| name | autopilot |
| description | In-session OMA autonomous delivery — invoke /oh-my-agy:autopilot with a goal; run deep-interview→ralplan→ultragoal→code-review→ultraqa HERE (no terminal required) |
| argument-hint | <product idea or task> |
autopilot (OMA / in-session)
You are already in the agent session
When invoked via /oh-my-agy:autopilot (canonical namespaced slash) or this autopilot skill, treat $ARGUMENTS as the goal and run the full delivery phases HERE in this conversation.
- Do not tell the user to open a terminal first.
- Do not require
oma autopilot start / drive / SID / CID / revision as the default happy path.
- Bare
/autopilot is left to OMC (or host) if present; OMA’s canonical form is /oh-my-agy:autopilot.
- Optional durable ledger via
oma CLI lives only in the Appendix--use it when the user wants cross-session resume, never as a gate to start work.
Purpose
End-to-end autonomous delivery inside the current agent session (Claude Code / Grok / Antigravity host). Sibling of OMC/OMX autopilot; tools are host-native. Workspace artifacts stay under .agy/.
Maps to the OMX five-phase contract.
Default loop (strict OMX)
deep-interview → ralplan → ultragoal → code-review → ultraqa → complete
Optional: team during ultragoal for multi-worker DAG (explicit only).
Non-clean code-review / ultraqa → return to ralplan with return_to_ralplan_reason — do not ad-hoc patch outside the loop.
Production / causal-trace evidence may still be required as a terminal gate before claiming complete.
Legacy phase aliases (compat)
| OMX canonical | Legacy OMA |
|---|
| deep-interview | requirements |
| ralplan | planning |
| ultragoal | executing |
| code-review | review |
| ultraqa | qa |
Canonical skill handoffs (in-session)
| # | Phase | Slash / skill | Workspace artifacts |
|---|
| 1 | deep-interview | /oh-my-agy:deep-interview · skills/deep-interview/SKILL.md | .agy/specs/… |
| 2 | ralplan | /oh-my-agy:ralplan · skills/ralplan/SKILL.md | .agy/plans/… (critic APPROVE) |
| 3 | ultragoal | /oh-my-agy:ultragoal · skills/ultragoal/SKILL.md | .agy/ultragoal/… + skills/verify (+ optional team) |
| 4 | code-review | /oh-my-agy:code-review · skills/code-review/SKILL.md | .agy/reviews/… → APPROVE+CLEAR or REQUEST CHANGES |
| 5 | ultraqa | /oh-my-agy:ultraqa · skills/ultraqa/SKILL.md | .agy/qa/… → PASS / FAIL / justified SKIP |
| 6 | complete | — | production evidence + mode inactive |
Within ultragoal, ralph / ultrawork remain component disciplines; they do not replace the phase skill.
Use when
- User invokes
/oh-my-agy:autopilot or says autopilot / full auto / build me / handle it all
- Multi-phase work needs clarify → plan → implement → review → QA in one session
Do not use when
- Single small fix → use
/oh-my-agy:ralph or direct work
- User only wants explanation → conversational answer
- User only wants a plan → stop after
ralplan (do not enter ultragoal)
In-session state (workspace mirror)
Happy path needs no session id. Optionally maintain a workspace note for continuity:
.agy/autopilot/state.json
Minimum fields:
{
"mode": "autopilot",
"active": true,
"current_phase": "deep-interview",
"phase_cycle": [
"deep-interview",
"ralplan",
"ultragoal",
"code-review",
"ultraqa"
],
"iteration": 1,
"goal": "<goal from $ARGUMENTS>",
"handoffs": {
"deepInterview": null,
"ralplan": null,
"ultragoal": null,
"codeReview": null,
"ultraqa": null
},
"return_to_ralplan_reason": null
}
Update current_phase and handoffs as you finish each phase. This file is a workspace convenience, not a hard dependency.
Steps (by phase) — run HERE
0. Start
- Parse
$ARGUMENTS (or the user message) as the goal.
- Write/refresh
.agy/autopilot/state.json with active: true, current_phase: "deep-interview".
- Enter Phase 1 immediately — no CLI preamble.
1. deep-interview (clarify)
- Follow
skills/deep-interview/SKILL.md (slash: /oh-my-agy:deep-interview).
- Restate goal, non-goals, constraints, acceptance.
- If still vague: max 3 focused questions or write explicit skip rationale authorized by user.
- Handoff:
.agy/specs/… → set handoffs.deepInterview → advance phase to ralplan.
2. ralplan (plan)
- Follow
skills/ralplan/SKILL.md (slash: /oh-my-agy:ralplan).
- Author + steelman + critic passes; write plan artifact.
- Do not implement until critic APPROVE.
- Handoff:
.agy/plans/… → advance phase to ultragoal.
3. ultragoal (implement + verify)
- Follow
skills/ultragoal/SKILL.md (slash: /oh-my-agy:ultragoal).
- Ledger under
.agy/ultragoal/; mandatory skills/verify/SKILL.md with fresh command output.
- Independent slices →
ultrawork; multi-worker → team (explicit only).
- Never reduce scope silently.
- Handoff → advance phase to
code-review.
4. code-review (merge readiness)
- Follow
skills/code-review/SKILL.md (slash: /oh-my-agy:code-review).
- Artifact under
.agy/reviews/ with APPROVE+CLEAR or REQUEST CHANGES.
- On REQUEST CHANGES → return to
ralplan (or fix + re-review; do not skip the artifact).
- On APPROVE+CLEAR → advance to
ultraqa.
5. ultraqa (adversarial QA)
- Follow
skills/ultraqa/SKILL.md (slash: /oh-my-agy:ultraqa).
- Artifact under
.agy/qa/. Docs-only may SKIP with written reason; code changes must run scenarios.
- On FAIL → return to
ralplan / fix loop; re-run ultraqa after fix.
- On PASS / justified SKIP → ready for complete.
6. Complete
- Mark inactive only when review + QA are clean (or QA skip justified) and any required production gate passes.
- Set
.agy/autopilot/state.json → active: false (or delete/archive).
- Summarize: what shipped, evidence commands, residual risks.
Stop conditions
- User: stop / cancel / abort → follow
cancel skill; stop advancing phases
- Same failure 3× with no new plan → report blocker
- Missing credentials / irreversible production action → ask user
Anti-patterns (forbidden)
- Opening with “run
oma autopilot start / drive first”
- Blocking work on missing SID / CID / revision in the happy path
- Claiming complete without review + QA (or justified QA skip) artifacts
- Silent scope reduction or skipping phases without written rationale
Final checklist
Appendix: optional oma CLI ledger
Use only when the user wants a durable outer ledger (cross-session resume, managed binding). Not required to start or finish in-session work.
Discover skills
oma skill list
oma skill show autopilot
Start / drive / status (optional)
oma autopilot start -- "Ship feature X"
oma autopilot status --session <id>
oma autopilot doctor --session <id>
oma autopilot drive --session <id> --conversation <cid> --expected-revision <n>
oma autopilot resume --session <id> --conversation <cid> --expected-revision <n>
Handoff & phase advance (optional)
oma autopilot handoff --session <id> --expected-revision <n> \
--key deepInterview|ralplan|ultragoal|codeReview|ultraqa \
--path <artifact>
oma autopilot advance --session <id> --expected-revision <n> --evidence <file>
oma autopilot consensus --session <id> --expected-revision <n> \
--role architect|critic --verdict approve|revise --note <text>
oma autopilot return-ralplan --session <id> --expected-revision <n> --reason <text>
oma autopilot review --session <id> --expected-revision <n> --evidence <review-evidence.json>
oma autopilot qa --session <id> --expected-revision <n> --evidence <qa-evidence.json>
oma autopilot checkpoint --session <id> --expected-revision <n> --evidence <file>
oma autopilot cancel --session <id> --expected-revision <n> --reason <text>
oma autopilot reset-breaker --session <id> --expected-revision <n>
When session-bound, keep CLI revision/handoff in sync with the .agy/ artifacts you already wrote in-session. The in-session artifact is the source of truth for phase quality; CLI records durability.