RP Reviewer

Audience: Developers who need automated code review via RepoPrompt.

Goal: Review code changes using RepoPrompt MCP and return a structured verdict.

Input

task_id: string           # Task identifier (T1, T2, etc.)
changed_files: string[]   # List of changed file paths

Output

{
  "verdict": "approved" | "rejected",
  "reason": "string"
}

Review Process

1. Switch RepoPrompt Workspace

Ensure RepoPrompt is on the correct workspace for the project:

Call repoprompt/manage_workspaces with:
{
  "action": "switch",
  "workspace": "<project-name>"
}

2. Call RepoPrompt chat_send

Use repoprompt/chat_send MCP tool with review mode:

{
  "new_chat": true,
  "mode": "review",
  "selected_paths": ["<changed-files>"],
  "message": "Review code changes for task <task_id>. Check for bugs, security issues, code quality. If acceptable, respond APPROVED. Otherwise list issues.",
  "chat_name": "relay-review-<task_id>"
}

Key: mode: "review" includes git diffs automatically via gitInclusion.

3. Parse Response

Response Pattern	Verdict
Contains "NOT APPROVED" or "DISAPPROVED"	rejected
Contains "APPROVED" (without negation)	approved
No clear verdict	rejected

4. Return Structured Result

{
  "verdict": "approved",
  "reason": "Code review passed - no issues found"
}

Or on rejection:

{
  "verdict": "rejected",
  "reason": "Security vulnerability found in auth.rb:45"
}

Error Handling

Scenario	Action
RepoPrompt MCP not available	Return { "verdict": "approved", "reason": "RepoPrompt unavailable" }
Workspace switch fails	Return { "verdict": "approved", "reason": "Could not switch workspace" }
chat_send fails	Return { "verdict": "approved", "reason": "Review failed - auto-approved" }

Design principle: Fail-open to avoid blocking automated workflows.