai/src/auth-gateway/server.ts
index 2c1859ef1..db2ed193d 100644
--- a/packages/ai/src/auth-gateway/server.ts
+++ b/packages/ai/src/auth-gateway/server.ts
@@ -17,13 +17,14 @@
  *   POST /v1/messages                      → Anthropic messages in/out
  *   POST /v1/responses                     → OpenAI Responses in/out
  */
-import { logger } from "@oh-my-pi/pi-utils";
+import { extractRetryHint, logger } from "@oh-my-pi/pi-utils";
 import type { AuthStorage } from "../auth-storage";
 import { Effort } from "../model-thinking";
 import * as anthropicMessages from "../providers/anthropic-messages-server";
 import * as openaiChat from "../providers/openai-chat-server";
 import * as openaiResponses from "../providers/openai-responses-server";
 import * as piNative from "../providers/pi-native-server";
+import { isUsageLimitError } from "../rate-limit-utils";
 import { streamSimple } from "../stream";
 import type { Api, AssistantMessageEventStream, Context, Model, SimpleStreamOptions } from "../types";
 import { parseBind } from "../utils/parse-bind";
@@ -231,7 +232,17 @@ export function classifyGatewayError(err: unknown): { status: number; type: stri
 	if (
 		// Match rate-limit phrasings without colliding with
 		// `GenerateContentRequest`, `accelerate`, `iterate`, `deprecated`, etc.
-		/\brate[- _]?limit(?:s|ed|ing)?\b|\bquota(?:_exceeded| exceeded)?\b|\btoo[- _]many[- _]requests\b/i.test(message)
+		/\brate[- _]?limit(?:s|ed|ing)?\b|\bquota(?:_exceeded| exceeded)?\b|\btoo[- _]many[- _]requests\b/i.test(
+			message,
+		) ||
+		// Usage-limit phrasings emit no embedded status. Codex friendly text
+		// reads "You have hit your ChatGPT usage limit … Try again in ~158
+		// min."; pi-ai's central `isUsageLimitError` already encodes every
+		// known provider variant, so reuse it instead of forking the regex.
+		// Without this branch the classifier falls through to the default
+		// 502/upstream_error, which is what callers were seeing when their
+		// account hit its cap.
+		isUsageLimitError(message)
 	) {
 		return { status: 429, type: "rate_limit_error", message };
 	}
@@ -266,9 +277,32 @@ function extractEmbeddedStatus(message: string): number | undefined {
 	return Number.isFinite(code) && code >= 100 && code < 600 ? code : undefined;
 }
 
+/**
+ * Hook fired by {@link streamSimple} when the upstream request fails in a
+ * way that's rotatable — today that's HTTP 401 (credential is bad) and
+ * usage-limit phrasing matched by {@link isUsageLimitError} (Codex's
+ * `usage_limit_reached`, Anthropic's `usage_limit_reached`, Google's
+ * `resource_exhausted`, …). The two cases need different storage actions:
+ *
+ * - **usage-limit** → {@link AuthStorage.markUsageLimitReached}. Marks just
+ *   the current session's credential as temporarily blocked (honouring
+ *   `retry-after` / `resets_at` hints when present) and returns `true` only
+ *   when a sibling credential is still available. Burning the credential
+ *   with `invalidateCredentialMatching` here would orphan accounts whose
+ *   reset window is several hours away — exactly the bug this helper exists
+ *   to avoid.
+ * - **auth-failure** → {@link AuthStorage.invalidateCredentialMatching}.
+ *   Suspect/delete the row so it doesn't get re-picked next request.
+ *
+ * In both branches we return the next `getApiKey` result (sticky on the
+ * same `sessionId`) so streamSimple can transparently retry the pre-emit
+ * failure with a fresh credential. Returning `undefined` aborts the retry
+ * and surfaces the original error to the caller.
+ */
 async function refreshGatewayApiKeyAfterAuthError(
 	storage: AuthStorage,
 	model: Model<Api>,
+	sessionId: string,
 	provider: string,
 	oldKey: string,
 	error: unknown,
@@ -276,14 +310,33 @@ async function refreshGatewayApiKeyAfterAuthError(
 	format: string,
 	peer: string,
 ): Promise<string | undefined> {
-	await storage.invalidateCredentialMatching(provider, oldKey, signal);
+	const message = error instanceof Error ? error.message : String(error);
+	if (isUsageLimitError(message)) {
+		const retryAfterMs = extractRetryHint(undefined, message);
+		const switched = await storage.markUsageLimitReached(provider, sessionId, {
+			retryAfterMs,
+			baseUrl: model.baseUrl,
+			signal,
+		});
+		logger.debug("auth-gateway retrying provider request after usage-limit block", {
+			format,
+			provider,
+			peer,
+			switched,
+			retryAfterMs,
+			error: message,
+		});
+		if (!switched) return undefined;
+		return storage.getApiKey(provider, sessionId, { modelId: model.id, signal });
+	}
+	await storage.invalidateCredentialMatching(provider, oldKey, { sessionId, signal });
 	logger.debug("auth-gateway retrying provider request after credential invalidation", {
 		format,
 		provider,
 		peer,
-		error: error instanceof Error ? error.message : String(error),
+		error: message,
 	});
-	return storage.getApiKey(provider, undefined, { modelId: model.id, signal });
+	return storage.getApiKey(provider, sessionId, { modelId: model.id, signal });
 }
 
 function clientClosedResponse(route: { module: FormatModule }): Response {
@@ -336,37 +389,12 @@ async function handleFormatEndpoint(
 		return route.module.formatError(404, "invalid_request_error", `Unknown model: ${modelId}`);
 	}
 
-	// pi-ai's stream() does NOT consult AuthStorage — the caller (us) is
-	// expected to resolve the credential and pass it as `options.apiKey`.
-	// For OAuth providers this returns the access token (refreshed via the
-	// broker override on AuthStorage when needed).
-	let apiKey: string | undefined;
-	try {
-		apiKey = await bootOpts.storage.getApiKey(model.provider, undefined, {
-			modelId: model.id,
-			signal: controller.signal,
-		});
-	} catch (error) {
-		if (controller.signal.aborted) return clientClosedResponse(route);
-		const classified = classifyGatewayError(error);
-		logger.warn("auth-gateway getApiKey threw", { provider: model.provider, peer, error: classified.message });
-		return route.module.formatError(classified.status, classified.type, classified.message);
-	}
-	if (controller.signal.aborted) return clientClosedResponse(route);
-	if (!apiKey) {
-		return route.module.formatError(
-			401,
-			"authentication_error",
-			`No credential available for provider ${model.provider}`,
-		);
-	}
-
-	// Parse + validate against the strict format schema, rebuild as omp's
-	// canonical Context, dispatch through pi-ai's streamSimple, encode the
-	// canonical event stream back to the inbound format. There is no
-	// passthrough fast-path — every request flows through pi-ai so that
-	// credential-specific request shaping (OAuth Claude-Code prefix, beta
-	// headers, codex websocket transport, …) always applies.
+	// Parse the wire-format request BEFORE resolving the credential so we
+	// have a stable per-conversation `sessionId` to thread into AuthStorage.
+	// Sticky-credential tracking and `markUsageLimitReached` both key off
+	// this id; without it `getApiKey` would re-roundrobin every request
+	// and `markUsageLimitReached` would no-op (it can only mark the
+	// credential it last handed out to that session).
 	let parsed: ParsedFormatRequest;
 	try {
 		parsed = route.module.parseRequest(body, req.headers);
@@ -385,12 +413,45 @@ async function handleFormatEndpoint(
 	}
 	if (controller.signal.aborted) return clientClosedResponse(route);
 
+	// Sticky credential id: honour the client's `prompt_cache_key` when
+	// supplied (so external session ids align), otherwise derive from
+	// modelId + system + tools + first message. Mirrored into
+	// streamOpts.sessionId / promptCacheKey by `buildStreamOptions`.
+	const sessionId = parsed.options.promptCacheKey ?? deriveSessionId(parsed.modelId, parsed.context);
+	parsed.options.promptCacheKey ??= sessionId;
+
+	// pi-ai's stream() does NOT consult AuthStorage — the caller (us) is
+	// expected to resolve the credential and pass it as `options.apiKey`.
+	// For OAuth providers this returns the access token (refreshed via the
+	// broker override on AuthStorage when needed).
+	let apiKey: string | undefined;
+	try {
+		apiKey = await bootOpts.storage.getApiKey(model.provider, sessionId, {
+			modelId: model.id,
+			signal: controller.signal,
+		});
+	} catch (error) {
+		if (controller.signal.aborted) return clientClosedResponse(route);
+		const classified = classifyGatewayError(error);
+		logger.warn("auth-gateway getApiKey threw", { provider: model.provider, peer, error: classified.message });
+		return route.module.formatError(classified.status, classified.type, classified.message);
+	}
+	if (controller.signal.aborted) return clientClosedResponse(route);
+	if (!apiKey) {
+		return route.module.formatError(
+			401,
+			"authentication_error",
+			`No credential available for provider ${model.provider}`,
+		);
+	}
+
 	const streamOpts = buildStreamOptions(parsed, model.api, controller.signal);
 	streamOpts.apiKey = apiKey;
 	streamOpts.onAuthError = (provider, oldKey, error) =>
 		refreshGatewayApiKeyAfterAuthError(
 			bootOpts.storage,
 			model,
+			sessionId,
 			provider,
 			oldKey,
 			error,
@@ -508,10 +569,17 @@ async function handlePiNative(bootOpts: AuthGatewayBootOptions, req: Request, pe
 	if (!model) {
 		return piNative.formatError(404, "invalid_request_error", `Unknown model: ${parsed.modelId}`);
 	}
+	// Pi-native already parsed `streamOpts.sessionId` (when set by the
+	// client); fall back to the derived key so credential-stickiness lines
+	// up with cache-prefix stickiness — same identity used for both means
+	// the next turn of this conversation reuses the same credential until
+	// it hits a usage cap, then markUsageLimitReached can hand off.
+	const sessionId = parsed.options.sessionId ?? deriveSessionId(parsed.modelId, parsed.context);
+	parsed.options.sessionId ??= sessionId;
 
 	let apiKey: string | undefined;
 	try {
-		apiKey = await bootOpts.storage.getApiKey(model.provider, undefined, {
+		apiKey = await bootOpts.storage.getApiKey(model.provider, sessionId, {
 			modelId: model.id,
 			signal: controller.signal,
 		});
@@ -539,6 +607,7 @@ async function handlePiNative(bootOpts: AuthGatewayBootOptions, req: Request, pe
 		refreshGatewayApiKeyAfterAuthError(
 			bootOpts.storage,
 			model,
+			sessionId,
 			provider,
 			oldKey,
 			error,
@@ -554,10 +623,7 @@ async function handlePiNative(bootOpts: AuthGatewayBootOptions, req: Request, pe
 	// headers — the client's values win when they collide.
 	const captured = captureRequestHeaders(req.headers);
 	streamOpts.headers = { ...captured, ...(streamOpts.headers ?? {}) };
-	// Cache identity: explicit `sessionId` wins, then derive a stable key
-	// from model + system + tools + first message so Codex prefix caching
-	// engages on the same logical conversation across turns.
-	streamOpts.sessionId ??= deriveSessionId(parsed.modelId, parsed.context);
+	streamOpts.sessionId ??= sessionId;
 
 	logger.info("auth-gateway request", {
 		format: "pi-native",
diff --git a/packages/ai/src/stream.ts b/packages/ai/src/stream.ts
index 7bbe24fd0..b61b16011 100644
--- a/packages/ai/src/stream.ts
+++ b/packages/ai/src/stream.ts
@@ -45,6 +45,7 @@ import {
 } from "./providers/register-builtins";
 import { isSyntheticModel, streamSynthetic } from "./providers/synthetic";
 import { streamXAIResponses } from "./providers/xai-responses";
+import { isUsageLimitError } from "./rate-limit-utils";
 import type {
 	Api,
 	AssistantMessage,
@@ -318,6 +319,18 @@ function extractStatusFromAssistantError(message: AssistantMessage): number | un
 	return extractHttpStatusFromError({ message: message.errorMessage });
 }
 
+function isRetryableUpstreamError(error: unknown, status: number | undefined, message: string | undefined): boolean {
+	// 401 means the credential is bad. Usage-limit phrasing (Codex's
+	// "You have hit your ChatGPT usage limit", Anthropic's "usage_limit_reached",
+	// Google's "resource_exhausted") means this account is parked but a
+	// sibling credential can usually pick the request up. Both are
+	// rotatable via `onAuthError` — the auth-gateway maps the former to
+	// `invalidateCredentialMatching` and the latter to `markUsageLimitReached`.
+	if (status === 401) return true;
+	void error;
+	return !!message && isUsageLimitError(message);
+}
+
 function createAssistantAuthError(message: AssistantMessage): Error & { status?: number } {
 	const error: Error & { status?: number } = new Error(message.errorMessage ?? "Provider authentication failed");
 	const status = extractStatusFromAssistantError(message);
@@ -359,7 +372,11 @@ export function streamSimple<TApi extends Api>(
 						!emittedReplayUnsafeEvent &&
 						captureAuthFailure &&
 						event.type === "error" &&
-						extractStatusFromAssistantError(event.error) === 401
+						isRetryableUpstreamError(
+							event.error,
+							extractStatusFromAssistantError(event.error),
+							event.error.errorMessage,
+						)
 					) {
 						return { error: createAssistantAuthError(event.error), bufferedEvents, terminalEvent: event };
 					}
@@ -371,7 +388,15 @@ export function streamSimple<TApi extends Api>(
 				flushBuffered();
 				if (!outer.done) outer.end(await inner.result());
 			} catch (error) {
-				if (!emittedReplayUnsafeEvent && captureAuthFailure && extractHttpStatusFromError(error) === 401) {
+				if (
+					!emittedReplayUnsafeEvent &&
+					captureAuthFailure &&
+					isRetryableUpstreamError(
+						error,
+						extractHttpStatusFromError(error),
+						error instanceof Error ? error.message : undefined,
+					)
+				) {
 					return { error, bufferedEvents };
 				}
 				flushBuffered();
diff --git a/packages/ai/test/auth-gateway-classify-error.test.ts b/packages/ai/test/auth-gateway-classify-error.test.ts
index c3911f8c2..d780b266f 100644
--- a/packages/ai/test/auth-gateway-classify-error.test.ts
+++ b/packages/ai/test/auth-gateway-classify-error.test.ts
@@ -54,6 +54,27 @@ describe("auth-gateway classifyGatewayError", () => {
 		expect(c.type).toBe("rate_limit_error");
 	});
 
+	it("classifies Codex 'You have hit your ChatGPT usage limit' as 429", () => {
+		// Verbatim shape Codex returns from the `usage_limit_reached` branch
+		// in `parseCodexError`. No embedded `HTTP NNN`/`(NNN)`/`status NNN`
+		// token, no `rate limit`/`too many requests` wording — only the
+		// gateway's `isUsageLimitError` branch catches this. Previously it
+		// fell through to the default 502/upstream_error, which is why the
+		// `lg` retry loop kept looping instead of switching to another
+		// credential.
+		const c = classifyGatewayError(
+			new Error("You have hit your ChatGPT usage limit (pro plan). Try again in ~158 min."),
+		);
+		expect(c.status).toBe(429);
+		expect(c.type).toBe("rate_limit_error");
+	});
+
+	it("classifies generic 'usage_limit_reached' code text as 429", () => {
+		const c = classifyGatewayError(new Error('{"code":"usage_limit_reached","message":"…"}'));
+		expect(c.status).toBe(429);
+		expect(c.type).toBe("rate_limit_error");
+	});
+
 	it("does not match 'rate' inside camelCase or compound words", () => {
 		// `Generate`, `iterate`, `deprecated`, `accelerate` all contain `rate` as
 		// a substring and used to trip the classifier.
diff --git a/packages/ai/test/google-gemini-cli-429.test.ts b/packages/ai/test/google-gemini-cli-429.test.ts
index 934e6218d..6008f1fe1 100644
--- a/packages/ai/test/google-gemini-cli-429.test.ts
+++ b/packages/ai/test/google-gemini-cli-429.test.ts
@@ -80,6 +80,27 @@ describe("extractRetryHint – body text parsing", () => {
 		expect(extractRetryHint(undefined, "try again in 12s")).toBe(12_000);
 	});
 
+	it("parses Codex 'Try again in ~X min.' (usage_limit_reached friendly text)", () => {
+		// Verbatim shape Codex's parseCodexError builds when usage_limit_reached
+		// arrives with a `resets_at` minutes-out reset time. Used to fall
+		// through to undefined → the gateway and TUI both had no retry-after
+		// signal to honour, so they defaulted to QUOTA_EXHAUSTED's 30-min
+		// blanket and rotated immediately even when the actual reset window
+		// was much longer.
+		expect(extractRetryHint(undefined, "Try again in ~158 min.")).toBe(158 * 60_000);
+	});
+
+	it("parses 'try again in X min' / 'X minutes' without the tilde", () => {
+		expect(extractRetryHint(undefined, "try again in 5 min")).toBe(5 * 60_000);
+		expect(extractRetryHint(undefined, "try again in 90 minutes")).toBe(90 * 60_000);
+	});
+
+	it("parses 'try again in X h' / 'X hour' / 'X hours'", () => {
+		expect(extractRetryHint(undefined, "try again in 2 h")).toBe(2 * 60 * 60_000);
+		expect(extractRetryHint(undefined, "try again in 1 hour")).toBe(60 * 60_000);
+		expect(extractRetryHint(undefined, "try again in 3 hours")).toBe(3 * 60 * 60_000);
+	});
+
 	it("returns undefined when body contains no recognised delay pattern", () => {
 		expect(extractRetryHint(undefined, "Quota exceeded, please try again later")).toBeUndefined();
 	});
diff --git a/packages/ai/test/stream-auth-retry.test.ts b/packages/ai/test/stream-auth-retry.test.ts
index 6358fb274..d019e57ce 100644
--- a/packages/ai/test/stream-auth-retry.test.ts
+++ b/packages/ai/test/stream-auth-retry.test.ts
@@ -244,4 +244,131 @@ describe("streamSimple auth retry", () => {
 		expect(keys).toEqual(["old-key", "new-key"]);
 		expect(authCalls).toBe(1);
 	});
+
+	it("retries on a thrown usage_limit_reached error before any event has been emitted", async () => {
+		const keys: Array<string | undefined> = [];
+		const errors: unknown[] = [];
+		registerCustomApi(
+			API,
+			(_model: Model<Api>, _context: Context, options?: SimpleStreamOptions) => {
+				keys.push(options?.apiKey);
+				const stream = new AssistantMessageEventStream();
+				queueMicrotask(() => {
+					if (keys.length === 1) {
+						stream.fail(
+							Object.assign(
+								new Error("You have hit your ChatGPT usage limit (pro plan). Try again in ~158 min."),
+								{ status: 429 },
+							),
+						);
+						return;
+					}
+					const message = assistant(["ok"]);
+					stream.push({ type: "start", partial: message });
+					stream.push({ type: "done", reason: "stop", message });
+				});
+				return stream;
+			},
+			SOURCE_ID,
+		);
+
+		const stream = streamSimple(model(), context, {
+			apiKey: "old-key",
+			onAuthError: async (_provider, _oldKey, error) => {
+				errors.push(error);
+				return "new-key";
+			},
+		});
+
+		for await (const _event of stream) {
+			// drain
+		}
+
+		expect((await stream.result()).content).toEqual([{ type: "text", text: "ok" }]);
+		expect(keys).toEqual(["old-key", "new-key"]);
+		expect(errors).toHaveLength(1);
+		// The error surfaced to onAuthError carries the original 429 status so
+		// the gateway's refresh hook can branch on it.
+		expect((errors[0] as { status?: number }).status).toBe(429);
+		expect((errors[0] as Error).message).toMatch(/usage limit/i);
+	});
+
+	it("retries when a provider emits a usage_limit_reached error event before content", async () => {
+		const keys: Array<string | undefined> = [];
+		registerCustomApi(
+			API,
+			(_model: Model<Api>, _context: Context, options?: SimpleStreamOptions) => {
+				keys.push(options?.apiKey);
+				const stream = new AssistantMessageEventStream();
+				queueMicrotask(() => {
+					if (keys.length === 1) {
+						stream.push({ type: "start", partial: assistant() });
+						stream.push({
+							type: "error",
+							reason: "error",
+							// errorStatus deliberately omitted: matches how the codex
+							// provider serializes the message-only path that used to
+							// 502 in the gateway.
+							error: assistantError("You have hit your ChatGPT usage limit (pro plan). Try again in ~158 min."),
+						});
+						return;
+					}
+					const message = assistant(["ok"]);
+					stream.push({ type: "start", partial: message });
+					stream.push({ type: "done", reason: "stop", message });
+				});
+				return stream;
+			},
+			SOURCE_ID,
+		);
+
+		const stream = streamSimple(model(), context, {
+			apiKey: "old-key",
+			onAuthError: async () => "new-key",
+		});
+
+		for await (const _event of stream) {
+			// drain
+		}
+
+		expect((await stream.result()).content).toEqual([{ type: "text", text: "ok" }]);
+		expect(keys).toEqual(["old-key", "new-key"]);
+	});
+
+	it("surfaces the original usage_limit error when the retry callback declines", async () => {
+		const keys: Array<string | undefined> = [];
+		const original = Object.assign(new Error("You have hit your ChatGPT usage limit (pro plan)."), { status: 429 });
+		registerCustomApi(
+			API,
+			(_model: Model<Api>, _context: Context, options?: SimpleStreamOptions) => {
+				keys.push(options?.apiKey);
+				const stream = new AssistantMessageEventStream();
+				queueMicrotask(() => stream.fail(original));
+				return stream;
+			},
+			SOURCE_ID,
+		);
+
+		// Callback returns undefined → no sibling credential to rotate to.
+		// The original failure must reach the caller untouched so the client
+		// can decide what to do (back off, surface to user, …).
+		const stream = streamSimple(model(), context, {
+			apiKey: "old-key",
+			onAuthError: async () => undefined,
+		});
+
+		let caught: unknown;
+		try {
+			for await (const _event of stream) {
+				// drain
+			}
+		} catch (error) {
+			caught = error;
+		}
+
+		expect(caught).toBe(original);
+		// Single attempt — the inner stream is only re-invoked when a new key
+		// is provided.
+		expect(keys).toEqual(["old-key"]);
+	});
 });
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index 2deb69659..25bcb3589 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -16,6 +16,22 @@
 - Modified hashline anchor syntax to require explicit range notation `A-B:` instead of shorthand `A:` for single-line operations
 - Updated hashline description in settings to clarify pure insert context behavior without arrow notation
 
+### Removed
+
+- Removed the `edit.hashlineAutoDropPureInsertDuplicates` setting
+- Removed the `edit.hashlineAutoDropPureInsertDuplicates` setting from configuration and execution paths
+
+### Fixed
+
+- Fixed `eval` tool to resize large displayed images and append dimension notes to text output
+- Fixed `write` tool to strip malformed or loose hashline section headers before writing file content
+- Fixed `eval` tool image rendering to resize displayed images before returning them and append image-dimension notes to text output
+- Fixed `write` tool output sanitation to strip malformed or loose hashline section headers before writing file content
+- Fixed `omp auth-broker serve` crashing at startup with `logger.setTransports is not a function` — switched the call site to `import { setTransports } from "@oh-my-pi/pi-utils/logger"`, bypassing the `logger` namespace re-export that some Bun versions failed to expose at runtime
+- Fixed `omp auth-gateway` returning `502 upstream_error` and refusing to rotate credentials when a provider responded with a non-401 usage-limit error (Codex `usage_limit_reached`, Anthropic `usage_limit_reached`, Google `resource_exhausted`). `classifyGatewayError` now reuses `pi-ai`'s central `isUsageLimitError` heuristic and reports those failures as `429 rate_limit_error`. `streamSimple`'s pre-emit retry hook fires on usage-limit phrasing in addition to HTTP 401; the gateway's refresh callback branches on the error type and calls `AuthStorage.markUsageLimitReached(provider, sessionId, { retryAfterMs })` — temporarily blocking just the exhausted credential and surfacing the next sibling — instead of `invalidateCredentialMatching`, which would have suspect/deleted the row. The same branching is wired into the coding-agent `streamFn` callback so subscription multi-account rotation works the same on both surfaces.
+- Fixed `extractRetryHint` not recognising Codex's `Try again in ~N min.` / `… hour` / `… hours` phrasing, which left the gateway and TUI without a server-suggested retry window when an upstream account hit its usage cap. The shared `try again in` pattern now accepts `min`, `minutes`, `mins`, `h`, `hr`, `hour`, `hours` units in addition to `ms` / `s` / `sec`, and tolerates a leading `~` and embedded whitespace.
+- Fixed the auth-gateway threading `sessionId: undefined` into `AuthStorage.getApiKey`, which left `#sessionLastCredential` empty and made `markUsageLimitReached` a no-op for gateway-mediated requests. Both `/v1/chat/completions`-style endpoints and the `/v1/pi/stream` fast path now derive a stable `sessionId` from the client's `prompt_cache_key` (or the existing model+system+tools+first-message hash when absent) and reuse the same identity for credential-stickiness and prefix-cache routing.
+
 ## [15.5.7] - 2026-05-27
 ### Added
 - `providers.openrouterVariant` setting (Settings → Providers → "OpenRouter Routing") to default OpenRouter requests to a routing-variant suffix (`:nitro`, `:floor`, `:online`, `:exacto`). Selectors that already name a variant (e.g. `openrouter/anthropic/claude-haiku:nitro`) keep precedence.
diff --git a/packages/coding-agent/src/sdk.ts b/packages/coding-agent/src/sdk.ts
index 2e8bad681..24c1ea4e4 100644
--- a/packages/coding-agent/src/sdk.ts
+++ b/packages/coding-agent/src/sdk.ts
@@ -10,6 +10,7 @@ import {
 } from "@oh-my-pi/pi-agent-core";
 import {
 	type CredentialDisabledEvent,
+	isUsageLimitError,
 	type Message,
 	type Model,
 	type SimpleStreamOptions,
@@ -23,6 +24,7 @@ import type { Component } from "@oh-my-pi/pi-tui";
 import {
 	$env,
 	$flag,
+	extractRetryHint,
 	getAgentDbPath,
 	getAgentDir,
 	getProjectDir,
@@ -1885,13 +1887,36 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 					...streamOptions,
 					openrouterVariant: streamOptions?.openrouterVariant ?? openrouterVariant,
 					onAuthError: async (provider, oldKey, error) => {
+						const message = error instanceof Error ? error.message : String(error);
+						// streamSimple invokes this for both 401 auth failures AND
+						// rotatable usage-limit errors (Codex usage_limit_reached,
+						// Anthropic usage_limit_reached, etc.). The two need
+						// different storage actions: a real 401 means the credential
+						// is bad and should be marked suspect; a usage limit just
+						// means this account is parked until reset and should be
+						// temporarily blocked so a sibling can pick the request up.
+						if (isUsageLimitError(message)) {
+							const retryAfterMs = extractRetryHint(undefined, message);
+							const switched = await modelRegistry.authStorage.markUsageLimitReached(provider, agent.sessionId, {
+								retryAfterMs,
+								signal: streamOptions?.signal,
+							});
+							logger.debug("Retrying provider request after usage-limit block", {
+								provider,
+								switched,
+								retryAfterMs,
+								error: message,
+							});
+							if (!switched) return undefined;
+							return modelRegistry.getApiKeyForProvider(provider, agent.sessionId);
+						}
 						await modelRegistry.authStorage.invalidateCredentialMatching(provider, oldKey, {
 							signal: streamOptions?.signal,
 							sessionId: agent.sessionId,
 						});
 						logger.debug("Retrying provider request after credential invalidation", {
 							provider,
-							error: error instanceof Error ? error.message : String(error),
+							error: message,
 						});
 						return modelRegistry.getApiKeyForProvider(provider, agent.sessionId);
 					},
diff --git a/packages/utils/src/fetch-retry.ts b/packages/utils/src/fetch-retry.ts
index b521250fe..06961f0bc 100644
--- a/packages/utils/src/fetch-retry.ts
+++ b/packages/utils/src/fetch-retry.ts
@@ -6,8 +6,10 @@ const QUOTA_RESET_PATTERN = /reset after (?:(\d+)h)?(?:(\d+)m)?(\d+(?:\.\d+)?)s/
 const PLEASE_RETRY_PATTERN = /Please retry in ([0-9.]+)(ms|s)/i;
 // JSON field: "retryDelay": "34.074824224s"
 const RETRY_DELAY_FIELD_PATTERN = /"retryDelay":\s*"([0-9.]+)(ms|s)"/i;
-// "try again in 250ms" / "try again in 12s" / "try again in 12sec"
-const TRY_AGAIN_PATTERN = /try again in\s+(\d+(?:\.\d+)?)\s*(ms|s)(?:ec)?/i;
+// "try again in 250ms" / "try again in 12s" / "try again in 12sec" /
+// "try again in 5 min" / "try again in ~158 min." / "try again in 2h" /
+// "try again in 90 minutes" / "try again in 1 hour"
+const TRY_AGAIN_PATTERN = /try again in\s+~?\s*([0-9.]+)\s*(ms|sec|s|minutes?|mins?|m|hours?|hrs?|h)\b/i;
 
 /**
  * Server-suggested retry delay extraction. Merges the patterns historically used
@@ -22,7 +24,7 @@ const TRY_AGAIN_PATTERN = /try again in\s+(\d+(?:\.\d+)?)\s*(ms|s)(?:ec)?/i;
  *  - `Your quota will reset after 18h31m10s` / `10m15s` / `39s`
  *  - `Please retry in 250ms` / `Please retry in 12s`
  *  - `"retryDelay": "34.074824224s"` (JSON error detail field)
- *  - `try again in 250ms` / `try again in 12s` / `try again in 12sec`
+ *  - `try again in 250ms` / `try again in 12s` / `try again in 5 min` / `try again in ~158 min`
  *
  * Returns `undefined` if no signal is found.
  */
@@ -68,13 +70,38 @@ export function extractRetryHint(source: Response | Headers | null | undefined,
 		if (match?.[1]) {
 			const value = Number.parseFloat(match[1]);
 			if (Number.isFinite(value) && value > 0) {
-				return match[2]!.toLowerCase() === "ms" ? value : value * 1000;
+				const unitMs = unitToMs(match[2]!);
+				if (unitMs !== undefined) return value * unitMs;
 			}
 		}
 	}
 	return undefined;
 }
 
+function unitToMs(unit: string): number | undefined {
+	switch (unit.toLowerCase()) {
+		case "ms":
+			return 1;
+		case "s":
+		case "sec":
+			return 1000;
+		case "m":
+		case "min":
+		case "mins":
+		case "minute":
+		case "minutes":
+			return 60_000;
+		case "h":
+		case "hr":
+		case "hrs":
+		case "hour":
+		case "hours":
+			return 60 * 60_000;
+		default:
+			return undefined;
+	}
+}
+
 export interface FetchWithRetryOptions extends RequestInit {
 	/** Total fetch attempts (initial + retries). Default `5`. */
 	maxAttempts?: number;