Decide whether you need a supervisor
Foreman is worth a bounded trial if your coding agent already does the engineering work and you want another process to decide when it should continue, receive guidance, stop or seek verification. Keep a person available to inspect those decisions. The project says assessment accuracy is unproven and its semantic scores need calibration.
Foreman supervises one coding worker at a time. Its Python policy selects actions from model-informed proposals; the assessment model does not choose individual files or tools. This separation gives operators a place to configure intervention rules, but it does not establish that the model judges completion correctly.
What our sandbox established
Our run on October 4, 2026 tested commit e5d1aa4. Installation and build succeeded in our sandbox; the test command exited 1. The harness records command success, exit status, elapsed time, timeout status and parsed test counts separately. A successful command therefore describes that step, not an authenticated deployment or completed software job.
The stored parser summaries report 259 passed, 1 failed and 260 total. They record no skipped tests or errors. These are stored parser summaries, not verified exhaustive counts of unique tests. We have not rerun the project or reconciled a complete log for this correction.
The raw failure identifies tests/test_integration.py::test_noisy_events_are_coalesced. Its assertion expected 50 persisted output events and found 24. The excerpt uses a fake model and worker. It establishes that assertion failure in our sandbox, without establishing whether scheduling, implementation or the expectation caused it. The raw pytest footer reports 40.71 seconds, while the harness records 47 seconds for the test command; these describe different timing scopes.
Our run did not test authenticated Codex, OpenCode or Hermes supervision, live steering against a real CLI, assessment calibration or completion of an external repository job. The pinned README describes the suite as offline. Its deterministic demo is a useful next step, but the supplied record does not establish that we separately ran the demo.
The audit record contains zero reported advisories from pip-audit. Scan coverage is unknown because the record does not supply scannedPackages; its packages field is zero. This is not evidence that every installed package was checked or that Foreman is secure.
The MrKeyoor testing methodology explains the distinction between command outcomes and workload evidence. That page supplies provenance context; its current procedure does not establish which methodology version governed our October 4 run.
Choose the backend for the intervention you need
At the tested commit, Foreman supported Codex, OpenCode and Hermes workers, with Jev supplying assessments. Codex App Server kept an active turn addressable for steering or interruption. The Codex exec transport could not accept live steering, and the other backends fell back to stop and retry. If guidance during a running turn is your reason to adopt Foreman, backend choice matters before installation.
The supplied current README describes additional capabilities: an OpenAI Decisions API adapter, Deep Agents Code workers and attached supervision for supported sessions including Pi and Pi Durable. Those are current documentation claims, not capabilities exercised by our older run. The 0.4.1 release page also identifies a different release commit, so the release tag and our tested checkout should not be treated as identical.
Set permissions before autonomous work
Foreman does not isolate local execution. Workers retain local-environment permissions; Codex requests its workspace-write sandbox, while OpenCode runs with --auto subject to its permission configuration. Configured evidence commands also inherit local permissions and Foreman's environment, although Foreman invokes them without a shell and bounds their time and retained output.
Review the repository, worker permissions and trusted evidence commands before starting a real job. Use a task whose changes you can inspect and whose stopped state you can recover deliberately. Persistence supports inspection, but the README explicitly excludes production-grade durable execution. The supplied evidence does not independently verify the older page's issue-specific account of escalation recovery, so this review makes no claim about that issue's current status.
Alternatives answer different questions
Aider is a terminal pair programmer for starting a project or editing an existing codebase. Choose it when you want to work directly with a coding assistant rather than add Foreman's supervisory layer. This is a choice of workflow, not evidence that Aider implements Foreman's policy or is a documented Foreman backend.
LangGraph is a low-level framework for building long-running, stateful agents. Choose it when you intend to build the workflow and its state transitions yourself. Its documentation describes durable execution and human oversight, but adopting a framework entails implementing your own application; it is not a drop-in Foreman replacement.
Your existing coding-agent CLI is a complement to Foreman: it remains the worker. Neither alternative comparison establishes compatibility with Foreman or measures which approach produces better code.
How the ratings should be read
These editorial ratings use a 1 to 5 rubric: 1 means substantial documented obstacles, 2 means major prerequisites or gaps, 3 means usable with material limits, 4 means strong supporting evidence with limited gaps, and 5 means unusually complete evidence for that dimension. They are judgments about the supplied evidence, not benchmarks or adoption measurements.
Setup ease is 3 because installation and build succeeded, while real use still requires credentials and a configured worker. Documentation is 5 for its detailed treatment of runtime, evidence, policy, backends and execution limits; this does not rate the truth of model judgments. Community is 3 on the narrow evidence of a published release and contribution guidance; contributor breadth, response times and adoption remain unknown. Maturity is 3 because the project has offline tests and explicit guardrails, while our test command failed and the README excludes calibrated accuracy, isolation and production-grade durability.

