caspar-ship

name	caspar-ship
description	👻 | Autonomous end-to-end delivery — brain dump → scope → TDD → sweep → rebase → PR, zero confirmation gates. Use when you want a reviewable PR produced hands-off from an informal description. Do NOT use when you want approval checkpoints (use the staged scope/plan/execute flow), or for non-code/non-git work.
user-invocable	true

ship

Take a brain dump and autonomously produce a reviewable PR — scope, implement with TDD, sweep, rebase, and open the PR with no user approval gates.

Inputs

• Brain dump from $ARGUMENTS. If empty, ask once for it, then proceed autonomously.
• Parse from it: INTENT_TYPE (feat = new behavior, fix = broken behavior), TARGET_BRANCH (default origin/main), SCOPE_SUMMARY (1–2 sentences), RELEVANT_FILES, CONTEXT.
• Current branch + tree state (read just-in-time via git status/git rev-parse, never assume).

Working Set

• The code area named in RELEVANT_FILES/SCOPE_SUMMARY (locate via @finder, understand via @analyst).
• Project test/lint commands (detect from repo).
• BRANCH_NAME for artifact paths.

Outputs + DONE

• docs/tasks/{BRANCH_NAME}/concepts/scope.md — lightweight (~20 lines): Objective · Type · In Scope · Out of Scope · Target Branch · Key Files.
• Conventional commits (one per task minimum), {INTENT_TYPE}({scope}): {desc}.
• A pushed branch + a PR (gh pr create) with title {INTENT_TYPE}({scope}): {SCOPE_SUMMARY} (<70 chars) and body sections: Summary · Changes · Test Plan · "Shipped autonomously via caspar-ship".

DONE when: all tasks completed via TDD, sweep done and committed, branch rebased onto TARGET_BRANCH with lint + full test suite actually run and passing, branch pushed, and the PR URL output as the final deliverable.

Method / guardrails

Fully autonomous — no approval gates; parse intent, build it, ship the PR. Run in order:

1. Worktree — if on main/master, EnterWorktree to isolate; else use the current branch. Capture BRANCH_NAME.
2. Quick scope — fan out @finder + @analyst (read-only), write the ~20-line scope.md. This is a light scope, not full caspar-scope.
3. Tasks — TaskCreate 3–8 tasks proportional to scope, dependency-ordered (foundational first), each with imperative subject + acceptance criteria. Ephemeral — no file artifact.
4. Execute (TDD) — per task sequentially: TaskUpdate→in_progress; load Skill(caspar-tdd); RED (failing test first) → GREEN (minimal impl) → REFACTOR; commit; TaskUpdate→completed. Skip TDD only for config/doc-only tasks. If new work surfaces, create a new task — do not scope-creep the current one.
5. Sweep (inline, no subagents) — run every check, then commit chore({scope}): sweep cleanup if it produced changes:
   • Diff sanity: no unintended/whitespace-only/merge-artifact edits; no out-of-scope staged files; no secrets/keys/credentials.
   • Logging: remove temp/debug logs; keep intentional error/warning/state logs; appropriate levels.
   • Hygiene: remove commented-out code; resolve/document TODO/FIXME/HACK added this session; remove hardcoded test values.
   • Dead code: orphaned imports, unused vars/functions, debug artifacts.
   • Lint (strict): fix all violations — no skipping, no eslint-disable/--no-verify; refactor structural issues, never suppress.
   • Test: run affected + full suite, fix failures. Do not write new tests here (done in step 4).
6. Rebase — git fetch origin; YOU MUST create a backup ref first: git branch backup/ship-$(date +%Y%m%d-%H%M%S) (the only rollback; never rebase without it). git rebase {TARGET_BRANCH}, auto-resolving conflicts favoring the target's conventions, tracking each decision for the PR. Verify after: lint (fix) + full test suite (fix) + confirm commit count and no unexpected changes.
7. PR — git push -u origin {BRANCH_NAME}, then gh pr create; output the PR URL. Ground the description in the actual change, not memory: What = behavioral summary from the diff (not a diff recitation); Why = from the brain dump / commits / linked issue (placeholder if none — never fabricated); Test Plan = what the diff's tests actually cover. Scale to size — trivial PRs get Summary + Why; larger/multi-area PRs add Changes, Test Plan, and a Feedback-Requested line directing the reviewer. Before opening, verify each claim maps to a real hunk and no secrets/PII are quoted into the body. (See Skill(caspar-create_pr) for the full grounding contract.)

Handoff

Terminal skill — output is the PR URL. End with a one-line Next Steps pointer (e.g. review the PR / caspar-code_review).

Escalate-If

• ARGUMENTS empty → ask once for the brain dump, then continue autonomously.
• Rebase conflicts are genuine semantic divergence (not resolvable by favoring the target) → stop, report, leave the rebase in progress with the backup-ref restore command.
• Tests still fail after rebase/sweep fixes → report the failure + restore command; do not force-push or open the PR.