| name | l4 |
| description | Terse CEO front-door — Level 4: dispatch ONE background, worktree-isolated sub-orchestrator "foreman" that runs dev-flow unattended and returns a verdict; CEO context stays clean and the authoritative qc verdict is held at depth 0. Use when: "/l4 <goal>", "L4 <goal>", you want a long autonomous run offloaded off the main thread. Presets involvement=just-results, scope=Hold, project red lines plus -x additions (override --mode / --expand / --solo). Not for: inline execution (→ /l3), hetero impl engine (→ /l5).
|
/l4 — CEO autonomy, dispatched foreman
Terse front-door into autopilot:ceo-agent at Level 4: dispatch ONE
background, worktree-isolated sub-orchestrator "foreman" that runs dev-flow
unattended; the CEO holds the depth-0 control loop and the authoritative
qc verdict.
Hard rules:
- Startup presets identical to
/l3; engine all-Claude (hetero implementer → /l5).
- Before any TaskCreate, branch, worktree, runner, or model effect, run
node <plugin>/scripts/session-mode.js set --level l4 --repo-root <repo>. --solo uses
set --level l3 --entry-level l4 --fallback solo; a recorded precondition degradation uses
--fallback precondition_failed. The command admits canonical Mission policy/graph/source
coverage before writing the marker. Topology may change; its admission digest may not.
- qc@depth-0 is THE gate: reviewer families/panel come from
scripts/resolve-review-loop.sh (qc_panel / required_review_families /
min_panel_size); resolver unavailable → fall back to 3 reviewers. A
homogeneous (all-Claude) panel must not drop below the resolver's
min_panel_size (default 3) — panel size is emitted separately from
families because lens diversity ≠ family decorrelation (same-family lenses
can share blind spots). Reviewers read the branch diff, synthesized +
fix-before-integrate — NEVER a CEO self-read, and distinct from the
foreman's first-pass qc.
- Merge-back and worktree GC are owned by depth 0; budget cap is fail-closed
(
TaskStop + escalate). Record the run-summary ledger in the final CEO Report.
--solo → the /l3 inline engine (also the automatic degradation when the
foreman returns precondition_failed).
- Only admitted graph nodes become implementation tasks. Plan phase headings, modules, tests,
review seats, and retries stay inside the owning deliverable.
Capability profile (shadow): /l4 fixes foreman topology only. When the host supplies a current
verified envelope/grant/profile payload, forward it unchanged; never infer guidance density from
the level or model name. A late mismatch requires a fresh-session handoff.
The canonical profile-session.js lane proves only no-effect context isolation. An adaptive child
also needs an independently witnessed effectful adapter for the exact grant; otherwise retain this
level's existing guided dispatch path and report the adaptive row as unverified.
MUST-READ: ../ceo-agent/references/level-front-door.md
(§ The foreman, § Depth-0 control loop, § Run-summary ledger) — dispatch mechanics,
outcome→action table, and worktree-base rules live there, not here.