| name | foreman |
| description | The orchestration pattern for staffing implementation across cheaper model seats — the boss (the orchestrating agent) writes specs and constitutions, dispatches, blind-verifies, and rules on disputes, but never implements. Includes the audition oracle for hiring seats, merge-train discipline (post-merge build oracle on main), and watch-what-you-order review. Load for any arc with several independent implementation units, or any grindable, oracle-covered build. |
Foreman — the boss never implements
The one-sentence version: the orchestrating agent (the boss) writes the spec and the constitution, staffs implementation on cheaper seats, verifies blind, rules on disputes — and never implements. Done right, the boss does more judging, not less.
When to load
- Any arc with ≥ ~3 independent implementation units (files, endpoints, fixtures, migrations, assets — units a worker can own end-to-end).
- Any grindable, oracle-covered build — even a single big unit — where retry-until-green is safe (cheap deterministic re-runs, no side effects per attempt).
When NOT to: single-file hands-on work (the org chart has overhead); intent-heavy prose (voice belongs to the human — it never gets delegated); ungrindable work where each attempt touches the world (rate limits, live sends, one-shot state) — that stays on the human cadence, no worker loops.
Seats and pools
Speak in seats and tiers, not personalities. A seat is a named role bound to a specific model and dispatch recipe; a pool is the budget it draws from.
| Seat | Role | Never does |
|---|
| Boss (the orchestrating/session model) | Spec, constitution, dispatch, dispute rulings, final full-context integration review | Implement. Not "rarely" — never. The moment the boss writes the code, the economics invert. |
| Reasoning seat (strongest worker tier) | Hard-reasoning units (concurrency, algorithms, gnarly refactors); shepherd-review of a risky merge | Bulk work the bulk tier can do |
| Bulk seat (capable mid tier) | Default implementer AND bulk work (ports, fixtures, renames, test authoring, checker duty) | — |
| External seats (your roster) | Per their audition scores — dispatch recipes live with the roster | Join without an audition |
The gotcha that motivates this table: subagent dispatches typically inherit the session model unless a model is set explicitly — an unstaffed fan-out runs N boss-priced workers, the exact opposite of the point. Every worker dispatch names its seat's model explicitly. On a subscription plan the saving is measured in boss-share of weighted usage rather than dollars; it is still the binding constraint.
The five moves
1. Constitution first. Before any dispatch, write the done-right standard: 5–15 checkable points, each naming its oracle where one exists (the test suite, the linter, a smoke count, exact file layout, banned patterns). The constitution is not the spec — the spec says what to build, the constitution says how everyone (worker, checker, boss) will know it's right. It goes verbatim into every worker and every checker prompt. Prompt the standard once; never steer task-by-task.
2. Staff explicitly. Pick seats from the table; say the staffing out loud in the green-light message ("6 units → bulk-seat workers, the reasoning seat on the scheduler unit, bulk-seat checkers, oracle = the smoke suite"). Boss-implemented units = zero.
3. Dispatch = spec + constitution + oracle name. Each worker gets its unit's spec, the constitution verbatim, and the name of the oracle that will judge it. Workers return artifacts (or write to assigned paths). Their self-report is for the boss's ledger only — checkers never see it.
4. Blind verify. Oracle first — deterministic, cheap, no judgment spent on what a script can catch. Then a checker seat with fresh context that receives the artifact + constitution only — never the worker's claims — and executes and measures: run it, refetch it, re-count it, screenshot it. Failures bounce back as specific defects ("test 7 fails: ParseRange("5-") returns 5,0 want error"), not "try again." Two bounces → a fresh worker, not a third retry (entropy collapse: a worker marinating in its own failure converges on the same wrong idea).
5. The foreman's report. At arc close: units delegated, bounces, what the checkers caught, disputes ruled, and the estimated boss-share saved. What bounced and why is the interesting part — report it plainly.
Merge trains end with a MAIN build oracle
Blind verify covers each PR — it does NOT cover the boss's own merge-conflict resolutions. On one all-night multi-arc run, every per-PR check was green while the boss's hand-stitched conflict resolution had dropped a single closing paren; the live deploy kept working off a pre-break image, and main sat unbuildable until a worker caught the foreman's own merge slip. The rule: after the LAST merge of a train, run the repo's build oracle against merged main itself (build, vet, or a CI dispatch) before declaring the train landed. The boss's fingers are the one worker no checker was watching — so the oracle watches them.
Watch what you order
Review worker output against intent, not just completion. A green oracle proves the unit does what the constitution measured; only the boss's full-context read proves it does what the human meant. Two consequences:
- The dispute rule — no rank above verification. A worker may contest a checker's fail; the boss rules on the spec, both directions — a checker enforcing a rule the constitution never named gets corrected, same as a cheating worker. And the boss's own output (specs, designs, the constitution itself) goes through the same oracles. Nothing outranks the check.
- The final integration review is real work. A latent defect can pass build + vet + tests + a checker and still be wrong; the full-context review by the boss is the ceiling of the whole pattern. Budget for it — don't spend the boss's context implementing and then skip the one thing only the boss can do.
Hiring — the audition oracle
Never hire a model seat without an audition. Every new seat passes a small, standardized, paid audition task before joining the roster; re-audition on any model or version bump. The audition is a spec-precision coding tryout judged by a deterministic pass/fail checker: table tests, compliance greps, and an examples block whose claims are executed against a reference implementation, not trusted. Scores and the exact dispatch recipe that worked go in the roster — the roster is data, and future staffing reads it instead of re-guessing.
Ceilings — what this pattern cannot do
- Delegation ceiling = oracle coverage. No oracle over a unit → the boss reviews it by hand, or builds the oracle first. Widen delegation by widening the verification floor, not by trusting workers harder.
- The full-context shepherd is the ceiling. See "Watch what you order" — some defects are only visible to a reviewer holding the whole picture.
- Intent stays the human's. Anything touching voice, design taste, or strategy is surfaced, not delegated — the human-in-the-loop bins apply unchanged.
