Foreman

No. Foreman does not replace the worker's reason, tool, and observe loop, does not choose its tools or files, and keeps missions broad. It runs a debounced observation loop beside the worker so the factory can think about the work without the subprocess ever going idle.

Only on the Codex App Server backend. That transport keeps the active Codex thread addressable, so Foreman can push a supervisory update into an in-flight turn. The OpenCode and Hermes backends, and the older FOREMAN_CODEX_BACKEND=exec transport, degrade to stop and retry, and Foreman runs one worker at a time.

The repository is never dumped wholesale. The root AGENTS.override.md or AGENTS.md is read for each observation and included only in the transient Jev request, without being persisted in factory state, and each observation is bounded by a 20,000-character diff, 12,000-character output tails, 30 recent events, and 10 workers of history.

Responsibilities own Jev checks plus deterministic logic that may propose directives, and all active checks go out in one parallel request as independent Noul questions. Python policy then selects exactly one directive from the proposals, so continue, steer, stop, retry, verify, or finish is decided in code from probabilities rather than by the model.

Each one is compact and bounded: the original job and current factory status; active worker summaries, recent worker history, output tails, exit status and elapsed time; git status, a bounded diff and changed file names; repository-root AGENTS.override.md or AGENTS.md instructions when present; verification results and recent persisted events; the prior Foreman result and intervention; and attempt and failure counts with elapsed factory time. Defaults are a 20,000-character diff, 12,000 characters per captured output tail, 30 recent events and 10 workers of history, all in FactoryConfig.

Real runs need a TypeSafe API key plus the worker CLI. Select the backend with FOREMAN_WORKER_BACKEND, one of codex (the default), opencode, hermes or deepagents: FOREMAN_WORKER_BACKEND=opencode foreman run --repo ./my-project --job "Add request retries". OpenCode runs non-interactively and streams output, Hermes runs hermes chat headless in single-query mode and decodes its newline-delimited JSON events, and Deep Agents runs the separately installed dcode CLI headlessly with bounded output and turn and time budgets.

Yes. A verifier is another worker using the selected backend, given an independent, deterministic verification mission rather than the original job. That is what feeds tests_sufficient and needs_verification, and the point is independence: the agent that wrote the code does not get to grade itself. Foreman currently runs one coding worker at a time, and the worker implementation is replaceable behind a small protocol rather than Codex-specific types.