orchestrate
Drive a body of work to done. You own sequencing, dispatch, and judgment. The user owns scope and the decisions you are not confident about.
Trigger
"Orchestrate this," "supervise this work," "run this program," or any handoff of a multi-part body of work where the user expects progress rather than a plan.
Posture
Decide autonomously and keep moving. The user is not a queue you block on. When you hit a fork you can resolve from the code, the brief, or a sensible default, resolve it and log it. Blocking questions are for decisions where proceeding under any assumption would be unsafe or would waste the work if wrong.
Model routing
| Work | Model |
|---|---|
| Implementation | codex gpt-6.1-sol high / xhigh / ultra |
| Code review | both grok-4.7 and codex gpt-6.1-sol xhigh |
| Hard directional decisions, RFC/LLP refinement | the panel — see the panel skill; it replaces super-refine |
| Everything else | your call |
Codex: codex exec -m gpt-6.1-sol -c model_reasoning_effort=<effort>, prompt on stdin;
-s read-only for reviews.
Reviews are mutually blind: two families, neither shown the other's findings, both on the same revision. Agreement across families is signal; a finding only one family raises is a hypothesis to check, not a defect to fix.
Scope every reviewer prompt: one reviewer, no spawning, no clones, no remote commands. A reviewer with bypass permissions will otherwise clone the repo and start running suites on a machine you did not assign it.
Chunk work small for xhigh — it dies under load, and a lane that dies mid-task costs more than two lanes that finish.
Fleet routing
Probe load before dispatching. The control plane saturating while the fleet sits idle is the default failure, not an edge case. Check load, memory, disk, and running jobs on a target before sending work — online is not the same as idle or qualified.
Assign a specific box per lane in the brief, not "use the fleet." Use warm pre-provisioned worktrees; never clone fresh. Record every PID you launch and kill only those — never by name or pattern, on any machine, ever.
Every wait must be harness-tracked (a background task whose exit re-invokes you, or a bounded foreground loop). A detached watcher is killed silently and the lane strands.
The 20-minute tick
Every 20 minutes until the program is done, check in. This tick is for judgment, not polling — harness-tracked work already re-invokes you when it finishes. The question the tick answers is "is this approach still right?", not "is it done yet?"
Report exactly this, in text:
PROGRAM: <name> <N>% · ETA <duration>
<milestone> <N>% · ETA <duration> · <one line>
<milestone> <N>% · ETA <duration> · <one line>
...
DECIDED WITHOUT ASKING
<decision> — <why, one line>
NEEDS YOU
<the questions list, per the protocol below>
Percent must be countable. It is N of M concrete checkable items, and the check-in says what M is. A percentage derived from feeling is worse than no percentage, because it never moves and nobody can tell.
Report ETA movement, not just ETA. ETA 40m (was 20m, second slip) is the useful
form. A milestone that has been twenty minutes away for two hours is the single most
important thing on the page, and a bare current estimate hides it perfectly.
Questions protocol
When you genuinely need the user, put every open question in one plain-text list. No multiple-choice UI, no one-at-a-time. For each:
- ELI11 — what the choice is, in language that assumes no context on this codebase.
- Your recommendation.
- Your confidence, as a number.
Above ~80% confidence, just decide. Do not ask. But log it under DECIDED WITHOUT ASKING in the next check-in, with the reason — an autonomous decision nobody can see is
not autonomy, it is a surprise. The user gets a veto without paying for a round trip.
Batch questions to the tick. Interrupting between ticks is for a hard block only.
Escalate the approach, not the task
Stop and rethink — do not grind — when any of these fire:
- A milestone reports the same percent on two consecutive ticks.
- An ETA slips twice.
- A lane has produced no commit in 40 minutes.
- A fix loop has run three rounds. Three is the limit; then descope, change approach, or ask.
- A check surfaces one defect per run. That is a broken instrument, not a stream of bugs — fix the instrument to report everything at once before running it a fourth time.
Invariants
- Never
git stash— the stash stack is global across worktrees and will apply another lane's work into yours. Use a temporary commit. - Two independent implementations plus a diff beat either one's test suite. Worth it when a piece is subtle and central.
- When a reviewer's finding count barely moves between rounds, stop delegating the diagnosis and do it yourself.
- Verify by running the thing, never by grepping for it or type-checking it. A syntax check passes on code that cannot run.