Skip to content

Harden AI review quality: subagent dispatch, inline-only output, no wall-of-text#19772

Merged
T-Gro merged 1 commit into
mainfrom
review-quality-hardening
May 20, 2026
Merged

Harden AI review quality: subagent dispatch, inline-only output, no wall-of-text#19772
T-Gro merged 1 commit into
mainfrom
review-quality-hardening

Conversation

@T-Gro
Copy link
Copy Markdown
Member

@T-Gro T-Gro commented May 20, 2026

Problem

Audit of 17 AI reviews showed 53% had zero inline comments — the agent dumps analysis as a wall-of-text review body instead of posting actionable inline comments at file+line.

Quality Tier Count %
🟢 Inline-focused (ideal) 3 18%
🟡 Inline + wall (OK) 4 24%
🔴 LGTM wall (useless avg 1500 chars) 5 29%
🔴 Feedback wall, no inline 4 24%

Changes

expert-reviewer.md

  • Wave 0: MUST dispatch independent subagents per dimension (not optional)
  • Wave 5: empty body for findings (just inline comments), LGTM for approve
  • New delivery rule: body >10 chars (other than LGTM) is a bug
  • Removed "See inline comments" boilerplate

reviewing-compiler-prs/SKILL.md

  • Replaced aspirational "Multi-Model Dispatch" with concrete Subagent Dispatch
  • Each dimension gets independent background agent with CHECK rules + filtered diffs
  • Subagents produce structured {file, line, severity, dimension, issue, suggestion}
  • Expert-reviewer consolidates, deduplicates, applies assessment gates, posts inline

…text

Evidence from 17 AI review audit:
- 53% had zero inline comments (wall-of-text body only)
- 5 reviews were LGTM verdicts buried in avg 1500 chars of prose
- Only 3/17 were inline-focused (ideal)

expert-reviewer.md:
- Wave 0: MUST dispatch independent subagents per dimension via task tool
- Wave 5: empty body for findings (just inline comments), 'LGTM' for approve
- New hard rule: body >10 chars (other than LGTM) is a bug
- Removed 'See inline comments' boilerplate — people have eyes

reviewing-compiler-prs/SKILL.md:
- Replaced aspirational 'Multi-Model Dispatch' with concrete 'Subagent Dispatch'
- Each dimension gets independent background agent with CHECK rules + filtered diffs
- Subagents produce structured {file, line, severity, dimension, issue, suggestion}
- Expert-reviewer consolidates, deduplicates, applies assessment gates

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@github-actions
Copy link
Copy Markdown
Contributor

✅ No release notes required

@github-actions github-actions Bot added the AI-Tooling-Check-Bypassed Tooling check: non-fork PR, not diff-analyzed label May 20, 2026
@T-Gro T-Gro requested a review from abonie May 20, 2026 12:59
@T-Gro T-Gro enabled auto-merge (squash) May 20, 2026 12:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

AI-Tooling-Check-Bypassed Tooling check: non-fork PR, not diff-analyzed

Projects

Archived in project

Development

Successfully merging this pull request may close these issues.

2 participants