-
Notifications
You must be signed in to change notification settings - Fork 329
[q] fix(deep-report): switch to claude engine and fix memory path (#23140) #23747
Description
Q Workflow Optimization Report
Issues Found (from live data)
Workflow: DeepReport - Intelligence Gathering Agent
Run IDs Analyzed: 23603033388, 23653655633, 23753118444, 23805651089
Issue 1: Engine blocked by OpenAI safety mitigations (Critical)
The codex engine (OpenAI) is being blocked in runs 23753118444 and 23805651089 with:
stream disconnected before completion: This user's access to this model has been temporarily limited for potentially suspicious activity related to cybersecurity.
The DeepReport prompt contains phrases like "Detect suspicious patterns", "Identify potential security concerns", and "Suspicious activity" which trigger OpenAI's safety classifier. This is a recurring failure — the same error appeared on March 30 and March 31.
Issue 2: Wrong repo-memory path in instructions
The workflow instructions told the agent to write memory files to:
- ❌
/tmp/gh-aw/repo-memory-default/memory/default/
But the compiled lock file sets GH_AW_MEMORY_DIR=/tmp/gh-aw/repo-memory/default/ with FILE_GLOB_FILTER=memory/deep-report/*.md. The correct path is:
- ✅
/tmp/gh-aw/repo-memory/default/memory/deep-report/
With the wrong path, the agent's memory writes were silently discarded — nothing was uploaded to the repo-memory-default artifact or pushed to the memory/deep-report branch.
Changes Made
.github/workflows/deep-report.md:
engine: codex→engine: claude— avoids OpenAI cybersecurity content blocks- Fixed memory path in Step 0 (check) and Step 4 (write):
/tmp/gh-aw/repo-memory-default/memory/default/→/tmp/gh-aw/repo-memory/default/memory/deep-report/
Expected Improvements
- ✅ Eliminates recurring engine failures caused by OpenAI safety mitigations
- ✅ Fixes silent memory loss — agent insights will now be correctly persisted and reused across runs
Validation
Workflow compiled successfully:
- ✅
deep-report.md— no errors, no warnings
Note: .lock.yml files are not included — they will be regenerated automatically after merge.
References
-
Issue: [aw] DeepReport - Intelligence Gathering Agent failed #23140
-
Run IDs: 23753118444, 23805651089
-
Fixes [aw] DeepReport - Intelligence Gathering Agent failed #23140
Warning
Protected Files — Push Permission Denied
This was originally intended as a pull request, but the patch modifies protected files. A human must create the pull request manually.
Protected files
The push was rejected because GitHub Actions does not have workflows permission to push these changes, and is never allowed to make such changes, or other authorization being used does not have this permission.
Create the pull request manually
# Download the patch from the workflow run
gh run download 23807356936 -n agent -D /tmp/agent-23807356936
# Create a new branch
git checkout -b fix/deep-report-engine-and-memory-path-10ceb60c656355fc main
# Apply the patch (--3way handles cross-repo patches)
git am --3way /tmp/agent-23807356936/aw-fix-deep-report-engine-and-memory-path.patch
# Push the branch and create the pull request
git push origin fix/deep-report-engine-and-memory-path-10ceb60c656355fc
gh pr create --title '[q] fix(deep-report): switch to claude engine and fix memory path (#23140)' --base main --head fix/deep-report-engine-and-memory-path-10ceb60c656355fc --repo github/gh-aw
- expires on Apr 2, 2026, 4:19 PM UTC