fix(code-review): short-circuit exhausted peer routes - #1344
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0b25a97a91
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if candidate.timestamp() <= now + 60: | ||
| candidate += timedelta(days=1) |
There was a problem hiding this comment.
When the provider reports an absolute reset less than 60 seconds in the future, this condition advances the reset to tomorrow before adding the intended one-minute buffer. For example, at 15:49:30 a resets 3:50pm response opens the circuit for the full 24-hour cap instead of 90 seconds, unnecessarily disabling Claude peer reviews for the rest of the day; only advance the date when the reset time has actually passed.
Useful? React with 👍 / 👎.
| try: | ||
| zone = ZoneInfo(zone_name) if zone_name else datetime.now().astimezone().tzinfo | ||
| except ZoneInfoNotFoundError: | ||
| zone = datetime.now().astimezone().tzinfo |
There was a problem hiding this comment.
On native Windows installations without the optional tzdata package, ZoneInfo("America/New_York") raises ZoneInfoNotFoundError, so this fallback silently interprets the provider's explicitly New York reset time in the machine's local timezone. A user outside that zone can therefore remain blocked for several extra hours—or, when the misinterpreted time has passed, up to the 24-hour cap—despite native Windows being a supported target; use a source that can resolve the named zone or fall back conservatively rather than changing its meaning.
AGENTS.md reference: AGENTS.md:L55-L55
Useful? React with 👍 / 👎.
Preflight Claude authentication and persist a privacy-safe, reset-aware session-quota circuit so later reviews exclude a known-exhausted route before packaging or egress. Keep transient 429, execution-context auth, and transport failures distinct while preserving fixed-recipient behavior. Fixes EveryInc#1343 Co-Authored-By: Codex <codex@openai.com> Agent-Provenance: host=conductor harness=codex model=gpt-5.6-sol
0b25a97 to
2343c6b
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2343c6bc54
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| def _classify(text: str) -> str: | ||
| lowered = text.lower() | ||
| if "session limit" in lowered or "usage limit" in lowered: |
There was a problem hiding this comment.
When Claude reports quota exhaustion as You've hit your limit · resets ..., the text contains neither session limit nor usage limit, so _classify falls through to other. cmd_record then never persists a retry epoch, and every subsequent review sends another doomed API-bearing request instead of short-circuiting the route; recognize the unqualified limit form while tying it to reset or 429 evidence to avoid false positives.
Useful? React with 👍 / 👎.
| MAX_QUOTA_TTL = 24 * 3600 | ||
| SAFE_ROUTE = re.compile(r"^[A-Za-z0-9._-]+$") | ||
| RESET_AT = re.compile( | ||
| r"\bresets?\s+(?:at\s+)?(\d{1,2}):(\d{2})\s*(am|pm)" |
There was a problem hiding this comment.
For round-hour diagnostics such as resets 4pm, this expression does not match because it requires :MM, so _retry_after uses the one-hour fallback rather than the provider's absolute reset. That can retry an exhausted route hours too early or suppress it after it is available; make the minute group optional and default it to 00.
Useful? React with 👍 / 👎.
1 participant
Fixes #1343
Problem
The code-review cross-model route treated an installed Claude CLI as available even after the account had exhausted its session quota. It discovered the known-unhealthy route only after packaging and launching the adversarial review, then repeated that late failure on later invocations.
Fix
claude auth status --jsonbefore review payload construction;ce-code-reviewretains its local adversarial reviewer;resets 3:50pm (America/New_York)diagnostic shape.Claude exposes authentication state but no zero-cost quota-remaining endpoint. The first newly exhausted session still requires an API-bearing request; every subsequent review learns the unhealthy route before code leaves the host or a peer job displaces the local fallback.
Validation
git diff --checkpassed.Security Disclosure
Security-relevant state handling changed. The circuit stores no prompt, diff, raw diagnostic, account, or credential data. Its bounded state is owner-private, symlink/ownership checked, lock-serialized, and atomically published. A state-integrity failure fails closed before provider egress.
Agent Disclosure
Codex · GPT-5.6 Sol · via Conductor authored the implementation and tests.
🤖 Generated with Codex · GPT-5.6 Sol · via Conductor