Prioritize review dispatch for bot-authored PRs that modify agent definitions

[fullsend-ai-retro]

What happened

PR #2857, a bot-authored fix that modifies the review agent's own severity calibration rules, was merged with only a single human approval and zero review agent involvement. The dispatch authorization gate (ADR 0054) rejects bot PRs because fullsend-ai-coder[bot] lacks write collaborator permission, and the label-based fallback doesn't fire because GitHub suppresses events from app tokens. Issue #2674 already tracks the general dispatch gap. This retro confirms the gap is actively causing unreviewed merges of agent-behavior-modifying code — PR #2857 changed intent-coherence.md and SKILL.md, files that directly control review agent behavior.

What could go better

Changes to agent definitions (skills, sub-agent prompts, dimension mappings) are among the highest-impact changes in this repo because they alter how all future reviews behave. These changes getting zero automated review creates a blind spot: if the code agent introduces a subtle error in severity guidance or prompt wording, no automated check catches it. Issue #2674 already proposes fixing the dispatch authorization gate, but that issue doesn't call out the specific risk of agent-definition changes going unreviewed. Confidence is high that this is a recurring pattern — PRs #2885 and #2881 (also bot-authored) similarly lacked review agent runs.

Proposed change

Add urgency context to issue #2674 noting that bot PRs modifying internal/scaffold/fullsend-repo/skills/pr-review/ are particularly high-risk to leave unreviewed. As an interim measure until #2674 is resolved, add a CODEOWNERS entry requiring a second human reviewer for paths under internal/scaffold/fullsend-repo/skills/pr-review/ — this ensures changes to review agent behavior get at least two human reviewers when the review agent itself cannot participate. The CODEOWNERS entry should require review from a team or individual with domain expertise in the review agent's behavior.

Validation criteria

Short-term: after adding the CODEOWNERS entry, verify that the next bot-authored PR touching internal/scaffold/fullsend-repo/skills/pr-review/ requires and receives two human approvals before merge. Long-term: after #2674 is resolved, verify the next 3 bot-authored PRs all receive review agent runs by checking for fullsend-ai-review[bot] comments.

---

Generated by retro agent from #2857

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 9:47 AM UTC · Completed 9:51 AM UTC
Commit: 0a95cac · View workflow run →

[fullsend-ai-triage]

Triage Summary

Severity: High — agent-definition changes going unreviewed is a governance blind spot.

This issue identifies a specific high-risk consequence of the dispatch authorization gap tracked in #2674: bot-authored PRs modifying review agent behavior files (severity calibration, skill definitions, sub-agent prompts) are merging with minimal human oversight and zero automated review.

Root Cause

The ADR 0054 authorization gate silently skips review dispatch for bot accounts because they lack write collaborator permission. While #2674 tracks the general dispatch gap, this issue highlights that agent-definition files have no compensating control.

Proposed Fix (Interim)

1. Add a CODEOWNERS entry requiring a second human reviewer for internal/scaffold/fullsend-repo/skills/pr-review/ paths
2. Add urgency context to dispatch: ADR 0054 authorization gate blocks review dispatch for coder-bot PRs #2674 noting the elevated risk for agent-definition changes

Related Work

• dispatch: ADR 0054 authorization gate blocks review dispatch for coder-bot PRs #2674 — root cause: dispatch auth gate blocks bot PRs (long-term fix)
• code-review skill: mandatory finding when diff modifies review agent behavior files #1309 — complementary: mandatory review finding when diffs touch review agent behavior files

Proposed Test Case

After adding the CODEOWNERS entry:
1. Create or observe a bot-authored PR modifying files under
   internal/scaffold/fullsend-repo/skills/pr-review/
2. Verify GitHub requires two human approvals before merge
3. Edge cases:
   - Nested paths (sub-agent definitions) are covered
   - Pattern applies equally to bot and human PRs

The interim CODEOWNERS mitigation is independently actionable now. The long-term resolution depends on #2674 landing the dispatch auth gate fix.

---

Labels: Issue concerns review agent dispatch governance for agent-definition files, touching dispatch and skills components with security implications.