fix(codex-executor): handle response.incomplete + raise output-token budget#38
Merged
Merged
Conversation
…budget ~43% of recent GPT-5.5 reviews in dotCMS/core were posting "❌ Codex Review failed — job failed before producing output." Root cause: on the Responses API, max_output_tokens caps the COMBINED reasoning + visible-answer tokens, not just the answer. At reasoning_effort=medium, GPT-5.5 sometimes spends the whole 2048-token budget thinking and the stream ends status=incomplete (incomplete_details.reason=max_output_tokens) with ZERO output_text.delta. The executor captured usage only on response.completed and only logged errors on response.failed/error, so an incomplete response fell through to an empty review -> sys.exit(1) -> generic failure sticky. Non-deterministic by reasoning load, not diff size (a 137-line diff failed while a 208-line diff passed). Fix: - Raise max_output_tokens default 2048 -> 8000 so medium-effort reasoning plus the answer both fit. - Capture usage/status from the response.incomplete (and response.failed) terminal events, not just response.completed. - Retry once when the answer is empty *because* of max_output_tokens: bump the budget (>=16000) and drop reasoning_effort to low so the visible answer fits. - If still empty, post a clear truncation diagnostic and exit 0 (sticky shows the reason) instead of a generic "review failed". Keep + flag partial answers. - Correct the stale "max_output_tokens does NOT cap reasoning tokens" comments. No consumer interface change. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
This was referenced Jun 13, 2026
riccardoruocco
pushed a commit
to riccardoruocco/core
that referenced
this pull request
Jun 16, 2026
## Summary Bumps the `dotCMS/ai-workflows` pin in the AI review workflows from `@v3.1.2` to **`@v3.1.4`** (orchestrator + backend reviewer). Supersedes the intermediate v3.1.3 bump — v3.1.4 includes everything in v3.1.3 plus the outcome-signaling improvements, so we go straight to it. ## What's in v3.1.3 + v3.1.4 **v3.1.3 — silent-failure fix ([ai-workflows#38](dotCMS/ai-workflows#38 ~43% of GPT-5.5 reviews were posting "❌ Codex Review failed — job failed before producing output." Root cause: `max_output_tokens` caps reasoning+answer combined, so medium-effort GPT-5.5 sometimes spent the whole budget reasoning and returned `incomplete` with no text. Fix: budget 2048→8000, handle `response.incomplete`, retry once with a bigger budget + lighter reasoning. **v3.1.4 — clear outcome signaling ([ai-workflows#39](dotCMS/ai-workflows#39 - Sticky header reflects the outcome: `🤖 Codex Review` / `⚠️ truncated` / `❌ no output` / `⏱️ canceled` - The job **fails (red ✗ in checks)** when no review is produced — surfaces the outcome without gating merges (advisory review) - Canceled / timed-out runs rewrite the sticky to `⏱️ Codex Review canceled` instead of leaving it stuck on `🔄 in progress` ## Validation - v3.1.3 before/after e2e: steve-quarterly-planning dotCMS#105 (@v3.1.2 failed, @v3.1.3 recovered) - v3.1.4 signaling e2e: steve-quarterly-planning dotCMS#106 (🤖+green confirmed; ⏱️ cancellation confirmed; retry makes the ❌ path a hardened safety net) Closes: dotCMS#36158 --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
~43% of recent GPT-5.5 automatic reviews in
dotCMS/coreposted "❌ Codex Review failed — job failed before producing output" (e.g. core PRs #36150, #36149, #36144, #36130, #36124, #36112). Those PRs got no review, not a false "clean".Root cause
On the OpenAI Responses API,
max_output_tokenscaps the combined reasoning + visible-answer tokens — not just the answer. Atreasoning_effort: medium, GPT-5.5 sometimes spends the entire2048budget reasoning, and the stream ends withstatus=incomplete(incomplete_details.reason=max_output_tokens) and zerooutput_text.deltaevents.mantle_review.pycaptured usage only onresponse.completedand only logged errors onresponse.failed/error, so anincompleteterminal event fell through everything → empty review →sys.exit(1)→ the generic failure sticky. The failure signature in the logs isTokens: in: ? · out: ?(usageNone) with no::error::line. It's non-deterministic by reasoning load, not diff size (a 137-line diff failed while a 208-line diff passed).Fix
max_output_tokensdefault2048 → 8000so medium-effort reasoning + the answer both fit.response.incomplete(andresponse.failed), not justresponse.completed.max_output_tokens: bump the budget (≥16000) and dropreasoning_efforttolowso the visible answer fits.max_output_tokensdoes NOT cap reasoning tokens" comments.No consumer interface change. → release as v3.1.3.
Validation
mantle_review.pycompiles.dotCMS/steve-quarterly-planning(linked after the tag is cut): reproduce an incomplete/empty review, confirm v3.1.3 produces a real review.