Honor Retry-After header in rate-limit retry policy and clarify terminal message

[TheGreatAxios]

Summary

• Retry policy respects retryAfterMs from provider adapters, pacing retries at min(retryAfterMs, MAX_BLIND_WAIT_MS) instead of fixed 500/1000ms backoff
• Codex, Grok, and OpenAI Responses adapters implement extractRetryAfterMs to extract Retry-After / Retry-After-Ms headers from 429 responses
• Terminal message for rate-limit errors changed from misleading "Rate limited — retrying…" to clear guidance: "Rate limited. Wait a moment and try again."

Verification

• bun run check passes (lint, typecheck, build, projects-dir-guard)
• Tests cover paced retry behavior, header extraction, and terminal message copy

Fixes CL-7523

[linear-code]

CL-7523

[TheGreatAxios]

Greybeard · Comment

Responses adapters extract Retry-After; Corbits policy paces remapped 429s instead of the 500/1000ms backoff.

Findings

• src/agent/retry-policy.ts:49 — MAX_BLIND_WAIT_MS owns abort-vs-sleep for quota windows, not a delay cap. min(retryAfterMs, 30s) then retry waits 30s into the existing 45s xAI/Codex fixtures, retries early, and burns the attempt budget. Honor retryAfterMs on remapped retryable 429s; abort if the value is a hang.
• src/agent/retry-policy.ts — hardcoded attempt >= 3 re-owns vendored MAX_ATTEMPTS. Ask defaultPolicy for abort-vs-retry; override only delayMs.

Notes

• extractRetryAfterMs on Codex/Grok/OpenAI Responses is the right adapter boundary.
• Terminal copy is the right surface for the "retrying" lie; that string only appears after abort.

[TheGreatAxios]

Critic · Comment

Retry policy honors Retry-After on retryable 429s (capped at 30s), Responses adapters extract the header, and terminal copy no longer claims an in-flight retry after abort.

Findings

• src/agent/retry-policy.ts:59 — new path hardcodes attempt >= 3 instead of reusing vendored MAX_ATTEMPTS via defaultPolicy. Sequence: retryable 429 with retryAfterMs set, attempt === 3 aborts here; if vendor bumps the cap, this branch silently stays at 3. Override delayMs on a defaultPolicy retry (or share a constant) rather than re-capping.
• src/provider/grok-responses-adapter.ts:255 / src/provider/openai-responses-adapter.ts:237 — extractRetryAfterMs is wired but only Codex asserts it (codex-responses-adapter.test.ts:67). Dropping the one-liner on Grok or OpenAI would not fail the suite.
• src/provider/codex-responses-adapter.ts:643 — extractResponsesRetryAfterMs matches vendored Chat Completions (vendor/intx-inference/src/providers/openai.ts:1011): numeric retry-after / retry-after-ms only. HTTP-date Retry-After still yields undefined and the 500/1000ms fallback. Same as vendor; not a regression.

Notes

• Remapped xAI/Codex 429 with retryAfterMs: 45_000 now waits MAX_BLIND_WAIT_MS (30s) rather than aborting or using 500ms. Documented by the updated tests; a header longer than 30s without quota body markers can still under-wait and burn the remaining attempts.
• GATEWAY_OVERLOAD_USER_MESSAGE still says “retrying…” and can surface terminally. Out of this diff.