tiennm99/ccs
1
0
Fork 0
mirror of https://github.com/tiennm99/ccs.git synced 2026-09-02 12:19:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

hotfix(ci): fix AI review performance — bypass permissions, cap turns, reduce thinking

Browse Source 
- Add --permission-mode bypassPermissions to eliminate 16+ permission denials per review
- Add --max-turns 20 to prevent runaway 41+ turn sessions
- Replace ultrathink with think (32K→16K thinking tokens) — sufficient for read-only review
- Expand --allowedTools with commonly-tried tools (wc, head, tail, find, git show)
- Remove unnecessary tools (Edit, LS — review shouldn't edit files)
- Reduce timeout from 20 to 15 minutes for faster failure detection

Evidence from run #23659037151: 11.6 min, 41 turns, 16 permission denials, $1.95, failed to write review.

Closes #818
This commit is contained in:
Tam Nhu Tran
2026-03-27 14:26:21 -04:00
parent aaed6f39e0
commit 0a7fef0368
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/ai-review.yml
+6 -4
View File
@@ -127,7 +127,7 @@ jobs:
name: Claude Code Review name: Claude Code Review
needs: prepare needs: prepare
if: needs.prepare.result == 'success' if: needs.prepare.result == 'success'
timeout-minutes: 20 timeout-minutes: 15
runs-on: ${{ fromJSON(needs.prepare.outputs.runs_on) }} runs-on: ${{ fromJSON(needs.prepare.outputs.runs_on) }}
permissions: permissions:
contents: read contents: read
@@ -147,7 +147,7 @@ jobs:
DISABLE_ERROR_REPORTING: '1' DISABLE_ERROR_REPORTING: '1'
DISABLE_TELEMETRY: '1' DISABLE_TELEMETRY: '1'
CLAUDE_CODE_MAX_OUTPUT_TOKENS: '64000' CLAUDE_CODE_MAX_OUTPUT_TOKENS: '64000'
MAX_THINKING_TOKENS: '32000' MAX_THINKING_TOKENS: '16000'
REVIEW_OUTPUT_FILE: pr_review.md REVIEW_OUTPUT_FILE: pr_review.md
REVIEW_COMMENT_FILE: .ccs-ai-review-comment.md REVIEW_COMMENT_FILE: .ccs-ai-review-comment.md
@@ -226,7 +226,7 @@ jobs:
show_full_output: true # Visible logs for debugging slow/failing reviews show_full_output: true # Visible logs for debugging slow/failing reviews
track_progress: false # Disabled - no progress comments, just final review track_progress: false # Disabled - no progress comments, just final review
prompt: | prompt: |
ultrathink think
REPO: ${{ github.repository }} REPO: ${{ github.repository }}
PR NUMBER: ${{ needs.prepare.outputs.pr_number }} PR NUMBER: ${{ needs.prepare.outputs.pr_number }}
@@ -257,7 +257,9 @@ jobs:
claude_args: | claude_args: |
--bare --bare
--model ${{ env.REVIEW_MODEL }} --model ${{ env.REVIEW_MODEL }}
--allowedTools "Edit,Glob,Grep,LS,Read,Write,Bash(gh pr diff *),Bash(gh pr view *),Bash(gh issue view *),Bash(git diff *),Bash(git log *),Bash(git status *),Bash(cat *),Bash(ls *)" --permission-mode bypassPermissions
--max-turns 20
--allowedTools "Glob,Grep,Read,Write,Bash(gh pr diff *),Bash(gh pr view *),Bash(git diff *),Bash(git log *),Bash(git show *),Bash(cat *),Bash(ls *),Bash(wc *),Bash(head *),Bash(tail *),Bash(find *)"
# Fallback: if Claude didn't write the review file, extract from execution output # Fallback: if Claude didn't write the review file, extract from execution output
- name: Extract review from execution output (fallback) - name: Extract review from execution output (fallback)