Remap Codex short 429s to retryable rate limits

[TheGreatAxios]

Summary

• Known-Codex short HTTP 429s now normalize to retryable rate limits instead of quota exhaustion
• Codex usage_limit_reached responses keep the reset/profile quota-exhausted copy
• Terminal credential failures now tell the operator to log in again instead of claiming re-authentication is in flight

Verification

• bun test ./src/agent/retry-policy.test.ts ./src/inference-gateway-error.test.ts ./src/inference-error-message.test.ts ./src/tui/stream-event-map.test.ts
• bun run check

Fixes CL-7173
Fixes CL-7174

[linear-code]

CL-7173

CL-7174

[TheGreatAxios]

Lens: critic. Verdict: no blocking findings.

This branch remaps known-Codex short 429s to retryable rate limits, keeps usage_limit_reached responses on the quota path, and changes terminal credential-failure copy to ask for login again. The branch includes retry-policy, normalization, message, and TUI copy coverage. I found no blocking or should-fix correctness issues.

[TheGreatAxios]

Review · Comment

Known-Codex short 429s remap to retryable rate limits; usage_limit_reached stays quota-exhausted; terminal credential copy asks the operator to log in again.

No findings.

[TheGreatAxios]

Critic · Comment

Known-Codex short 429s remap to retryable on the same dual-path as xAI. Terminal credential copy no longer claims a refresh is in flight.

Findings

• src/inference-error-message.ts:25 — every credential_failure uses "Authentication failed — log in again." There is still one string for recovered and terminal 403s, and no test pins the difference. A leftover recovered 403 row (handoff expires on anything other than inference.retry / inference.start) tells the operator to log in even if the turn continues. Retraction belongs to PR #711; this branch still lacks the recovered-vs-terminal copy test CL-7174 listed.

Notes

• Remap, backoff, usage_limit_reached, and non-Codex 429s match CL-7173.

[TheGreatAxios]

Neckbeard · Comment

Codex short-429 transcript copy reuses the xAI remap string.

Findings

• src/inference-gateway-error.ts:52 — Codex returns XAI_RATE_LIMIT_USER_MESSAGE while the JSDoc still says known-xAI. Rename or alias so a later xAI-only edit cannot retarget Codex.
• src/inference-error-message.ts:98 — two consecutive ifs return the same constant; the comment already treats them as one case.
• src/agent/retry-policy.test.ts:76 — title says "bare 429"; the fixture is ChatGPT usage-limit prose.

[TheGreatAxios]

Review · Comment

Remaps known-Codex HTTP 429s that lack a usage_limit_reached body to retryable, and changes the credential_failure transcript line to "Authentication failed — log in again."

Findings

• src/inference-gateway-error.ts:301-314 — the remap only exempts bodies parseCodexUsageLimitError recognizes, which needs an exact code field (src/auth/codex/usage-limit-error.ts:76-79). A prose-only "usage limit" body with an hour-long Retry-After becomes retryable, bypassing the MAX_BLIND_WAIT_MS abort (src/agent/retry-policy.ts:41-47) and hitting the exhausted account twice more before aborting. The xAI path (:241-254) keeps quota_exhausted on prose markers. CL-7173 asks for this, so it is a deliberate choice — but src/inference-gateway-error.test.ts:342-353 pins the remap of usage-limit prose, so if OpenAI ships the real cap as prose this becomes a silent regression relative to the xAI approach. Could the Codex path keep the marker check for the quota_exhausted case as well?
• src/inference-error-message.ts:99 — on the final aborted attempt the transcript still reads "Rate limited — retrying…". Same shape as the pre-existing xAI line at :98; noting because it is the same class CL-7174 is about.

Notes

• credential_failure copy is global, not Codex-only. Verified no refresh-on-401 in the inference path (grep -n "401\|credential\|refresh" src/agent/director.ts), so "log in again" is accurate.
• Checks: bun run typecheck, bun run lint, bun test src/agent/retry-policy.test.ts src/inference-gateway-error.test.ts src/inference-error-message.test.ts src/tui/stream-event-map.test.ts (74 pass). Repro of the prose-body case: retryAfterMs 3.6M → retry/retry/abort under codex, immediate abort under openai.
• Commit 57b479b4 "Pin TUI copy for Codex short 429s and credential failure" also touches src/agent/retry-policy.test.ts.