Back off Codex review triggers on usage limits

[Komzpa]

Summary

• stop the Codex label sync from posting more @codex review comments when the same GitHub comment sender received a recent Codex usage-limit reply
• allow review requests again only when that same sender account has a newer normal Codex response within the 24-hour window
• make no-data runs probe once, reread the PR timeline, and suppress the remaining review requests if that probe hits the usage limit

Tests

• uv run pytest tests/unit/test_sync_codex_ok_labels.py -q
• uv run ruff check .github/scripts/sync_codex_ok_labels.py tests/unit/test_sync_codex_ok_labels.py
• uv run ruff format --check .github/scripts/sync_codex_ok_labels.py tests/unit/test_sync_codex_ok_labels.py
• make test-unit

[chatgpt-codex-connector]

You have reached your Codex usage limits. You can see your limits in the Codex usage dashboard.

[Komzpa]

@codex review

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Soju06]

The backoff logic itself looks sound (per-sender attribution, newer-normal-response unlatch, probe-and-reread), but splitting classification from apply introduced a log misattribution bug:

Stale number in the apply loop — .github/scripts/sync_codex_ok_labels.py:1687 and the except handler at :1706 still format f"{repo}#{number}", but the apply loop now iterates for decision in decisions: while number is the leftover from the classification loop (always the last classified PR). With --all-open over PRs 710 and 714, both status lines print apply Soju06/codex-lb#714 ..., and any exception during apply is attributed to #714 regardless of which PR failed. These lines are the script's audit trail, so this actively misleads. Should be decision.number. Note test_main_stops_codex_review_triggers_after_probe_hits_usage_limit passes only because it asserts on #714, the last PR — asserting apply Soju06/codex-lb#710 appears in output would catch it.

Minor: the PR description says no-data runs "probe once", but the sleep+reread at :1678-1682 runs after every posted @codex review, not just the first — with N eligible PRs and no limit hit that's N x 10s plus N extra timeline fetches per run. Consider probing only until the first non-limit observation, or only when latest_normal_response_at is still None.

[chatgpt-codex-connector]

To use Codex here, create an environment for this repo.