Merge pull request #710 from corbitsdev/cl-7173-treat-codex-short-429s-as-retryable-rate-limits

Remap Codex short 429s to retryable rate limits

Author: TheGreatAxios

src/agent/retry-policy.test.ts

@@ -69,7 +69,23 @@ describe("createCorbitsRetryPolicy", () => {
         raw: { error: { message: "Too Many Requests" } },
       },
     });
-    // Remapped to retryable → default backoff, not abort on moderate Retry-After.
+    // Remapped to retryable -> default backoff, not abort on moderate Retry-After.
+    expect(decision).toEqual({ kind: "retry", delayMs: 500 });
+  });
+
+  test("stamped Codex usage-limit 429 retries as retryable, not long-quota abort", async () => {
+    const policy = createCorbitsRetryPolicy({ providerId: "codex/abk-labs" });
+    const decision = await policy({
+      attempt: 1,
+      elapsedMs: 0,
+      error: {
+        category: "quota_exhausted",
+        message: "You have hit your ChatGPT usage limit",
+        statusCode: 429,
+        retryAfterMs: 45_000,
+        raw: "You have hit your ChatGPT usage limit",
+      },
+    });
     expect(decision).toEqual({ kind: "retry", delayMs: 500 });
   });
 

src/inference-error-message.test.ts

@@ -65,4 +65,27 @@ describe("inferenceErrorMessage", () => {
     });
     expect(line).toBe("Quota exhausted — usage limit reached.");
   });
+
+  test("known-Codex short 429 shows rate-limit line, not usage-limit copy", () => {
+    const line = inferenceErrorMessage({
+      category: "quota_exhausted",
+      message: "You have hit your ChatGPT usage limit",
+      statusCode: 429,
+      providerId: "codex/abk-labs",
+      raw: "You have hit your ChatGPT usage limit",
+    });
+    expect(line.toLowerCase()).toMatch(/rate limit/);
+    expect(line).not.toContain("Quota exhausted");
+    expect(line.toLowerCase()).not.toContain("the usage limit has been reached");
+    expect(line.toLowerCase()).not.toContain("usage limit reached");
+  });
+
+  test("credential_failure tells the user to log in again", () => {
+    const line = inferenceErrorMessage({
+      category: "credential_failure",
+      message: '{"error":{"code":401}}',
+    });
+    expect(line.toLowerCase()).not.toContain("re-authenticating");
+    expect(line.toLowerCase()).toMatch(/log in again|sign in again/);
+  });
 });

src/inference-error-message.ts

@@ -13,16 +13,16 @@ import {
 import { codexProfileFromProviderName, isCodexProviderName } from "./config/codex-providers.js";
 import {
   gatewayOverloadUserMessage,
+  isCodexShortRateLimitInferenceError,
   isGatewayOverloadInferenceError,
   isXaiShortRateLimitInferenceError,
-  XAI_RATE_LIMIT_USER_MESSAGE,
+  RATE_LIMIT_USER_MESSAGE,
   type InferenceErrorLike,
 } from "./inference-gateway-error.js";
 
 const FRIENDLY_BY_CATEGORY: Record<string, string> = {
-  // Re-authentication runs on its own; keep the transcript line short and free
-  // of the provider's raw 401 JSON.
-  credential_failure: "Session expired — re-authenticating…",
+  // Committed auth death — do not claim a refresh is in flight.
+  credential_failure: "Authentication failed — log in again.",
   quota_exhausted: "Quota exhausted — usage limit reached.",
   context_overflow:
     "Context window full — compaction could not keep up. Try /clear to start fresh.",
@@ -93,8 +93,10 @@ function codexUsageLimitLine(error: InferenceErrorLike): string | undefined {
 export function inferenceErrorMessage(error: InferenceErrorLike): string {
   if (isGatewayOverloadInferenceError(error)) return gatewayOverloadUserMessage(error);
   // Dual-path: harness may still emit intx's quota_exhausted for a known-xAI
-  // short 429; FRIENDLY_BY_CATEGORY would otherwise say "Quota exhausted".
-  if (isXaiShortRateLimitInferenceError(error)) return XAI_RATE_LIMIT_USER_MESSAGE;
+  // or known-Codex short 429; FRIENDLY_BY_CATEGORY would otherwise say
+  // "Quota exhausted".
+  if (isXaiShortRateLimitInferenceError(error) || isCodexShortRateLimitInferenceError(error))
+    return RATE_LIMIT_USER_MESSAGE;
 
   const category = classifyInferenceErrorCategory(error);
   if (category === "quota_exhausted") {

src/inference-gateway-error.test.ts

@@ -318,4 +318,70 @@ describe("normalizeInferenceErrorForRetry", () => {
     };
     expect(normalizeInferenceErrorForRetry(err)).toEqual(err);
   });
+
+  test("known-Codex bare 429 without usage_limit_reached remaps to retryable", () => {
+    const bare = {
+      category: "quota_exhausted" as const,
+      message: "Too Many Requests",
+      statusCode: 429,
+      retryAfterMs: 5_000,
+      raw: { error: { message: "Too Many Requests" } },
+    };
+
+    expect(normalizeInferenceErrorForRetry(bare)).toEqual(bare);
+
+    const viaProviderId = normalizeInferenceErrorForRetry({
+      ...bare,
+      providerId: "codex/abk-labs",
+    });
+    expect(viaProviderId.category).toBe("retryable");
+    expect(viaProviderId.retryAfterMs).toBe(5_000);
+    expect(viaProviderId.message.toLowerCase()).toMatch(/rate limit/);
+    expect(viaProviderId.message.toLowerCase()).not.toMatch(/quota exhausted|usage limit reached/);
+  });
+
+  test("known-Codex 429 with ChatGPT usage-limit prose remaps to retryable", () => {
+    const normalized = normalizeInferenceErrorForRetry({
+      category: "quota_exhausted",
+      message: "You have hit your ChatGPT usage limit",
+      statusCode: 429,
+      providerId: "codex/abk-labs",
+      raw: "You have hit your ChatGPT usage limit",
+    });
+    expect(normalized.category).toBe("retryable");
+    expect(normalized.message.toLowerCase()).toMatch(/rate limit/);
+    expect(normalized.message.toLowerCase()).not.toMatch(/quota exhausted|usage limit reached/);
+  });
+
+  test("known-Codex 429 with empty body remaps to retryable", () => {
+    const normalized = normalizeInferenceErrorForRetry({
+      category: "quota_exhausted",
+      message: "Too Many Requests",
+      statusCode: 429,
+      providerId: "codex/abk-labs",
+    });
+    expect(normalized.category).toBe("retryable");
+  });
+
+  test("known-Codex usage_limit_reached 429 stays quota_exhausted", () => {
+    const normalized = normalizeInferenceErrorForRetry({
+      category: "quota_exhausted",
+      message: "Too Many Requests",
+      statusCode: 429,
+      providerId: "codex/abk-labs",
+      raw: {
+        detail: {
+          error: {
+            code: "usage_limit_reached",
+            message: "You have reached your usage limit. Try again later.",
+            plan_type: "workspace_member",
+            resets_in_seconds: 3435,
+          },
+        },
+      },
+    });
+    expect(normalized.category).toBe("quota_exhausted");
+    expect(normalized.retryAfterMs).toBe(3_435_000);
+    expect(normalized.message).toContain('Codex profile "abk-labs"');
+  });
 });

src/inference-gateway-error.ts

@@ -49,8 +49,8 @@ const GATEWAY_OVERLOAD_TEXT_MARKERS = [
 /** User-visible line while the harness retries a transient gateway overload. */
 export const GATEWAY_OVERLOAD_USER_MESSAGE = "Inference gateway overloaded — retrying…";
 
-/** User-visible line while the harness retries a short known-xAI HTTP 429. */
-export const XAI_RATE_LIMIT_USER_MESSAGE = "Rate limited — retrying…";
+/** User-visible line while the harness retries a short known-provider HTTP 429. */
+export const RATE_LIMIT_USER_MESSAGE = "Rate limited — retrying…";
 
 /** Body markers that mean a real usage/quota window, not a short rate limit. */
 const XAI_QUOTA_BODY_MARKERS = [
@@ -246,7 +246,67 @@ export function normalizeXaiRateLimitError(error: InferenceErrorWithGoContext):
 
   return {
     category: "retryable",
-    message: XAI_RATE_LIMIT_USER_MESSAGE,
+    message: RATE_LIMIT_USER_MESSAGE,
+    statusCode: 429,
+    ...(error.raw !== undefined ? { raw: error.raw } : {}),
+    ...(error.retryAfterMs !== undefined ? { retryAfterMs: error.retryAfterMs } : {}),
+  };
+}
+
+function parseCodexUsageLimitFromError(
+  error: InferenceErrorLike,
+): ReturnType<typeof parseCodexUsageLimitError> {
+  const candidates: unknown[] = [];
+  if (error.raw !== undefined) candidates.push(error.raw);
+  if (typeof error.message === "string" && error.message.trim().startsWith("{")) {
+    candidates.push(error.message);
+  }
+
+  for (const candidate of candidates) {
+    const parsed = parseCodexUsageLimitError(candidate);
+    if (parsed !== undefined) return parsed;
+  }
+  return undefined;
+}
+
+function isKnownCodexProviderId(providerId: string | undefined): boolean {
+  return providerId !== undefined && isCodexProviderName(providerId);
+}
+
+/**
+ * True when a known-Codex HTTP 429 looks like a short rate limit rather than a
+ * `usage_limit_reached` window. Used by both retry normalization and transcript
+ * copy — FRIENDLY_BY_CATEGORY would otherwise paint every quota_exhausted 429 as
+ * "Quota exhausted" even when the policy remaps it to retryable.
+ *
+ * Discrimination is the existing Codex usage-limit parser, not Retry-After length
+ * and not ChatGPT usage-limit prose without `usage_limit_reached`.
+ */
+export function isCodexShortRateLimitInferenceError(error: InferenceErrorLike): boolean {
+  if (!isKnownCodexProviderId(error.providerId)) return false;
+  if (error.statusCode !== 429) return false;
+  if (error.category !== "quota_exhausted" && error.category !== "retryable") return false;
+  if (parseCodexUsageLimitFromError(error) !== undefined) return false;
+  return true;
+}
+
+/**
+ * intx defaults bare 429 → quota_exhausted. For known-Codex contexts a bare 429
+ * (or usage-limit prose without `usage_limit_reached`) reclassifies as retryable
+ * so short ChatGPT 429s are not painted as a committed usage-limit window.
+ *
+ * Nested `detail.error.code === usage_limit_reached` stays quota_exhausted via
+ * `normalizeCodexUsageLimitError`. Unknown / non-Codex providers are never remapped.
+ */
+export function normalizeCodexRateLimitError(error: InferenceErrorWithGoContext): InferenceError {
+  if (error.statusCode !== 429) return error;
+  if (error.category !== "quota_exhausted") return error;
+  if (!isKnownCodexProviderId(error.providerId)) return error;
+  if (parseCodexUsageLimitFromError(error) !== undefined) return error;
+
+  return {
+    category: "retryable",
+    message: RATE_LIMIT_USER_MESSAGE,
     statusCode: 429,
     ...(error.raw !== undefined ? { raw: error.raw } : {}),
     ...(error.retryAfterMs !== undefined ? { retryAfterMs: error.retryAfterMs } : {}),
@@ -266,17 +326,7 @@ function normalizeCodexUsageLimitError(error: InferenceErrorWithGoContext): Infe
     return error;
   }
 
-  const candidates: unknown[] = [];
-  if (error.raw !== undefined) candidates.push(error.raw);
-  if (typeof error.message === "string" && error.message.trim().startsWith("{")) {
-    candidates.push(error.message);
-  }
-
-  let parsed = undefined as ReturnType<typeof parseCodexUsageLimitError>;
-  for (const candidate of candidates) {
-    parsed = parseCodexUsageLimitError(candidate);
-    if (parsed !== undefined) break;
-  }
+  const parsed = parseCodexUsageLimitFromError(error);
   if (parsed === undefined) return error;
 
   const profile =
@@ -298,7 +348,8 @@ function normalizeCodexUsageLimitError(error: InferenceErrorWithGoContext): Infe
  * Reclassify gateway overload errors so the default retry policy treats them as
  * transient instead of aborting on protocol_mismatch. Also normalizes OpenCode
  * Go quota/rate-limit shapes (including HTTP 400 mis-status), known-xAI short
- * 429s, and Codex usage limits (nested detail.error with resets_in_seconds).
+ * 429s, Codex usage limits (nested detail.error with resets_in_seconds), and
+ * known-Codex short 429s that are not usage_limit_reached.
  */
 export function normalizeInferenceErrorForRetry(
   error: InferenceErrorWithGoContext,
@@ -312,6 +363,9 @@ export function normalizeInferenceErrorForRetry(
   const codexNormalized = normalizeCodexUsageLimitError(error);
   if (codexNormalized !== error) return codexNormalized;
 
+  const codexRateLimit = normalizeCodexRateLimitError(error);
+  if (codexRateLimit !== error) return codexRateLimit;
+
   if (!isGatewayOverloadInferenceError(error)) return error;
   if (error.category === "retryable" || error.category === "timeout") return error;
 

src/tui/stream-event-map.test.ts

@@ -314,7 +314,7 @@ describe("inference.error text", () => {
 
   test("a classified failure gets its written line, not the provider body", () => {
     expect(message({ category: "credential_failure", message: '{"error":{"code":401}}' })).toBe(
-      "Session expired — re-authenticating…",
+      "Authentication failed — log in again.",
     );
     expect(message({ category: "quota_exhausted", message: "429" })).toBe(
       "Quota exhausted — usage limit reached.",
@@ -371,6 +371,29 @@ describe("inference.error text", () => {
     expect(event.message).not.toContain("Quota exhausted");
   });
 
+  test("ctx.providerId Codex + ChatGPT usage-limit 429 shows rate-limit copy", () => {
+    const ctx = createStreamMapContext({ providerId: "codex/abk-labs" });
+    const [event] = mapProductionEvent(
+      {
+        type: "inference.error",
+        data: {
+          error: {
+            category: "quota_exhausted",
+            message: "You have hit your ChatGPT usage limit",
+            statusCode: 429,
+            raw: "You have hit your ChatGPT usage limit",
+          },
+        },
+      },
+      ctx,
+    );
+    expect(event?.type).toBe("error");
+    if (event?.type !== "error") return;
+    expect(event.message.toLowerCase()).toMatch(/rate limit/);
+    expect(event.message).not.toContain("Quota exhausted");
+    expect(event.message.toLowerCase()).not.toContain("usage limit reached");
+  });
+
   test("bare quota_exhausted 429 without ctx/provider still shows Quota exhausted", () => {
     expect(
       message({