Skip to content

Evidence for #2940: medium-severity ADR 0062 finding merged unresolved on PR #2850 #3878

Description

@fullsend-ai-retro

What happened

On PR #2850, the review agent flagged a medium-severity factual-accuracy finding across 4 consecutive reviews (SHAs 307cb63 through 0f47a19): the Exploration section references 'ADR 0062' for auto-merge, but the accepted ADR 0062 on disk covers dispatch version skew (unrelated). The referenced PR #2791 is an open draft proposing to supersede ADR 0062. The agent provided clear remediation (assign a new ADR number or clarify the pending status). The author explicitly dismissed only the low-severity broken-reference inline finding ('accepting as is') but never acknowledged the medium finding. Two human reviewers (Roming22, ralphbean) approved based on content quality without engaging with the agent's ADR finding. The PR was merged by the author on Jul 9.

What could go better

The COMMENTED verdict made the medium finding invisible in GitHub's merge readiness UI. A CHANGES_REQUESTED verdict for medium-severity factual-accuracy findings would require explicit dismissal before merge, creating an audit trail for the decision to ship with the known inconsistency. This reinforces #2940. Confidence: medium — the finding is factually correct, but reasonable people could decide the inline PR link to #2791 provides sufficient context for readers. The key insight is about the process gap (no forced acknowledgment), not necessarily whether the finding should have blocked merge.

Proposed change

This is additional evidence for #2940. When implementing that issue's verdict escalation logic, this PR's factual-accuracy category should qualify for CHANGES_REQUESTED — the finding was correct, actionable, had a clear remediation, and was silently ignored. The specific case (ADR number collision where on-disk ADR content doesn't match what the roadmap claims) is a high-confidence factual error that should not be dismissible by omission.

Validation criteria

On the next 5 PRs where the review agent identifies a medium-severity factual-accuracy finding, track whether the finding receives an explicit response (fixed, dismissed with rationale, or overridden) before merge. The baseline from PR #2850 is 0% explicit response rate for medium findings. Target: 80%+ of medium-severity factual-accuracy findings receive explicit acknowledgment before merge.


Generated by retro agent from #2850

Metadata

Metadata

Assignees

No one assigned

    Labels

    agent/reviewReview agentduplicateThis issue or pull request already existsready-for-triageRetro-filed issue awaiting triage agent

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    Status
    Done

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions