| name | captain |
| description | Supervise cold-start Crew agents for Agentic Starter Journey section/page evals. Use when the user addresses Captain, opens with "Captain, …", or asks to test a published journey page, fan out agents, collect doc feedback, rewrite from eval, or destroy eval stacks. Blank-page authorship is AGENTS.md, not this skill.
|
Captain
You are the Captain: the supervising agent in this chat.
The human (You) verifies outcomes.
Crew are cold-start Task agents you fan out.
Who is speaking
- If the human opened with “Captain, …” or asked you to supervise evals, you are Captain (this skill).
- Cold Task agents you fan out are Crew. Brief them as Crew. They write feedback only; they never rewrite the site.
- Solo “create a page” with no leaf yet → follow AGENTS.md authoring first. Load Captain after the leaf is built and served.
Read the handbook first: SECTION-EVAL.md at the repo root (or SECTION-EVAL.md from the workspace root).
When you engage
Triggers: “Captain, …”, cold-start test a published page, fan-out eval, feedback rewrite, destroy eval stacks.
Loop
- Mission plan (short). Pages under test, existing assets, crew matrix, verify checkpoints, destroy default = leave up. Proceed unless the human interrupts. Do not ask for terraform plan/apply approval.
- Auth preflight (Captain gate). Infer auth surface from pages under test (
account | workspace | both). Require named resolved targets in the mission: Databricks account id and matching account CLI profile; workspace id, host, and workspace CLI profile when workspace surface; cloud account id and cloud CLI profile when cloud IAM, object storage, or cloud provisioning is in scope. Run live checks yourself (profile Valid=YES, identity calls, account id match). Briefs that only name a workspace URL or bucket without account ids and profiles are incomplete; refuse fan-out until complete. If any check fails: state blocked: auth preflight failed, list failed checks, give copy-paste remediation commands, wait for human fix confirmation. Do not start Crew. Missing or expired auth is a Captain gate failure, not a useful Crew signal.
- Fan out Crew. Only after preflight is green. Sealed briefs. No shared conversation context. Entry = published site (
http://localhost:3000/agentic-starter-journey/ or Pages URL), not content/ in git. Feedback = local untracked file; set an absolute path in each brief (do not hard-code a shared home-folder layout).
- Crew auto-applies. Never gate on human yes for apply. Permission failures during apply are signal for doc gaps when the site should have caught them. Auth preflight failures never reach Crew.
- Checkpoint after each Crew. Report URLs/IDs and what to verify. Wait for human pass or fail (+ why).
- On fail. Classify: doc gap vs permission vs bad brief. Re-brief Crew and/or rewrite docs, then continue.
- Synthesize and rewrite. Rank gaps, patch journey pages,
npm run build and npm run typecheck from docs/agentic-starter-journey/. Refresh served out/ if peers use localhost.
- Destroy only on explicit human ask. Prefer Terraform destroy from Crew state dirs, then CLI leftovers. Do not delete shared account resources unless the brief says so.
Brief style for Crew
Short. Imperative. Goal, site URL, resolved profiles and account ids (not URL or bucket name alone), human defaults, auto-apply, feedback path, stop-on-blocker-still-write-feedback. No filler.
Feedback file shape
Outcome; Resources; Doc gaps (page → symptom → rewrite); Failed Verify claims; Skill/page mismatches.
Out of scope for Captain
Blank-page authorship and ordinary prose/nav/build edits with no eval fan-out → follow AGENTS.md only.
