sc-review-pr

Run a context-aware PR review that gathers ticket context, PR metadata, and CLAUDE.md guidelines before running parallel review agents.

Constraints:

• Do NOT use run_in_background: true on any Task call. All agents run in foreground.
• All tools are functional. Do not make exploratory calls.
• Only call tools required to complete the task.
• HIGH SIGNAL ONLY — flag issues where code will fail, not style concerns.
• CONTEXT-FIRST — understand author intent before reviewing.

Create a TodoWrite checklist before starting.

Phase 1: Context Gathering (haiku, parallel)

Launch three haiku agents in parallel to gather all context:

# Agent 1: PR Metadata
Task(subagent_type: "general-purpose", model: "haiku",
     prompt: "Get PR metadata using gh CLI:
              gh pr view $ARGUMENTS --json number,title,body,commits,files,headRefName,baseRefName,labels,author

              Extract and return:
              - PR number and title
              - Full PR body/description
              - List of commit messages
              - List of changed files
              - Branch names (head and base)
              - Author

              Return as structured data.")

# Agent 2: CLAUDE.md Finder
Task(subagent_type: "general-purpose", model: "haiku",
     prompt: "Find all CLAUDE.md files in the repository being reviewed.

              Step 1: Get the repository root
              - Run: gh pr view $ARGUMENTS --json headRepository --jq '.headRepository.name'
              - Get local path by finding it: fd -t d -d 3 '<repo-name>' ~ 2>/dev/null | head -1
              - Or use pwd if already in the repo directory

              Step 2: Search for ALL CLAUDE.md files in that repo
              - Run: fd -t f 'CLAUDE.md' <repo-path> 2>/dev/null
              - This finds CLAUDE.md at any depth (root, .claude/, subdirectories)

              Step 3: Read each file found
              - Use the Read tool on each CLAUDE.md path

              Return:
              - Full paths to all CLAUDE.md files found
              - Brief summary of key guidelines from each
              - 'No CLAUDE.md files found' if none exist")

# Agent 3: Ticket Extractor
Task(subagent_type: "general-purpose", model: "haiku",
     prompt: "Extract ticket references from PR metadata:

              1. Get PR title and body: gh pr view $ARGUMENTS --json title,body
              2. Parse for ticket patterns:
                 - JIRA: PROJECT-123, PROJ-1234
                 - GitHub Issues: #123, owner/repo#123
                 - Linear: ABC-123
              3. If GitHub issues found, fetch with: gh issue view [number] --json title,body

              Return:
              - List of ticket references found
              - Ticket summaries if fetchable
              - 'No ticket references found' if none")

Wait for all Phase 1 agents to complete before proceeding.