in-the-loop-labs/analyze

Analyze Changes

Perform a three-level code review analysis and return curated suggestions.

Tools

This skill includes a scripts/git-diff-lines script that annotates git diff output with explicit OLD and NEW line numbers. Each subagent should invoke it by name (git-diff-lines) — the orchestrating agent must ensure the script's directory is on PATH (e.g., via PATH="<skill-dir>/scripts:$PATH").

1. Gather context

Determine what's being reviewed:

• Local changes: git diff --name-only HEAD for changed files.
• PR changes: Determine the merge base (git merge-base main HEAD), then diff against it. Get PR title and description if available.
• If the user specifies a different diff range, use that.

Collect:

• The list of changed files
• PR metadata (title, description, author) if applicable
• Whether this is local or PR mode
• Any custom review instructions the user provided

2. Get analysis prompts

Obtain the prompt instructions for each analysis level. Use the tier argument (default: balanced).

If get_analysis_prompt is in your available tools (i.e., the pair-review MCP server is connected):

• Call get_analysis_prompt for each level you will run: level1, level2, and (unless skipLevel3 is true) level3
• Call get_analysis_prompt with promptType: "orchestration" for the orchestration step
• Pass the user's tier argument to each call
• Pass any custom review instructions (from the customInstructions argument, or gathered from user context in Step 1) as the customInstructions parameter — this injects them into the rendered prompt
• These return the full prompt text to use as Task agent instructions

Otherwise (standalone mode — no MCP connection):

• Read the static reference files from this skill's references/ directory:
  • references/level1-{tier}.md
  • references/level2-{tier}.md
  • references/level3-{tier}.md (unless skipLevel3)
  • references/orchestration-{tier}.md

3. Run analysis levels in parallel

Launch two or three Task agents simultaneously (subagent_type: "general-purpose"), depending on skipLevel3. Each task must:

1. Use the prompt obtained in Step 2 as its core instructions
2. Use the git-diff-lines script to get the annotated diff
3. Analyze the changes per its framework
4. Return valid JSON (no markdown wrapping) with the schema defined in the prompt

Pass each task the following context in its prompt:

• The analysis prompt from Step 2
• Review type (local or PR)
• The list of changed files
• PR title and description (if PR mode)
• Instructions to invoke git-diff-lines (ensure the script's directory is on PATH)

Level 1 — Analyze changes in isolation (diff only) Level 2 — Analyze changes in file context (full files) Level 3 — Analyze changes in codebase context (architecture, dependencies) — skip if skipLevel3 is true

4. Orchestrate results

Launch one more Task agent (subagent_type: "general-purpose") that:

1. Uses the orchestration prompt from Step 2 as its core instructions
2. Receives the JSON output from all completed levels (pass empty [] for Level 3 if skipped)
3. Merges, deduplicates, and curates suggestions
4. Returns final curated JSON

5. Push results to server

Push the orchestrated JSON to the pair-review web UI so suggestions appear inline. This step does not require MCP — it uses a direct HTTP POST with a fallback to http://localhost:7247 when the MCP get_server_info tool is unavailable.

1. Determine server URL:

   • If the get_server_info MCP tool is available, call it and use the url field
   • Otherwise, try reading the port from config: cat ~/.pair-review/config.json 2>/dev/null | jq -r '.port // empty'
   • If neither works, default to http://localhost:7247

2. Build the POST body from the orchestrated output:

   • For local mode: set path (absolute working directory from pwd) and headSha (from git rev-parse HEAD)
   • For PR mode: set repo (owner/repo) and prNumber
   • Set provider and model to describe what ran the analysis (e.g., the AI provider and model used)
   • Include summary, suggestions, and fileLevelSuggestions from the orchestrated JSON

3. POST via curl to ${SERVER_URL}/api/analyses/results:

   curl -s --connect-timeout 3 --max-time 10 \
     -X POST "${SERVER_URL}/api/analyses/results" \
     -H "Content-Type: application/json" \
     -d @- <<'PAYLOAD'
   { ... }
   PAYLOAD
   

   A successful import returns HTTP 201 with { runId, reviewId, totalSuggestions, status: "completed" }.

4. Graceful degradation: If the request fails (server not running, timeout, etc.), log a short warning and continue to the Report step. The push is best-effort.

5. Note in the report whether results were successfully pushed to the pair-review UI.

6. Report

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

• File and line reference
• Type (bug, improvement, security, etc.)
• Title and description
• Suggested fix (if applicable)
• Confidence level