Fix Design Decision Gate false failures caused by the LLM invocation cap - #52836
Merged
pelikhan merged 2 commits intoAug 15, 2026
Merged
Conversation
Co-authored-by: pelikhan <4175913+pelikhan@users.noreply.github.com>
Copilot
AI
changed the title
[WIP] Investigate Design Decision Gate merge-blocking failure
Fix Design Decision Gate false failures caused by the LLM invocation cap
Aug 15, 2026
Contributor
PR TriageCategory: bug | Risk: high | Score: 82/100
Recommended action:
|
Contributor
PR Triage
Merge-blocking Design Decision Gate false failures with real token spend impact. CI unknown but high urgency — recommend expedited human review. Automated triage — see [PR Triage Report] for full context.
|
Contributor
There was a problem hiding this comment.
Pull request overview
Addresses #52736 by reducing false failures when workflows exhaust their LLM invocation allowance after producing useful output.
Changes:
- Raises the Design Decision Gate limit from 20 to 30.
- Suppresses invocation-cap failures after terminal safe output.
- Adds coverage for Claude, Codex, and Copilot harnesses.
Show a summary per file
| File | Description |
|---|---|
.github/workflows/design-decision-gate.md |
Raises the turn limit. |
.github/workflows/design-decision-gate.lock.yml |
Recompiles the workflow. |
actions/setup/js/claude_harness.cjs |
Adds cap-failure suppression. |
actions/setup/js/claude_harness.test.cjs |
Tests Claude suppression. |
actions/setup/js/codex_harness.cjs |
Adds cap-failure suppression. |
actions/setup/js/codex_harness.test.cjs |
Tests Codex suppression. |
actions/setup/js/copilot_harness.cjs |
Adds cap-failure suppression. |
actions/setup/js/copilot_harness.test.cjs |
Tests Copilot suppression. |
Review details
💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.
- Files reviewed: 8/8 changed files
- Comments generated: 3
- Review effort level: Balanced
| pull-requests: read | ||
| issues: read | ||
| max-turns: 20 | ||
| max-turns: 30 |
Comment on lines
+543
to
+545
| if (nonRetryableGuard.maxRunsExceeded && safeOutputsPath && hasExpectedSafeOutputs(safeOutputsPath, { logger: log })) { | ||
| log(`attempt ${attempt + 1}: invocation cap saturated but safe-outputs already contain expected output — suppressing terminal verdict (false-red: core work succeeded)`); | ||
| return { action: "stop", exitCode: 0 }; |
Comment on lines
+657
to
+659
| if (nonRetryableGuard.maxRunsExceeded && safeOutputsPath && hasExpectedSafeOutputs(safeOutputsPath, { logger: log })) { | ||
| log(`attempt ${attempt + 1}: invocation cap saturated but safe-outputs already contain expected output — suppressing terminal verdict (false-red: core work succeeded)`); | ||
| return { action: "stop", exitCode: 0 }; |
pelikhan
deleted the
copilot/deep-report-investigate-merge-blocking-failure
branch
August 15, 2026 13:02
Contributor
|
🎉 This pull request is included in a new release. Release: |
3 tasks
sameerkhansf
added a commit
to sameerkhansf/Sameer-Khan
that referenced
this pull request
Sep 8, 2026
Run 34261048698 was the first run to reach the post-writing path since #76 — the gate correctly printed CLEAR — and it failed on the invocation cap: attempt 1: 56 turns / 93 invocations, then "Upstream error from Nvidia: Service temporarily overloaded" (transient, classified partial_execution, budget remaining) attempt 2: harness restarts the session from scratch, dies on "429 Maximum LLM invocations exceeded (120 / 120)" max-turns is not only a turn count: it is the AWF api-proxy's hard per-run LLM invocation cap, and that budget is pooled across every harness retry attempt (github/gh-aw#52836, #45827). At 120 a single pass fits and a retry cannot, so any transient upstream blip on the free NVIDIA endpoint kills the run. 250 leaves room for one full pass plus a retry. This is a symptom fix with a known ceiling. gh aw audit has flagged "about 50% of this run's turns appear to be data-gathering that could move to deterministic steps" on every run, and the log shows the cause: repeated `curl <hf-readme> | grep -i "parameters|context|license|price"` against the same files. Pre-fetching candidate READMEs in the `steps:` block is the DeterministicOps fix and the next lever if the cap is hit again. Claude-Session: https://claude.ai/code/session_01Vru5xPN55JPU7dyVS3uswd Co-authored-by: Claude Opus 5 (1M context) <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
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.
Three independent monitors flagged the merge-blocking
Design Decision Gate 🏗️workflow failing at a high rate while still burning real token spend — aTurns=0-style crash signature that turned out to be something else entirely.Root cause
max-turnsdoubles as the AWF API proxy's hard per-run LLM invocation cap. This gate's ADR-review task typically needs ~20-22 turns to complete, butmax-turnswas set to 20 — so it reliably hit429 Maximum LLM invocations exceeded (20/20). Worse, the agent had often already written a valid safe-output (e.g. anadd_commentwith the completed ADR review) before hitting the cap, but every harness (claude/codex/copilot) treated the cap error as an unconditional hard failure and discarded that completed work.Changes
max-turnsfrom 20 → 30 indesign-decision-gate.md(recompiled lock file) to give headroom above the observed completion point.claude_harness.cjs,codex_harness.cjs,copilot_harness.cjs: when the invocation cap is saturated but safe-outputs already contain a terminal/expected result, exit 0 instead of failing — mirroring the existing suppression pattern already used for "numerous permission-denied" and "partial execution" cases.