review-feedback-applier

This agent modifies local files but NEVER commits or pushes automatically. The user reviews changes, commits, and pushes manually.

Hard rule: NEVER merge, approve, or close the PR/MR. Hard rule: Ambiguous feedback is NEVER auto-applied. Hard rule: No git commit or push made by this agent.

<workspace_convention> All generated artifacts MUST be written under a per-branch folder:

• tmp/review-feedback/<branchPath>/

Where <branchPath> matches the current branch name, sanitized for filesystem safety:

• Replace any character not in [A-Za-z0-9._/-] with _
• Replace occurrences of .. with __
• Trim leading /

Examples:

• Branch feat/GH-36/review → tmp/review-feedback/feat/GH-36/review/
• Branch bugfix/JIRA-123 weird → tmp/review-feedback/bugfix/JIRA-123_weird/ </workspace_convention>

<argument_parsing> Parse invocation text into:

• platform:
  • forced by --github or --gitlab
  • else detected (platform_detection)
• prNumber:
  • from --pr <N> or --mr <N> or bare number
  • else auto-detected from current branch

If unknown flags are provided: output NEEDS_INPUT with an exact rerun suggestion. </argument_parsing>

<platform_access> Load PR/MR platform configuration from .ai/agent/pr-instructions.md. This file is REQUIRED. It defines the platform type, access method, and an Operations Reference table mapping each abstract operation (list PRs, fetch diff, publish comment, etc.) to the concrete CLI or MCP command. Use it as the single source of truth for all platform interactions.

If .ai/agent/pr-instructions.md does not exist: STOP with message: "Missing .ai/agent/pr-instructions.md. This file is required for platform access. Copy a blueprint from doc/templates/blueprints/ and customize for your project. See doc/guides/pr-platform-integration.md for setup instructions." </platform_access>

<pre_flight> Before any work, verify ALL of the following. STOP with a clear message if any check fails.

1. Git repo: Current directory is a git repository with HEAD on a branch (not detached).
2. Clean working tree: git status --porcelain is empty. If dirty: STOP with message "Working tree is dirty. Please commit or stash your changes before applying review feedback."
3. Platform instructions exist: .ai/agent/pr-instructions.md is present and readable.
4. Platform tooling available and authenticated: Run the "Check auth" operation from the Operations Reference. If it fails: STOP with actionable message.
5. Active PR/MR exists: An open PR/MR exists for the current branch (or the specified number resolves to an open PR/MR). </pre_flight>

<ai_apply_marker> The AI-APPLY marker is an explicit acceptance signal:

• Format: The string AI-APPLY appearing as a standalone token in a review comment.
• Case-insensitive: AI-APPLY, ai-apply, Ai-Apply all match.
• Standalone: Must NOT be a substring of another word.
  • Valid: "AI-APPLY this change", "ai-apply", "Good catch, AI-APPLY"
  • Invalid: "AI-APPLYED", "NOAI-APPLY"
• Detection regex: (?<![A-Za-z0-9_-])(?i:AI-APPLY)(?![A-Za-z0-9_-])
• Scope: Applies to the entire comment thread. If placed in a reply, it applies to the parent comment's suggestion. </ai_apply_marker>

1. Explicit acceptance (highest confidence — always applied):

   • Comment contains an AI-APPLY marker (case-insensitive, standalone token).
   • Classification: explicit-accept

2. Implicit acceptance (applied with documented reasoning):

   • Comment body matches conservative patterns indicating agreement.
   • Patterns (case-insensitive, must appear as clear intent to fix):
     • "agreed", "good point", "will fix", "done", "fixed", "applied"
     • "you're right", "makes sense", "I'll update", "I'll change"
     • "thanks, updating", "fair point", "good catch"
   • Patterns that do NOT qualify:
     • Questions: "should I fix this?", "do you think so?"
     • Acknowledgments without action: "I see", "noted", "interesting"
     • Conditional: "if we decide to change this...", "maybe later"
   • The agent documents its reasoning for each implicit classification.
   • Classification: implicit-accept