Autonomy Command — Suit Default Operating Model
v4.0 · Codex terms 67–68 · Skill: orchestrator · temperament-aware ceremony
What it is
The autonomy command is the default operating model when the 0xRay suit is worn. Users describe goals in plain language ("fix this", "continue P0.1", "ship it") — not framework keywords. The lead agent takes the helm, creates phased plans with todos, delegates to the best subagents, and loops until tests are green.
Problem it solves: keyword-gated skills, unused todo lists, buck-passing, skipped researcher/architect/code-reviewer on major work, and "not my job" syndrome.
Lead cadence: a peer wearing the suit continues from Station + Repertoire without a command stream. Dummy tests SKILLS.md. Same-sess bodies keep a Station card. Branch + worktree + PR. Ship when mill-gate D + foundry gate PASS — green CI is not ship. OTP then poll the published version. Loop until Station is done. Live ticks, not idle pacers. Disk SSOT: grok-bot/ops/LEAD-CADENCE.md. Playbook stays the orchestrator skill (Codex 69 — no new skill).
On guided/strict hosts, intake (analyze-complexity) is required before spawn. On frontier, the engine is available but spawn warns instead of denying. Governance (Codex 11/29/69) is always on.
Kernel-level activation (v4.0 — temperament)
Lead-dev mode is wired through multi_agent_orchestration in features.json and host hooks (no separate autonomy_kernel block):
| Layer | Mechanism | What it does |
|---|---|---|
| SessionStart hook | dist/integrations/grok/hooks/session-start.js | Boots lead-dev mode; writes session-boot.json |
| PreToolUse hook | pre-tool-use.js | Codex deny + synthesis/pending/spawn gates + per-suite triage hint |
| Orchestrator MCP | analyze-complexity | Persists phased plan + todos to lead-dev-plan.json |
| features.json | multi_agent_orchestration.lead_dev_mode | Kernel config SSOT (threshold, triage, consults) |
| Skill | autonomy-command | Human-readable playbook (synced on grok install) |
"multi_agent_orchestration": {
"lead_dev_mode": true,
"per_suite_test_triage": true,
"phased_plan_threshold": 25,
"auto_consult_major_work": true,
"auto_chain_delegations": true
}
First substantive task in every session: xray-orchestrator → analyze-complexity({ tasks: [...] })
When it activates
| Condition | Result |
|---|---|
npm install 0xray + npx 0xray grok install | Hooks + MCP + skills synced |
| Grok SessionStart | Autonomy kernel boot JSON on stdout |
| Substantive user task | Agent calls analyze-complexity → phased todos |
| Full test suite command | PreToolUse emits per_suite_triage_required hint |
Tier 1 consumers (repertoire, groover, chrono-warp-drive) may add npm run confirm:suit for layered boot checks.
The seven rules
- Phased plan + todos — Every multi-step effort gets phases with detailed todos. Assign the best subagent per item. Monitor output. Iterate until done.
- Take the helm — Loop: implement → test → fix → repeat. Do not stop for permission or status theater.
- Per-suite test triage — After major changes, run individual test suites, triage, fix, rerun. Full suite only when all affected suites pass.
- Delegate, don't abdicate — Lead stays on the main thread, reactive to the user. Subagents execute. Lead monitors and updates todos.
- Read all output — Console, stderr, test logs. Triage every failure. Rerun until clean.
- No buck-passing — Never defer errors as "pre-existing." Add a todo and resolve.
- Resolve all errors — Zero open failures at phase completion.
Lead-dev loop
INTAKE → CLASSIFY (analyze-complexity + trap check)
→ PLAN (complexity >25: strategist + architect-tools)
→ DISPATCH (orchestrate-task / Task subagent)
→ MONITOR (lead reads all output; TodoWrite)
→ TEST (per-suite triage → full suite gate)
→ DONE or loop
Hard stops (human only): secrets/OAuth, production deploy/cron, explicit user redirect.
Subagent routing
| Task | Subagent | MCP |
|---|---|---|
| Strategic phasing | strategist | orchestrate-task |
| Major refactor / architecture | architect-tools | context-analysis, architecture-assessment |
| Research / prior art | researcher | search_codebase, get_documentation |
| Implementation | backend-engineer / frontend-engineer | orchestrate-task |
| Test failures | bug-triage | triage → fix → rerun |
| Pre-merge review | code-review | check_best_practices |
| Trap-classified work | architect + repertoire | repertoire__get_task_confidence (stdio) |
| Codex gate | enforcer | codex-enforcement |
Before major planning or refactors: invoke researcher + architect-tools + code-review — automatically, not only when the user names them.
Agent panel consensus (2026-06-18)
| Agent | Recommendation |
|---|---|
| Strategist | Default ON after suit boot; complexity >25 requires phased plan before edits |
| Architect-tools | Document sync via AGENTS-consumer.md postinstall; architectural review on major refactors |
| Researcher | Use search_codebase / get_documentation when governance LLM unavailable |
| Code-reviewer | Per-suite triage; lead owns every error; skill doc passes 100/100 best practices |
Operationalization in core repos
| Repo | Boot check | AGENTS.md |
|---|---|---|
| xray | npx 0xray health | Framework + AGENTS-consumer.md (ships on postinstall) |
| repertoire | npm run confirm:suit | § Default operating mode |
| groover | npx 0xray health | § Default operating mode |
| chrono-warp-drive | npx 0xray health | § Default operating mode |
Skill source: src/skills/autonomy-command/SKILL.md · synced to ~/.grok/skills/ on npx 0xray grok install.
Related
- Grok floor
- SKILLS.md —
autonomy-commandcatalog entry