Skip to content

Review agent should downgrade protected-path for platform bot managed-file updates #3164

Description

@fullsend-ai-retro

What happened

On PR #29, the review agent posted CHANGES_REQUESTED 3 times driven by a high-severity protected-path finding on .github/workflows/fullsend.yaml. The file contains an explicit managed-file header ('This file is managed by fullsend. Do not edit it directly.') and the PR author is fullsend-ai-fullsend[bot]. The review agent's own info-level authorization finding correctly stated 'automated bot performing template synchronization' and 'change is mechanical', but the protected-path finding ignored that signal. Human reviewer approved on Jul 6 without addressing any findings, confirming them as false positives. The CHANGES_REQUESTED verdict contributed to a 27-day merge delay for a 5-line change.

What could go better

The review agent should consume its own authorization-inference signal when calibrating protected-path severity. When the platform bot modifies a file that declares itself as managed by fullsend, the protected-path finding should be informational, not blocking. This is distinct from #2588 (dep bots like Renovate/Dependabot) and #3061 (Dockerfile digest rotations) — the platform bot updating its own managed infrastructure is the strongest possible authorization case. Confidence is high: the review agent's own findings demonstrate the contradiction, and the human override confirms the blocking verdict was unnecessary.

Proposed change

Extend the review agent's protected-path severity calibration to recognize when fullsend-ai-fullsend[bot] modifies files containing managed-file headers. When both conditions are met, downgrade the protected-path finding from high to info severity, changing the verdict from CHANGES_REQUESTED to COMMENT. The detection signal already exists in the authorization-inference finding; the severity assignment just needs to consume it.

Validation criteria

Next 5 PRs from fullsend-ai-fullsend[bot] modifying files with managed-file headers should receive COMMENT verdicts (not CHANGES_REQUESTED) when the only blocking finding is protected-path. Non-managed files modified by the same bot should retain standard severity.


Generated by retro agent from fullsend-ai/experiments#29

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    Status
    Todo

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions