csdlc-intern

CSDLC Intern — Sub-agent Handoff

Hand off this story to a sub-agent: $ARGUMENTS

This is Step 3 of the CSDLC pipeline. A refined, scoped story gets handed to an "intern" (sub-agent) for execution. The intern works in an isolated worktree, implements the story, and reports back.

Procedure

1. Read the story. $ARGUMENTS is either:

   • A story ID (e.g., FND-E12-S3) → search Foundry for the story section in the epic doc
   • A Foundry path → get_page to read it
   • A local file path → read the file
   • Nothing → ask the user which story to dispatch

2. Verify the story is ready. Check that it has:

   • Clear scope
   • Acceptance criteria (specific, testable)
   • Target files/areas identified
   • Boundaries defined (what NOT to touch)

   If any are missing, flag them and ask: "This story is missing [X]. Want me to fill it in, or should we /refine it first?"

3. Craft the agent prompt. Use the csdlc-craft-agent-prompt skill to assemble the prompt. The prompt includes:

   • Role and context (story ID, repo, base branch)
   • Task context (story content + relevant design doc sections)
   • Acceptance criteria (copied directly)
   • Explicit boundaries (do-not-touch rules)
   • Verification steps (build, test commands)
   • Files to read first

4. Confirm with the user. Present a summary:

   Ready to dispatch:
   Story: {ID} — {title}
   Repo: {path}
   Branch: {branch-name}
   Scope: {one-line summary}
   
   Send the intern? (or review the full prompt first)
   

5. Spawn the intern. Use the Agent tool with isolation: "worktree":

   • Pass the crafted prompt
   • The intern works in an isolated worktree — no contamination of the main branch
   • The intern implements, tests, and opens a PR (if configured)

6. Report results. When the intern completes:

   • Summarize what was done (files changed, tests passed/failed, PR link)
   • Write a result annotation to Foundry on the story doc:

     Story {ID} executed.
     Status: complete / failed
     Branch: {branch}
     PR: {url}
     Tests: {summary}
     

   • Suggest next steps: /review to verify, or dispatch the next story

Execution modes (from PROCESS.md)

Mode	When	How
Direct sub-agent	Simple, well-scoped stories	Spawn intern with full prompt. Intern implements and reports back.
Managed agent	Complex stories, audit trail needed	Spawn a lightweight manager (Sonnet) that delegates to Claude Code for implementation.

Default to direct sub-agent. Use managed agent when the story is complex enough to warrant orchestration.

What NOT to do

• Don't execute without user confirmation (unless /shipit is running the batch).
• Don't modify the story doc — that's the AI Lead's job, not the dispatch skill's.
• Don't skip the prompt review step. A bad prompt produces rework.