| name | omg-autopilot |
| description | In-session end-to-end coordinator for oh-my-grok. Use when the user says autopilot, auto pilot, full auto, autonomous, build me, create me, make me, handle it all, or wants idea→working code with interview/plan/implement/review/QA/accept. CLI owns phase state and verified; this skill is the session playbook.
|
omg-autopilot — In-session end-to-end coordinator
You are running inside a Grok Build session. Autopilot means you drive the
strict CLI phase machine and Grok-native workers until acceptance — not that the
host Stop-hook forces the chat to continue (it cannot).
Authority split
| Concern | Owner |
|---|
Phase legality, stamps, verified | omg CLI only |
| Spec / plan / code proposals | Session + spawn_subagent |
| Outer “don’t stop” across many turns | Re-invoke this skill / user “continue” / optional omg ralph outer loop |
HARD RULES (non-negotiable)
- Fan-out only via Grok
spawn_subagent (depth = 1; children must not spawn).
- Always set
capability_mode on spawn:
- implementers (
omg-executor, write general-purpose): read-write (no Execute)
- critic / verifier / explore / plan:
read-only
- If spawn is DENIED for capability_mode: RETRY IMMEDIATELY same turn with the correct mode. Do not abandon multi-agent; do not solo-fallback after one deny.
- Never invoke
claude / codex / omc team / agy / cursor-agent as default workers.
- Grok tool names:
read_file, search_replace, run_terminal_command, spawn_subagent, grep, list_dir.
- Never write
passes / verified under .omg/state/. Only CLI after acceptance.
- No Stop hard-pin: PreToolUse is fail-open soft-guard. Do not claim OMC-style “chat cannot end until done.”
- Cancel with
omg cancel — never self-matching pkill -f.
Use when
- User says:
autopilot, auto pilot, full auto, autonomous, build me, create me, make me, handle it all, end to end, from idea to working code.
- Multi-phase work: requirements → plan consensus → implement → review → QA → accept.
- User wants hands-off orchestration inside this session and will re-prompt “continue” if the turn ends.
Do not use when
- Single tiny fix → work directly or
omg-ralph one story.
- Plan-only / critique-only →
omg-ralplan.
- Parallel burst only →
omg-ultrawork.
- Abort →
omg-cancel.
- User wants brainstorm without shipping → answer conversationally; do not start autopilot state.
Persistence honesty (read this)
| Want | How on Grok / OMG |
|---|
| Strict phases + destination gates | This skill + omg autopilot * |
| Outer retry until verified | omg ralph "…" or user re-invokes this skill after each turn |
| Host-forced continuation on Stop | Not available — see docs/research/stop-continuation/ |
If the session ends mid-phase: run omg autopilot status --run RUN and resume the playbook from the current phase.
CLI phase machine (normative)
interview → ralplan → implement → review → (rework) → qa → acceptance → verified
Illegal transitions fail closed. Destination gates (CLI-enforced):
| Enter phase | Required evidence / stamp |
|---|
ralplan from interview | interview_complete: true |
implement | consensus: true |
qa | CLI stages/structured_review.json clean |
acceptance | CLI stages/ultraqa.json status clean |
verified | Only omg autopilot complete after same-process accept — never transition … verified |
Session playbook
0. Bootstrap
omg doctor
omg setup
omg autopilot status --run RUN
If no run yet:
omg autopilot start "GOAL TEXT"
Record run_id from output. Prefer run_terminal_command for all omg invocations.
1. Phase interview (unless skip)
- Follow omg-deep-interview / CLI:
omg interview start "…" or continue with printed resume_command
omg interview answer … / pressure-pass / close
- When complete:
omg autopilot transition --run RUN --phase ralplan \
--evidence-json '{"interview_complete":true}' \
--reason "interview closed"
2. Phase ralplan
- Follow omg-ralplan playbook without product code:
- draft plan under run
stages/ + .omg/artifacts/
spawn_subagent critic read-only
- revise
spawn_subagent verifier read-only → stage artifact must contain whole-word APPROVE
- Prefer CLI when available:
omg ralplan "…"; then transition with evidence.
omg autopilot transition --run RUN --phase implement \
--evidence-json '{"consensus":true}' \
--reason "ralplan APPROVE"
3. Phase implement
- Decompose into stories / independent slices.
- Prefer omg-ultrawork patterns for parallel slices; omg-ralph one-story discipline for sequential must-finish.
- Spawn implementers with
capability_mode=read-write; worktrees for write-heavy work.
- Write notes under
.omg/artifacts/. Do not claim verified.
omg autopilot transition --run RUN --phase review --reason "implementation ready for review"
4. Phase review
- Prefer omg-dual-review / native critic→verifier (read-only).
- Or CLI:
omg review --run RUN --diff-text "…" --code-reviewer-json '…' --architect-json '…'
- On REQUEST CHANGES:
omg autopilot transition --run RUN --phase rework --reason "review findings"
omg autopilot transition --run RUN --phase review --reason "rework done"
When CLI stamps review clean:
omg autopilot transition --run RUN --phase qa --reason "structured review clean"
5. Phase qa
omg qa freeze --run RUN --scenarios-json '[{"id":"t1","command":"python3 -m pytest -q -m not live"}]'
omg qa run --run RUN
omg qa status --run RUN
omg autopilot transition --run RUN --phase acceptance --reason "ultraqa clean"
6. Phase acceptance → verified
omg accept --run RUN --yes
omg autopilot complete --run RUN
omg autopilot status --run RUN
Only then report success with evidence (commands + outputs).
7. Blocked / cancel
omg autopilot transition --run RUN --phase blocked --reason "…"
omg cancel
Capability cheat sheet
| Role | capability_mode | Notes |
|---|
| Implementer | read-write | No shell/Execute |
| Critic / verifier / explore | read-only | No product edits |
| Shell / tests for verified | CLI omg accept / omg qa | Never child self-verify run |
Anti-patterns
- Thin “done” prose without CLI stamps
transition --phase verified
- Skipping interview/ralplan gates by lying in evidence-json
- Self-approve after implement (skip dual-review)
- Infinite self-loop without CLI status (burn tokens; prefer status + continue)
- External agent CLIs as workers
- Claiming Stop hooks keep the session alive
Optional durable multi-story
When implement/QA spans more than one story that must survive process or run
boundaries (depends_on, checkpoints, cross-session resume), load omg-ultragoal
and drive omg goal * ledger in parallel with (or after) autopilot phases.
Grok has no host /goal — only the repo ledger under .omg/ultragoal/.
Still finish run acceptance via omg accept / omg autopilot complete before
omg goal link-run + omg goal verify.
Related skills
omg-using — router
omg-deep-interview, omg-ralplan, omg-ultrawork, omg-ralph
omg-ultragoal — durable multi-story goal ledger (no host /goal)
omg-dual-review, omg-ultraqa, omg-cancel
- Security:
docs/security-model.md
CLI quick reference
omg autopilot start "goal"
omg autopilot start "goal" --skip-interview
omg autopilot transition --run RUN --phase PHASE --evidence-json '{…}' --reason "…"
omg autopilot status --run RUN
omg accept --run RUN --yes
omg autopilot complete --run RUN
omg cancel