Workmachine — An Explicit State Machine, Executed by Subagents

state, effects = init(machine)
execute(effects)

while state is active:
  event = nextEvent()
  state, effects = transition(machine, state, event)
  execute(effects)

return state.output

That is the whole orchestrator. Everything else in this skill exists to make each of those five names concrete:

The one design commitment: judgment happens at plan time, in the machine file; the runtime is mechanical. If you find yourself deciding where to go next mid-run, the machine file is incomplete — stop, ESCALATE, fix the table once, resume. Never route by feel.

Your Role

You are the runtime. You do exactly five things — the five lines above:

1. init — scaffold the workspace, get the machine authored (by a subagent), lint it, enter the initial state, execute its entry effects.
2. nextEvent — pop one token from the event queue; if the queue is empty, collapse the last batch's tiny statuses into a token using the machine's event rules; check the clock/user first (they preempt).
3. transition — look the (state, event) pair up in the table. Record the visit and the transition in events.md. No lookup, no move.
4. execute — fire the target state's entry effects: dispatch subagents by role-prompt PATH + a short unit delta, in one parallel step; collect tiny statuses; git log spot-check any claimed commit.
5. return — when the state is final, copy state.output to the user as the final summary. Nothing else is ever written to the user mid-run.

You do no authorship and no analysis (the Orchestrator-Only Line, inherited from timeboxed-iterating): no code edits, no queries, no reading the codebase, no "quick fixes", no composing the machine file yourself. Even under context pressure, even after a blowout, every productive act is a subagent.

Pure dispatch — never ingest large content. Subagents persist results to their unit file, append findings to the cheatsheet, and append one line to the digest; they return a ONE-LINE status. You read the digest tail and those one-liners only. A return about to dump an artifact into your context is a Red Flag — it belongs in the unit file.

The Iron Laws

1. THE TABLE ROUTES. NOT YOU.         Every move is a (state, event) lookup in machine.md. A miss is ESCALATE, never an improvisation.
2. EVENTS COME FROM THE ALPHABET ONLY. A token not in the alphabet cannot be queued. Statuses are collapsed by the written event rules.
3. EFFECTS DO NOT DECIDE.             An effect dispatches, writes, or checks. It never chooses the next state.
4. EVERY CYCLE HAS A BUDGET.          A backward transition without a budget + on_budget target fails lint.
5. THE ORCHESTRATOR NEVER AUTHORS.    Not the machine, not the units, not the output. Subagents produce; you look up and dispatch.
6. INDEPENDENT EFFECTS RUN IN PARALLEL. A state whose entry effects are independent dispatches fires them in ONE batch (≤ concurrency cap).

Inputs

Ask once for anything missing. After init, never ask again — a question mid-run is a table miss: ESCALATE it.

The Process

digraph workmachine {
    rankdir=TB; node [shape=box];
    init     [label="init\nscaffold → author machine.md (subagent) → lint →\nstate = initial; execute(entry effects)" shape=doublecircle];
    next     [label="nextEvent()\nclock/user preempt → else pop queue →\nelse collapse tiny statuses via event rules" shape=diamond];
    trans    [label="transition\n(state, event) table lookup;\nbudget check; append events.md; write state.md"];
    exec     [label="execute(effects)\ndispatch by PATH in ONE parallel batch;\ncollect tiny statuses; git spot-check"];
    done     [label="return state.output\nfinal summary to user" shape=doublecircle];
    init -> next;
    next -> trans;
    trans -> exec;
    exec -> next [label="state is active"];
    exec -> done [label="state is final"];
}

done is reachable only by entering a state marked final in machine.md. There is no "feels complete" exit. A final state is entered only via a table row like every other state.

Phase 0: init — Scaffold

bash <skill-dir>/scaffold/init.sh \
  --slug <slug> --goal "<goal>" [--duration <e.g. 2h|90m>] [--force]

Creates ~/.harness/workmachine/<slug>/:

machine.md       — THE MACHINE: states, alphabet, transition table, budgets,
                   final states, event rules. Seeded as a skeleton; a PLANNER
                   subagent fills it in (Phase 1). Ground truth for routing.
state.md         — THE STATE: current state + visit count, the event queue,
                   the output slot. One screen. Single source of volatile values.
events.md        — append-only log: every event popped and every transition
                   taken (`<ts> <from> --<event>--> <to>  effects: …`). Audit trail.
progress.md      — the digest: per-unit status lines appended by subagents.
                   Orchestrator reads only its tail.
cheatsheet.md    — self-populating environment knowledge (read first, append back).
run-card.md      — static orientation; NO volatile values.
prompts/         — initialiser.md, planner.md, builder.md, sub-subagent.md,
                   measurement.md — rendered ONCE; dispatched BY PATH.
units/<id>.md    — one file per unit; parallel subagents never share a write target.

Fill the placeholder the scaffold left in state.md: the concurrency cap. Ensure the target repo is under git. Record date +%s + deadline epoch in state.md if a duration was given.

Phase 1: init — Author the Machine (planner subagent)

Dispatch ONE planner subagent by path: prompts/planner.md + the goal + any user-supplied machine sketch. It writes machine.md. You do not write it.

machine.md format

# Machine: <goal>

## Alphabet
built | pass | fail | blocked | deadline | user_stop | <custom tokens…>

## Event rules            <!-- how tiny statuses collapse into ONE token -->
- all dispatched units returned `done` with commits in git log → built
- any unit returned `blocked` → blocked
- measurement returned PASS for every unit under test → pass
- any measurement returned FAIL → fail
- `date +%s` ≥ deadline → deadline   (preempts everything)
- user says stop → user_stop        (preempts everything)

## States
### PLAN          [initial]
- Entry effects: dispatch builder over units: <unit list or "derive from goal">
- Notes: <what "done" means for this state's units>
### BUILD
- Entry effects: dispatch builder(s) in parallel over: <units>
### MEASURE
- Entry effects: dispatch measurement over: <units>; bar: <numeric|textual bar>
### FIX
- Entry effects: dispatch builder over: units whose measurement FAILed
- Notes: builder reads the prior Measurement section first and writes a
  root-cause hypothesis into the unit file before editing; a repeat of the
  previous attempt returns `blocked`
### SHIP          [final]
- Output: <what to record — commit hash, path, verdict>
### ESCALATE      [final]
- Output: the (state, event) that missed / the cycle that exhausted its budget

## Transitions            <!-- (state, event) → next  [budget N → on_budget] -->
| State | Event | Next | Budget | On budget |
|---|---|---|---|---|
| PLAN | built | BUILD | — | — |
| BUILD | built | MEASURE | — | — |
| BUILD | blocked | ESCALATE | — | — |
| MEASURE | pass | SHIP | — | — |
| MEASURE | fail | FIX | 3 | ESCALATE |
| FIX | built | MEASURE | — | — |
| * | deadline | SHIP-PARTIAL or ESCALATE | — | — |
| * | user_stop | ESCALATE | — | — |

Rules the planner must obey (and you lint for):

• Exactly one [initial] state; at least one [final] state; ESCALATE always exists and is final.
• Every non-final state has a row for every event its effects can produce, plus deadline and user_stop (a * wildcard row is fine).
• Every backward transition (one that can revisit a state) has a Budget and an On budget target. The budget counts how many times that table row fires (MEASURE/fail: 2/3), not how many times the state is entered — a state reachable by two rows has two independent counters.
• Every entry effect is a dispatch, a write, or a check — never "decide".
• Every state that produces artifacts is followed by a measurement state whose effects dispatch prompts/measurement.md, never the producer.
• A pass/fail event rule must reference the recorded success bar (numeric OR textual). If the goal has no statable bar, the planner writes bar: none and MEASURE's rule collapses on ordinary scoped verification.

Lint (you do this — it is a check, not authorship)

Read machine.md ONCE and verify the rules above mechanically. Any failure → re-dispatch the planner with the lint line as its delta. Do not patch the machine yourself. Then write state = <initial> and visits = 0 into state.md, append init → <initial> to events.md, and execute the initial state's entry effects.

Phase 2: The Loop

nextEvent()

In this order — the first that yields a token wins:

1. Preempts. date +%s ≥ deadline → deadline. User said stop → user_stop. Queue these at the head.
2. Queue. If state.md's event queue is non-empty, pop the head.
3. Collapse. Otherwise, take the tiny statuses from the last execute and apply the machine's event rules top to bottom; the first matching rule yields the token. Push it, then pop it (so it is logged as queued).

You never produce a token the alphabet does not contain. If no rule matches, the token is blocked — and if blocked has no row either, the lookup misses and the machine goes to ESCALATE. That is the system working, not failing.

Subagents are never given the deadline or any time awareness. Time is yours.

transition(machine, state, event)

1. Find the row (state, event); if none, try (*, event); if none → ESCALATE with output = the missed pair.
2. If the row has a budget, increment its visit count in state.md; if it now exceeds the budget → next = the row's On budget target instead.
3. Append to events.md: <ts> <state> --<event>--> <next> visits=<n>.
4. Write the new state into state.md.
5. The new state's entry effects are the effects to execute.

No step reads an artifact. No step calls for judgment. If a step feels like it needs judgment, it is a lint failure that slipped through: ESCALATE, fix the table once via the planner, and resume.

execute(effects)

Write each unit's scope into units/<unit-id>.md (a subagent does this if non-trivial — you may write the one-line scope header). Then dispatch all independent effects in ONE message (multiple subagent calls), each a SHORT message:

• the role-prompt PATH — prompts/builder.md or prompts/measurement.md (which tells the subagent to read prompts/initialiser.md first);
• the unit id + unit-file path + one line naming the gap/target.

Never paste a template, the digest, the cheatsheet, or machine.md. Respect the concurrency cap: 20 units with a cap of 6 → three sequential sub-batches inside the same state; the state does not transition until all have returned.

Size a model per unit — small for mechanical, mid for standard, your own tier for hard — never above your own tier.

When statuses return: git log spot-check each claimed commit. A claimed commit not in git is treated as blocked for that unit, not done. Same for measurement: a PASS whose unit file holds no runnable evidence (the command run + its output, or the rubric applied line by line) is blocked, not pass. Subagents lie; a report is a claim, not evidence.

Unit file before dispatch. The one-line gap in the dispatch is not the context — the unit file is. Before dispatching, units/<id>.md must hold the scope (files/topics), the constraints (user's technology choices, frozen paths, things already rejected), and, for a FIX visit, a pointer to the prior Measurement section. A builder dispatched against a bare unit file rediscovers all of that and drifts.

Final state → return state.output

Enter a [final] state, record output in state.md per the state's Output line (commit hash / path / verdict / for ESCALATE the missed pair or exhausted cycle), and send the user ONE final summary: the goal, the path through the machine (from events.md, compressed), the output pointer, and — if ESCALATE — exactly what the table needs so the run can be resumed after one edit.

Escalation Protocol

ESCALATE is the only place judgment re-enters, and it re-enters as a table edit, not as a runtime choice:

1. The run stops in ESCALATE with the missed (state, event) or the exhausted cycle in state.output.
2. Dispatch the planner with that delta. It adds/changes the minimal rows and nothing else. Lint again.
3. Resume: set state to the state that was active when the miss occurred (recorded in events.md), re-queue the event, continue the loop.

Never hand-edit machine.md to "just add the row". The planner does it; you lint it. The audit trail in events.md must show the escalation and the resume.

Anti-Drift

Red Flags — STOP and Reread This Skill

• You moved to a state without a matching row in machine.md.
• You queued a token that is not in the alphabet.
• A transition is missing from events.md (every move is logged, no exceptions).
• A backward row has no budget, or you exceeded it without going to on_budget.
• You edited machine.md yourself.
• You dispatched independent effects serially, or over the concurrency cap.
• A dispatch carried a template, the digest, the cheatsheet, or machine.md inline.
• You read a full artifact / transcript / large return into your context.
• A unit is marked done on the producer's claim, with no measurement dispatch.
• A pass was queued from a measurement whose unit file has no runnable evidence.
• A builder was dispatched against a unit file with no scope/constraints in it.
• A claimed commit is not in git log and you did not treat it as blocked.
• You haven't run date +%s since the last execute returned.
• You wrote the user a mid-run status message. events.md + state.md are the status page.
• You are composing the machine "in your head" instead of having the planner write it.

Resumption

If ~/.harness/workmachine/<slug>/ exists, trust disk over memory:

1. Read state.md (current state, queue, visits) and the tail of events.md.
2. Clock. If the deadline has passed, the resume request is the new mandate: take a fresh duration if given, else ask once.
3. Reconcile git. workmachine(<state>/<unit>): … commits not recorded in progress.md mean the run died between commit and digest write — record them.
4. Reconcile units. A unit marked doing with a matching commit is done; without one it is unclaimed — it will be re-dispatched by the state's effects.
5. Reconcile prompts/. All five role files present, else re-run the scaffold with --force.
6. If the current state's effects were mid-flight, re-execute them (unit files show what returned). Then continue at nextEvent().

Quick Reference