fix(#45): tighten review verdict criteria for medium findings

[fullsend-ai-coder]

Replace the subjective "multiple medium-severity findings which could affect the intended outcome" criterion with an objective rule: two or more medium findings, at least one identifying a functional bug, require request-changes.

Add a self-consistency check to the pr-review synthesis step: if the summary text states findings should be addressed before merge, the verdict must be request-changes, not comment.

Update comment-only criteria to clarify it applies only to stylistic, advisory, or process-related medium findings.

Changes:

• agents/review.md: Outcome section verdict criteria
• skills/pr-review/SKILL.md: Step 6f verdict logic and
  self-consistency check

Note: pre-commit could not run (shellcheck-py failed to install due to sandbox network restrictions). Secret scan passed.

---

Closes #45

Post-script verification

• Branch is not main/master (agent/45-fix-review-verdict-consistency)
• Secret scan passed (gitleaks — 2540ede96537b32447321564bec60226efca120e..HEAD)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[ben-alkov]

/fs-review

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 8:54 PM UTC · Completed 9:09 PM UTC
Commit: e8381e3 · View workflow run →

[fullsend-ai-review]

Review

Findings

Medium

• [logic-error] agents/review.md:238 — The new verdict criteria create a logical gap for exactly one medium finding that identifies a functional bug. The request-changes criterion requires "two or more" medium findings with at least one functional bug. The comment-only criterion explicitly excludes functional bugs ("not when any medium finding identifies a functional bug"). A single medium functional-bug finding satisfies neither condition, creating a contradiction. The same gap exists in the parallel criteria in skills/pr-review/SKILL.md.
  Remediation: Add explicit handling — either lower the request-changes threshold to include one functional-bug medium, broaden comment-only to permit it, or add a dedicated bullet for this case.

• [protected-path] agents/review.md, skills/pr-review/SKILL.md — Both files are under protected paths (agents/, skills/). The PR links to issue Review agent verdict contradicted its own analysis on PR #32: chose comment despite saying findings should be addressed before merge #45 and explains the rationale. Human approval is always required for protected-path changes.

Low

• [stale-reference] skills/code-review/SKILL.md:279 — The code-review skill retains the old verdict criterion ("Multiple medium findings which could affect the intended outcome") that this PR replaces in agents/review.md and skills/pr-review/SKILL.md. The skill's override clause (line 298) deferring to agents/review.md mitigates functional impact, but the textual inconsistency may cause confusion.

• [pattern-inconsistency] agents/review.md:238 — The new multi-part conditional introduces semantic content qualification into verdict criteria that previously used only severity counts. This is intentional per issue Review agent verdict contradicted its own analysis on PR #32: chose comment despite saying findings should be addressed before merge #45 but diverges from the established pattern.

• [pattern-inconsistency] skills/pr-review/SKILL.md:779 — Same pattern divergence in the parallel verdict criteria text.

---

Labels: PR fixes a review verdict bug and modifies protected agent/skill definitions

[fullsend-ai-review]

See the review comment for full details.

[ben-alkov]

/fs-fix lower the request-changes threshold to include one functional-bug medium finding; update outdated language in "skills/code-review"; ignore all other findings

[fullsend-ai-coder]

🤖 Finished Fix · ✅ Success · Started 9:58 PM UTC · Completed 10:05 PM UTC
Commit: e8381e3 · View workflow run →

[fullsend-ai-coder]

🔧 Fix agent — iteration 1 (human-triggered)

Addressed all 2 human-directed findings. Lowered the request-changes threshold from two-or-more to one-or-more functional-bug medium findings across agents/review.md and skills/pr-review/SKILL.md, and updated outdated verdict language in skills/code-review/SKILL.md to match. All other review findings ignored per human instruction.

Fixed (3):

1. lower request-changes threshold to one functional-bug medium finding (agents/review.md): Changed 'two or more medium-severity findings where at least one identifies a functional bug' to 'one or more medium-severity findings identifying a functional bug' so a single functional-bug medium finding triggers request-changes
2. lower request-changes threshold to one functional-bug medium finding (skills/pr-review/SKILL.md): Changed 'Two or more medium findings, at least one of which identifies a functional bug' to 'One or more medium findings identifying a functional bug' and updated the comment-only bullet to 'One or more medium findings that are all stylistic/advisory/process-related'
3. update outdated language in skills/code-review (skills/code-review/SKILL.md): Replaced outdated verdict criteria ('Multiple medium findings which could affect the intended outcome of the PR' and 'One medium finding (but no critical/high)') with new criteria matching the lowered threshold: one or more medium findings identifying a functional bug triggers request-changes; stylistic/advisory/process-related medium findings get comment-only

Tests: passed

_{Updated by fullsend fix agent}

[rh-hemartin]

@ben-alkov you are assigned, is this ready to be merged?

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 9:28 PM UTC · Completed 9:46 PM UTC
Commit: 9cd9f43 · View workflow run →

[fullsend-ai-retro]

Retro: PR #49 — Tighten review verdict criteria for medium findings

Timeline

1. Jul 7 — Retro agent files issue #45 from PR docs: add CLAUDE.md, COMMITS.md, CONTRIBUTING.md, LICENSE, and repo workflows #32 retro, identifying that the review agent chose COMMENTED despite its summary saying findings "should be addressed before merge." The issue proposes a "two or more medium findings" threshold for request-changes.
2. Jul 7 — Triage agent classifies as Medium/Bug, restates the proposed fix verbatim. Code agent creates PR #49, faithfully implementing the issue's "two or more" threshold.
3. Jul 9 — Human triggers /fs-review. Review agent correctly identifies a logic-error: the proposed criteria leave a single medium functional-bug finding unhandled (falls through both branches). Verdict: COMMENTED.
4. Jul 9 — Human triggers /fs-fix with precise instructions to lower the threshold. Fix agent pushes corrected commit (run 29053069905).
5. Jul 20 — Human approves. Jul 30 — Merged.

Assessment

What went well: The review agent caught a genuine logic error that the code agent introduced — the "two or more" threshold created a gap where exactly one medium functional-bug finding satisfied neither request-changes nor comment-only. The fix agent correctly addressed all directed findings in a single iteration. The pipeline worked end-to-end.

Root cause of the rework: The logical gap originated in the retro agent's issue #45 proposal, not in the code agent's implementation. The retro agent proposed decision-rule criteria with two branches that didn't cover all cases. The triage agent restated it without catching the gap. The code agent faithfully implemented the incomplete specification per AGENTS.md rule #3 (surgical changes). The review agent was the safety net that caught it.

Review verdict irony: The review agent found a medium functional-bug finding but issued COMMENTED — demonstrating the exact behavior that issue #45 was filed to fix. Under the pre-PR rules, this was technically correct (the old criteria didn't distinguish functional from stylistic medium findings). Now that PR #49 is merged, this class of verdict inconsistency should be resolved going forward.

Time-to-merge (23 days): Substantive work completed by day 2 (Jul 9). Human approval came Jul 20 (11-day gap). Then sat 10 more days until merged Jul 30. Existing issues #395 and #304 already cover the automation angle for reducing post-approval merge latency.

Existing issue evidence: This retro provides additional evidence for #383 (code agent should trace rule changes against cited incidents) — if the code agent had tested the proposed verdict criteria against the PR #32 scenario that motivated issue #45, it might have noticed the gap for one-medium-functional-bug cases.

Proposals filed

• Retro agent should validate edge-case completeness when proposing decision-rule changes (in fullsend-ai/agents)