in-the-loop-labs/analyze

Analyze Changes (Server-Side via MCP)

Perform a three-level code review analysis using pair-review's built-in analysis engine. Unlike the code-critic:analyze skill (which spawns analysis agents directly within the coding agent's context), this skill delegates all analysis work to the pair-review server via MCP tool calls. Use this skill when the pair-review MCP server is connected; use code-critic:analyze when it is not.

Prerequisites: The pair-review MCP server must be connected with the start_analysis, get_ai_analysis_runs, and get_ai_suggestions tools available.

1. Gather context

Determine what is being reviewed:

• Local changes: Run git rev-parse HEAD to get the HEAD SHA, and use the current working directory as the path.
• PR changes: Determine the repository (owner/repo) and PR number. If the user provides a PR URL, extract these from it. If available, run gh pr view --json number,headRepository or similar.
• If the user specifies a particular scope, use that.

Collect:

• Whether this is local or PR mode
• For local: the absolute path and HEAD SHA
• For PR: the owner/repo string and PR number

2. Start analysis

Call the start_analysis MCP tool with the appropriate parameters:

For local mode:

{
  "path": "/absolute/path/to/repo",
  "headSha": "abc123...",
  "tier": "balanced",
  "skipLevel3": false,
  "customInstructions": "optional instructions"
}

For PR mode:

{
  "repo": "owner/repo",
  "prNumber": 42,
  "tier": "balanced",
  "skipLevel3": false,
  "customInstructions": "optional instructions"
}

Pass the tier, skipLevel3, and customInstructions arguments from the skill invocation.

The tool returns immediately with { analysisId, runId, reviewId, status: "started" }.

3. Poll for completion

Analysis typically takes 1-5 minutes depending on the size of the changes and the tier.

Poll using get_ai_analysis_runs with limit: 1 and the review coordinates from step 1. Wait approximately 10 seconds between polls.

For local mode:

{
  "path": "/absolute/path/to/repo",
  "headSha": "abc123...",
  "limit": 1
}

For PR mode:

{
  "repo": "owner/repo",
  "prNumber": 42,
  "limit": 1
}

Verify that runs[0].id matches the runId returned from step 2 to ensure you are polling the correct analysis (another tab or user could have triggered a concurrent run).

Check the status field in runs[0]:

• "running": Analysis is still in progress — keep polling
• "completed": Analysis finished — proceed to fetch results using the same runId
• "failed": Analysis failed — report the failure to the user

Keep polling until status is completed or failed. Report progress to the user between polls.

Error handling

• If start_analysis returns an error, report it to the user with the error message. Common causes: the PR has not been loaded in pair-review yet, the worktree is missing, or invalid parameters.
• If polling shows status: "failed", report the failure to the user. The error field in the run object contains the failure reason.
• If get_ai_suggestions returns an empty list after a completed analysis, this means no issues were found -- report a clean result.

4. Fetch results

Once the analysis is complete, call get_ai_suggestions to retrieve the curated suggestions.

Pass the runId from step 2's start_analysis response (same value used for polling):

{
  "runId": "the-runId-from-step-2"
}

If runId is unavailable, fall back to resolving by review coordinates:

For local mode:

{
  "path": "/absolute/path/to/repo",
  "headSha": "abc123..."
}

For PR mode:

{
  "repo": "owner/repo",
  "prNumber": 42
}

This returns the final orchestrated suggestions from the latest analysis run.

5. Report

Present the curated suggestions to the user, organized by file. For each suggestion show:

• File and line reference
• Type (bug, improvement, security, performance, praise, etc.)
• Title and description
• Confidence level

Group suggestions by file for easy navigation. Highlight critical issues (bugs, security) prominently.

If no suggestions were found, report that the analysis completed cleanly.