Operate

/operate — run long, without silent drift

Goal: let an agent work unattended across many iterations (or wake on a schedule) and guarantee it halts rather than drifting — spinning on a dead end, burning budget, or quietly breaking the repo. This skill is the discipline layer: the platform /loop and /schedule are the engine; /harness-claude:operate adds guardrails, durable state, and a drift check wired to the harness's own eval skills.

Opt-in. No default-pipeline skill or hook starts a run. Running long is always explicit. Git boundary: a run never commits or pushes unless you explicitly arm it for that in the objective; branch creation for the run's non-trivial work follows rules/git.md (branch-at-first-write).

How it works

Each iteration the operator does one increment of work, then runs a checkpoint:

node scripts/operate/step.js --id <run-id> [--spec <path>] [--cmd '<check>' ...] \
     [--max-iterations <n>] [--max-fails <n>] [--budget-ms <n>]

step.js loads the durable run state (.claude/runs/<id>.json), runs the drift check — harness-claude:health (test/lint pulse) and, when --spec is set, harness-claude:eval (acceptance-criteria gate) — updates and persists state, evaluates the guardrails, and exits:

• 0 = continue — schedule the next iteration.
• 1 = halt — a guardrail tripped (drift | budget | iteration-cap); stop and report.
• 2 = usage/error — fix the invocation (e.g. missing --id).

The state file is the sole source of truth across firings — each /loop firing may be a fresh context, so iteration count, budget spent, and the consecutive-fail counter persist there and resume automatically. Config flags are honored only when the run is first created; later firings ignore them so counts accumulate rather than reset.

Do this

1. Frame the run. Name a one-line objective, a --max-iterations ceiling, a drift threshold --max-fails (default 2 — one failure can be transient, two is a trend), and a --budget-ms wall-clock cap. Point --spec at the spec the run must keep satisfying when one exists (that arms the harness-claude:eval half of the drift check).
2. Pick a trigger.
   • Interval / autonomous: drive iterations with the platform /loop (fixed interval) or let the model self-pace; each firing advances the work one increment, then calls step.js.
   • Scheduled / cron: use the platform /schedule to wake on a cadence; each wake runs one checkpointed iteration.
3. Supervise by exit code. Continue on 0; on 1, stop the loop, read the printed summary, and surface the halt reason and the failing criterion — never restart blindly past a drift halt. Delegate the per-iteration work to the harness-claude:loop-operator agent.
4. Report on halt. Always end with the run summary: iterations run, budget spent, last verdict, halt reason. (step.js prints it; relay it.)

Guardrails (each can HALT — priority: drift > budget > iteration-cap)

• drift — --max-fails consecutive failing checkpoints → stop and surface. The "can't silently diverge" guarantee. Outranks the benign ceilings. On a drift halt the failing check itself is recorded secret-safely in the eval artifact .claude/eval/continuous/latest.json (check name + command + pass, never output) — read it to see what failed after the fact.
• budget — wall-clock (--budget-ms) exceeded. (Token/cost is best-effort in v1 — a plain /loop doesn't expose token spend; lean on wall-clock + iteration count.)
• iteration-cap — --max-iterations reached.

0 is a live cap, not "off". --max-iterations 0 halts before any work, --max-fails 0 halts on the first checkpoint, --budget-ms 0 halts at once. To disable a cap, omit the flag.

Security & trust

• step.js shells the eval runners, which execute the repo's own check commands / spec eval fences with your privileges — same trust boundary as harness-claude:health / harness-claude:eval. Don't /operate a repo whose checks you don't trust.
• Secret-safe: the run state records only metrics, verdicts, and the halt reason — never a check's output (the eval runners stream output to the tty and capture nothing).

Notes

• Self-contained: pure Node (scripts/operate/{step,state,guardrails}.js) + the existing eval runners. No new dependency, no new MCP server.
• Distinct from a one-shot harness-claude:verify: this is a bounded, guarded, many-iteration run with durable state.

Exit criterion

A run has reached a clean halt (done, iteration-cap, budget, or drift) with its summary reported, and on drift the failing criterion surfaced — never a run left silently looping.