Review agent should suppress governance CHANGES_REQUESTED when human approvals already satisfy the gate

[fullsend-ai-retro]

What happened

On PR #5641, the review bot's only HIGH finding across 14 review rounds was [protected-path] — a governance gate flagging that files under skills/ were modified and require human review. Two humans approved the PR (ascerra on Aug 3, waynesun09 on Aug 5), both explicitly reviewing the protected-path files. Despite this, the review bot continued filing CHANGES_REQUESTED — 11 times total, including 3 times after both human approvals existed. The bot's verdict also oscillated: DISMISSED at 06:57, CHANGES_REQUESTED at 12:09, DISMISSED at 17:40 on Aug 5, all with the same protected-path finding. The author's final triage comment explicitly noted: 'Process gate only, human reviews already APPROVED. No code change.'

What could go better

The protected-path finding's purpose is to ensure human review of sensitive paths. Once qualifying human approvals exist on those exact files, the governance requirement is satisfied — filing CHANGES_REQUESTED at that point adds friction without value. This is distinct from #5369 (which covers governance findings on bot PRs, not human-approved PRs) and from #1922 (which covers stale findings from unchanged code, not governance findings whose implied action is already completed). The bot's protected-path finding was technically correct every time — the files were modified — but its recommendation ('needs human review') was already fulfilled. Confidence is high that this caused unnecessary friction: the author spent time triaging a finding that was already resolved by the review state. Uncertainty: it's possible the review agent intentionally re-flags protected paths as an audit trail, in which case the fix is to change the verdict type (COMMENT not CHANGES_REQUESTED) rather than suppress the finding entirely.

Proposed change

In the review agent's verdict logic (likely in agents/review.md or policies/review.yaml in fullsend-ai/agents), add a check: when the only HIGH+ findings are governance-type gates (like protected-path) whose requirement is 'human review of protected files,' query the PR's review state for existing human APPROVED reviews. If qualifying approvals exist, downgrade the finding's impact on the verdict from CHANGES_REQUESTED to COMMENT/informational. The finding should still be surfaced (it's useful as an audit trail) but should not block the merge indicator. This could be implemented as a policy rule: if finding.category == 'governance' and pr.has_human_approvals(on_protected_paths=True): verdict_impact = 'comment'.

Validation criteria

On the next 5 PRs in fullsend-ai/fullsend that modify protected paths (e.g., skills/, harness/) and already have at least one human APPROVED review, the review bot should not file CHANGES_REQUESTED solely based on protected-path. The finding should still appear in the review summary as informational. Measure: zero CHANGES_REQUESTED reviews where the only HIGH finding is protected-path and human approvals exist.

---

Generated by retro agent from fullsend-ai/fullsend#5641

[fullsend-ai-prioritize]

RICE Priority Score: 1.8

Score breakdown

Dimension	Score	Reasoning
Reach	1.5	This affects all developers who modify protected paths (skills/, harness/) in fullsend-ai/fullsend and already have human approvals. The issue demonstrates repeated friction on a real PR (#5641) and the pattern will recur for any contributor touching protected directories. Multiple active contributors across the org are likely affected, but it's not platform-wide — only PRs touching specific protected paths trigger this.
Impact	1.5	The bot filed CHANGES_REQUESTED 11 times on a single PR, including 3 times after two human approvals existed, causing the author to spend time triaging findings that were already resolved. This blocks merge indicators without value and creates unnecessary noise. The workaround (manual triage/dismissal) exists but is tedious and recurring, making this a significant pain point for affected workflows.
Confidence	0.8	High confidence. The issue is exceptionally well-documented with specific PR references, exact timestamps showing verdict oscillation, direct quotes from the author's triage comment, and clear validation criteria. The scope is well-defined: modify verdict logic for governance-type findings when human approvals satisfy the gate. The only uncertainty is whether the fix should suppress vs. downgrade the finding, which the issue itself acknowledges.
Effort	1	The change is well-scoped: add a conditional check in the review agent's verdict logic (likely agents/review.md or policies/review.yaml) to query PR review state for existing human approvals when governance-type findings are the only HIGH+ findings, then downgrade verdict impact from CHANGES_REQUESTED to COMMENT. Requires understanding the review agent's verdict pipeline and testing across representative PRs, but the logic is straightforward and the issue provides a clear implementation sketch.

Formula: (1.5 x 1.5 x 0.8) / 1 = 1.8