feat(model-eval): per-provider preflight tolerance (one bad candidate no longer aborts the pilot) — part of #2819 - #2824
Conversation
…te can't abort the pilot The maint-78 pilot aborts if ANY candidate produces no schema-valid rows in preflight. In run 30112277194 that meant github/codex-mini-latest (404) and the Anthropic temperature-400s killed the whole run even though OpenAI was fine. Add partition_usable_candidates(): drop unusable candidates per provider (keeping a provider only when its incumbent AND >=1 candidate are usable), and fail only when no provider has a usable incumbent+candidate pair. Dropped providers/models are reported to stderr. This lets the pilot still compare the providers that work. Implements the per-candidate preflight-tolerance part of #2819. ruff/black/mypy (CI-pinned) clean; 15 tests pass incl. 4 new (github-incumbent drop, no-usable-candidate drop, fatal-when-none, no-op-when-all-usable). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
|
Warning Review limit reachedYou’ve reached a temporary PR review limit under our Fair Usage Limits Policy. Next review available in: 36 minutes Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (2)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
Workflow source neededPR #2824 needs either a linked GitHub issue or one valid non-issue Workflow Source before PR metadata automation can manage it safely. Please do one of:
Once a valid source is present, this warning will not be reposted. |
Automated Status SummaryHead SHA: 39ab37a
Coverage Overview
Coverage Trend
Top Coverage Hotspots (lowest coverage)
Low Coverage Files (<50.0%)
Updated automatically; will refresh on subsequent CI/Docker completions. Keepalive checklistScopeNo scope information available Tasks
Acceptance criteria
|
![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: 7487bbd7ca
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 unusable: | ||
| print( | ||
| "pilot preflight error: candidates produced no schema-valid rows: " | ||
| + ", ".join(unusable), | ||
| file=sys.stderr, | ||
| ) | ||
| return 1 | ||
| candidates, dropped, fatal = partition_usable_candidates(candidates, unusable) |
There was a problem hiding this comment.
When every model passes preflight, this guard skips partition_usable_candidates() entirely, so providers without a candidate are still sent through the full pilot. In the checked-in config/model_eval_candidates.json, github-models has only the codex-mini-latest incumbent; once that model is available, the workflow will spend 30 evaluations on an unpaired baseline and include it in an artifact that the new invariant says should contain only incumbent-plus-candidate providers. Apply the partition even when unusable is empty, while retaining the conditional only for failure reporting.
Useful? React with 👍 / 👎.
… id → 404) (#2825) * fix(llm): namespace github-models model ids for the GA endpoint The GitHub Models GA endpoint (models.github.ai/inference) requires publisher-namespaced ids (e.g. openai/gpt-5). A bare id returns "404 page not found" — which is why github-models/codex-mini-latest failed the maint-78 pilot (run 30112277194). Add _github_model_id(): default a bare id to the openai/ publisher; already-namespaced ids pass through unchanged. Scope/honesty: this fixes the bare-id-on-GA-endpoint 404 CLASS. If a specific id (e.g. codex-mini-latest) is not in the GitHub Models catalog at all, namespacing won't help and the durable fix is an owner-reviewed change of the github-models selection to a catalogued id (openai/gpt-5 is already catalogued). Per-candidate preflight tolerance (#2824) already prevents this from aborting a whole pilot. Updated two tests that asserted verbatim pass-through (a pre-GA assumption) to the namespaced expectation. ruff/black/mypy (CI-pinned) clean; 63 tests pass incl. 4 new. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore(autofix): formatting/lint * chore(codex-autofix): apply updates (PR #2825) * fix(ci): restage black-formatted test + update github-model assertion for GA namespacing Two follow-ups to the #2825 merge: - tests/tools/test_langchain_client.py: commit the black-formatted version (the post-merge black ran in the worktree but the staged pre-format version was committed, so CI black flagged 'would reformat'). - tests/scripts/test_task_decomposer.py: the github client model kwarg is now publisher-namespaced (openai/...), so assert against _github_model_id(...) of the configured model rather than the bare id. ruff/black (CI-pinned) clean on the whole tree; affected test files pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
1 participant
What
Makes the maint-78 verifier pilot resilient: a single unusable candidate/provider no longer aborts the whole run.
Today
main()returns 1 if any candidate produces no schema-valid rows in preflight. Run30112277194(2026-07-24) aborted becausegithub-models/codex-mini-latest404'd and the newer Anthropic models 400'd on temperature — even though the OpenAI candidates were fine.How
partition_usable_candidates(candidates, unusable)drops unusable candidates per provider, keeping a provider only when its incumbent AND ≥1 candidate are usable (you can't benchmark without a baseline or without something to compare). Dropped providers/models are logged to stderr.Why
Implements the per-candidate preflight-tolerance part of #2819. Combined with #2822 (Anthropic temperature) and a github-404 fix, a pilot can now compare whatever providers work instead of all-or-nothing.
Verification
ruff==0.15.20/black==26.5.1 -l100/mypy==2.1.0(CI-pinned) clean;pytest tests/tools/test_run_model_eval_pilot.py→ 15 passed, incl. 4 new: github-incumbent-404 → drop github + keep openai/anthropic; provider with no usable candidate dropped; fatal when nothing survives; no-op when all usable.🤖 Generated with Claude Code