GitHub - thruwire/foreman: Software factory foreman based on TypeSafe's Jev model

Foreman

Foreman watches the software factory floor with TypeSafe AI's Jev , placing a fast decision model above slower coding agents.

Give it a ticket, specification, bug report, or any free-form software job.

A Codex or OpenCode worker does the software engineering while Foreman independently assesses whether the implementation is complete, requirements are satisfied, tests are sufficient, verification is needed, or human input is required.

SOFTWARE FACTORY Codex/OpenCode Codex/OpenCode Tests worker worker │ │ │ │ └──────────────────┼──────────────────┘ │ factory evidence ▼ FOREMAN Jev │ ▼ implementation_complete .91 tests_sufficient .34 requirements_satisfied .79 worker_stuck .

02 needs_verification .82 work_off_track .06 agents_md_drift .01 meaningful_progress .94 ready_to_finish .21 │ ▼ continue / steer / stop / retry verify / finish

Generative models work. Foreman watches the work.

Foreman is an architectural experiment, not a claim that this design is already better than a conventional coding-agent harness.

Documentation

• Theory: semantic supervision

• Why Jev fits the experiment

• What Foreman is proving

• Runtime and event flow

• Live steering

• Worker backends

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup and pull-request guidance. Please also read the Code of Conduct and report vulnerabilities according to the Security Policy .

What is Foreman?

Foreman is a native Python asyncio runtime with two concurrent loops:

CODING AGENT LOOP FOREMAN LOOP reason watch │ │ ▼ ▼ tool assess │ │ ▼ ▼ observe ───────── factory events ────────► Jev │ │ ▼ ▼ edit decide (Python policy) │ │ ▼ ▼ test ◄──────────── intervention ────────── intervene │ └── continue

Foreman does not replace a worker's reason/tool/observe loop and does not choose its individual tools or files. Missions stay broad.

The important property is that the worker does not have to stop working for the factory to think: worker output and lifecycle events flow into an independent, debounced observation loop while the subprocess remains active.

Why build this?

Coding agents are relatively slow, stateful generative systems. Supervisory questions such as “is this worker stuck?” or “does this now need independent verification?” are narrower.

Jev is interesting here because TypeSafe describes it as accepting structured state and typed questions, returning probabilistic decisions, and evaluating multiple questions independently in one parallel request.

Foreman explores whether that shape supports frequent semantic supervision without rebuilding the coding agent itself.

The factory floor

V1 runs one coding worker at a time. By default, a real worker is a Codex App Server thread and turn launched over its JSONL protocol:

codex app-server --listen stdio:// thread/start → turn/start → turn/steer or turn/interrupt

The App Server transport keeps the active Codex thread addressable, allowing Foreman to send a supervisory update into an in-flight turn.

App Server notifications and stderr are bounded in memory, persisted as factory events, and made visible to Foreman before the worker exits. A verifier is another worker using the selected backend with an independent, deterministic verification mission.

The prior stable codex exec transport remains available through

FOREMAN_CODEX_BACKEND=exec , but it cannot accept live steering.

The worker implementation is replaceable; the runtime depends on a small worker protocol rather than Codex-specific types.

Worker backends

The semantic-supervision loop is agent-agnostic. Select the worker backend with FOREMAN_WORKER_BACKEND ( codex , the default, or opencode ):

FOREMAN_WORKER_BACKEND=opencode foreman run --repo ./my-project --job " Add request retries "

The OpenCode backend runs opencode run non-interactively and streams its output like the Codex exec backend. Live steering into an active turn is only available with the Codex App Server backend; other backends degrade to stop/retry.

See Worker backends for the Worker protocol and how to add your own.

What Foreman watches

Each observation is compact and bounded. It contains:

• 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;

• bounded repository-root AGENTS.override.md or AGENTS.md instructions when present;

• verification results and recent persisted events;

• the prior assessment and intervention;

• attempt/failure counts and elapsed factory time.

Foreman never dumps the repository into Jev. Repository instructions are read for each observation and included only in the transient Jev request; their contents are not persisted in factory state.

Defaults are a 20,000-character diff, 12,000 characters per captured output tail, 30 recent events, and 10 workers of history. The limits live in FactoryConfig and can be changed for experiments.

What Foreman assesses

The first five dimensions describe the overall job :

• implementation_complete : probability that required implementation work is complete.

• tests_sufficient : probability that relevant coverage and passing verification are sufficient.

whole, which is broader than code completion.

• requirements_satisfied : probability that the repository satisfies the free-form request as a

• needs_verification : probability that an independent verification pass is warranted.

• ready_to_finish : probability that the factory should consider the job complete.

The remaining five describe the factory floor now :

• meaningful_progress : probability that the current or latest worker is advancing the job.

• worker_stuck : probability that the worker is looping, repeatedly failing, or unable to advance.

• work_off_track : probability that work is drifting from the original job or is unrelated.

inconsistent with the target repository's root AGENTS.override.md or AGENTS.md instructions.

• agents_md_drift : probability that the worker's behavior or repository work is materially

• needs_human : probability that judgment, credentials, clarification, or permission is needed.

Every dimension is one Jev Noul question, whose result is the probability of “yes.” All ten are sent in one request. Values are validated, normalized to [0, 1] , stored in state.json , and recorded in the event timeline.

What Foreman can do

Jev only assesses. A deterministic Python policy decides which action is permitted:

• CONTINUE : let an active worker keep working.

• START_WORKER : begin a coding pass because work remains.

• START_VERIFIER : launch one independent verification pass.

• STEER_WORKER : send Jev-informed guidance into the active Codex turn.

• STOP_WORKER : gracefully terminate a stuck or off-track process.

• RETRY_WORKER : launch a fresh coding worker after a stopped attempt.

• FINISH : declare the job complete.

• ESCALATE : stop autonomous work and request human attention.

The ordering is safety-first: human need, iteration bounds, AGENTS.md drift, off-track/stuck workers, retry handling, completion, verification, then continued work.

A worker that crosses one of those drift, stuck, or off-track thresholds is steered once by default. It receives a grace period before a repeated high score causes Foreman to stop it. State tracks steering and verification so policy does not oscillate.

Transient supervisor (Jev) failures are tolerated up to FOREMAN_MAX_CONSECUTIVE_ASSESSMENT_FAILURES consecutive misses — the workers keep running while the assessment is retried — and only then does the run escalate.

Default policy thresholds are:

Decision input Threshold

needs human 0.80

off track 0.80

AGENTS.md drift 0.80

worker stuck 0.80

needs verification 0.65

implementation before verification 0.75

ready to finish 0.75

requirements satisfied 0.75

tests sufficient 0.75

Why Jev?

The integration follows TypeSafe's current official Python SDK:

• package: typesafe-sdk ;

• async client: AsyncTypeSafeClient ;

• authentication: TYPESAFE_API_KEY ;

• model: jev-latest ;

• call: await client.system_one(state=..., questions=...) ;

• question types: Noul , Choice , and Score ;

• timeout: configurable per client/call (the SDK default is 10 seconds);

exceptions;

• errors: typed API, authentication, rate-limit, connection, timeout, and response-validation

transient 5xx failures within its assessment timeout.

• retries: the SDK supports status-aware backoff and Retry-After ; Foreman retries 429 and

TypeSafe's primitives documentation says questions in a single call are evaluated independently and in parallel. Noul is the right primitive for these ten yes/no probabilities; Choice and Score remain available for future experiments.

The public docs describe HTTP 429 handling but do not publish a single numeric rate limit, so Foreman does not invent one. Its minimum assessment interval defaults to five seconds and is configurable.

Requirements

Python 3.11 or newer.

• A TypeSafe API key for real runs. The deterministic demo and tests need neither service.

Codex CLI on PATH , a version that provides codex app-server for live steering, and Codex authentication ( codex login , then verify with codex login status ).

• For the default Codex backend: the

provider and model.

• For the OpenCode backend: the OpenCode CLI on PATH with an available

Installation

From a fresh checkout:

python3.12 -m venv .venv

source .venv/bin/activate

python -m pip install --upgrade pip

python -m pip install ' .[dev] '

cp .env.example .env

Put the key in .env :

TYPESAFE_API_KEY = your-key-here

.env is ignored by Git. Foreman never writes the key into logs, observations, state, or events.

Running Foreman

foreman run \ --repo ./my-project \ --job " Add rate limiting to the API and make sure it is properly tested. "

The terminal shows worker lifecycle messages and grouped job/factory-floor assessments. It makes explicit when the coding agent is working and Foreman is independently watching, without animated noise.

Deterministic demo

The simulation exercises the same runtime, policy, persistence, event stream, and UI with deterministic model and worker implementations:

foreman demo --repo .

It needs no API key, network, coding-agent CLI, or external repository. The sequence progresses from continued implementation, through independent verification, to FINISH .

Persistence and inspection

Each repository gets local, ignored state:

.foreman/runs/<run-id>/ ├── state.json └── events.jsonl

state.json is atomically replaced and contains enough typed state to recover a run. events.jsonl is an append-only timeline. Inspect either through the CLI:

foreman runs --repo ./my-project foreman inspect < run-id > --repo ./my-project

Runtime configuration

The most useful environment overrides are:

Variable Default Meaning

FOREMAN_ASSESSMENT_MIN_INTERVAL_SECONDS 5 Debounce/coalescing floor

FOREMAN_PERIODIC_ASSESSMENT_SECONDS 30 Assessment during quiet work

FOREMAN_JEV_TIMEOUT_SECONDS 10 Semantic assessment timeout

FOREMAN_WORKER_TIMEOUT_SECONDS 3600 Per-worker timeout

FOREMAN_OVERALL_TIMEOUT_SECONDS 7200 Whole-run timeout

FOREMAN_MAX_WORKERS 3 Total workers, including verifier

FOREMAN_MAX_RETRIES 1 Fresh attempts after a stop

FOREMAN_MAX_ITERATIONS 20 Semantic decision ceiling

FOREMAN_MAX_CONSECUTIVE_ASSESSMENT_FAILURES 3 Tolerated supervisor failures before escalation

FOREMAN_CODEX_BACKEND app-server app-server for steering or exec fallback

FOREMAN_WORKER_BACKEND codex codex or opencode worker backend

FOREMAN_STEERING_ENABLED true Allow Jev-informed active-turn guidance

FOREMAN_MAX_STEERS_PER_WORKER 1 Steering attempts before stop/retry

FOREMAN_STEERING_GRACE_SECONDS 30 Time to recover before another intervention

Policy thresholds and observation bounds are typed FactoryConfig fields and can be configured by applications embedding Foreman.

Tests

python -m pytest

The suite is offline: no credentials, network, Codex process, or external repository is required.

It covers models, serialization, persistence/recovery, Jev translation and failure handling, every policy branch, subprocess streaming/termination, concurrent assessments, intervention delivery, the complete simulated factory, and a stuck-worker recovery scenario.

Process safety and security

Foreman enforces steering, worker, retry, iteration, worker-timeout, overall-timeout, and concurrency limits. Workers receive only the supplied repository as their working root.

Stop requests first interrupt the active App Server turn, then terminate the process after a bounded grace period. Ctrl-C cancels the run, terminates active workers, and persists a final cancelled state.

Workers still run with the permissions of the local environment. Codex requests its workspace-write sandbox. OpenCode runs with --auto , which approves permission requests that are not explicitly denied; Foreman does not add a sandbox around it.

Configure restrictive OpenCode permission rules before use. Foreman does not make untrusted repositories or jobs safe, so review the worker configuration and repository before running either backend.

Limitations

• Jev assessment accuracy is unproven for this use case and the semantic scores need calibration.

• False positives can stop useful workers; false negatives can allow bad work to continue.

• Repository observations are necessarily incomplete and bounded.

• The selected coding agent remains responsible for software-engineering reasoning and tool use.

• Codex App Server is currently experimental and its protocol may change between CLI releases.

• V1 runs one coding worker at a time.

• Local execution is not isolated.

• Persistence is useful for inspection, not production-grade durable execution.

correctness.

• A verifier reports evidence through the same observation channel; there is no formal proof of

• This is an architectural experiment, not a production software factory.

Future experiments

Natural next steps include simultaneous workers, per-worker and factory-wide assessments, alternate coding agents or fast decision models, dynamic assessment frequency, calibrated policies, durable execution, and isolated worker environments.

They are intentionally outside this small V1.

— 本文由 AI 根据公开来源辅助整理，命令、版本与许可证请在使用前到原始页面复核。