Review agent: detect issues introduced by fix commits

[fullsend-ai-retro]

What happened

PR #1178 went through 3 review rounds. In round 2 (May 21), the fix for the empty-Sanitized edge case introduced a dead-code guard (if len(result.Findings) == 0 inside a if len(result.Findings) > 0 block). This was only caught in round 3 (May 26) by waynesun09's 4-agent review squad. The dead code itself was a logic error that could have masked a real sanitization bypass. Related existing issues: #1088 (self-review loop), #902 (circuit breaker for review-fix loop), #1368 (approval with residual findings).

What could go better

When a review agent evaluates a fix commit that addresses prior review findings, it should specifically check whether the fix introduced new issues — particularly dead code, unreachable branches, or logic contradictions near the changed lines. Currently, each review round treats the diff as a fresh review without focused attention on fix-induced regressions. Medium confidence: this pattern (fix introduces new bug caught in next round) occurred once here but is architecturally predictable for multi-round reviews. The existing issues (#1088, #902) acknowledge the broader problem but don't specifically address fix-regression detection.

Proposed change

In the review agent's prompt or skill (likely in the reusable-review.yml workflow or the review agent definition in fullsend-ai/fullsend), add a review dimension for fix-induced regressions. When the review detects that the current push addresses prior review comments (e.g., by checking if the PR has prior CHANGES_REQUESTED reviews), the agent should specifically scan for: (1) dead code or unreachable branches in the modified regions, (2) logic contradictions between guard conditions, (3) whether the fix actually resolves the original finding or just moves the problem. This could be a focused sub-prompt added to the review skill's checklist rather than a new agent.

Validation criteria

Over the next 10 PRs that go through 2+ review rounds, track whether fix-induced regressions are caught in the same round they're introduced (round N) rather than requiring an additional round (round N+1). Success: at least 80% of fix-induced issues caught in the same review round as the fix.

---

Generated by retro agent from #1178

[fullsend-ai-triage]

Triage Summary

Severity: medium | Category: feature

This issue proposes adding fix-regression detection to the review agent so that issues introduced by fix commits are caught in the same review round rather than requiring an additional cycle.

Problem

The review agent's code-review skill evaluates all dimensions uniformly, with no special attention to whether a push is a fix for prior review findings. The pr-review skill (step 2a) already computes the delta since the prior review and passes prior findings forward, but code-review doesn't cross-reference fixes against original findings or scan for fix-induced defects like dead code or logic contradictions.

Recommended Approach

Add a fix-regression sub-section to the Correctness dimension in skills/code-review/SKILL.md. When prior review context is present, the skill should:

1. For each prior finding marked as addressed, verify the fix actually resolves the issue
2. Scan modified regions for dead code, unreachable branches, and guard-condition contradictions introduced by the fix
3. Check whether the fix merely moves the problem rather than eliminating it

The pr-review skill's step 2a already provides the necessary context (prior findings + changed-file list). The scaffold review agent definition (internal/scaffold/fullsend-repo/agents/review.md) may also need minor updates to reference this new sub-dimension.

Proposed Test Case

Scenario: Fix-induced dead code detected in same review round

Setup:
  - PR with a finding: "missing nil check on line 42"
  - Fix commit adds nil check inside an existing `if x != nil` block,
    creating an unreachable branch

Expected:
  - Review agent flags the dead code / unreachable branch as a new
    finding in the same round as the fix

Edge cases:
  - Fix introduces unreachable else branch
  - Fix duplicates a parent guard condition
  - Fix resolves original finding but introduces unrelated logic error
    in the same function

Related Issues

• code agent: invoke review agent in a pre-PR self-review loop (max 3 iterations) #1088 (self-review loop) — broader scope, not blocking
• Add circuit breaker to review-fix feedback loop #902 (circuit breaker for review-fix loop) — complementary, not blocking
• Review agent approval with residual findings drives unnecessary author rework #1368 (approval with residual findings) — related but distinct problem

---

Labels: This is a review agent feature request targeting the code-review and pr-review skills, with medium priority given it optimizes multi-round reviews but is not blocking functionality.