Foreman
Supervise slower coding agents with parallel Jev assessments and deterministic intervention policy.
What Jev decides
The probability of completion, sufficient tests, satisfied requirements, useful progress, drift, a stuck worker, verification need, readiness to finish, and need for a human.
Foreman watches Codex or OpenCode workers, asks ten Noul questions about completion, tests, progress, drift, verification, and human need, then lets Python policy decide whether to continue, steer, verify, stop, retry, finish, or escalate.
- Jev role
- Continuously assess whether autonomous software work is complete, sufficiently verified, progressing, or in need of intervention.
- Input
- The job, bounded diff and output tails, repository instructions, recent factory events, and worker history.
- Output
- Ten validated Noul probabilities in a single assessment.
- Decision frequency
- On relevant factory events and periodically while a worker is active, subject to a configurable minimum interval.
- Fallback
- Let workers continue through transient assessment failures, then escalate after the configured consecutive-failure limit; deterministic policy owns all actions.
How it works
Overview
Foreman is an experimental asynchronous supervisor for software work. Worker agents edit and test the repository while a separate loop collects bounded factory evidence and asks Jev to assess the overall job and the current worker state.
All ten dimensions are independent Noul questions sent in one request. Jev only supplies probabilities; deterministic Python policy owns every intervention and applies safety-first thresholds. The repository includes an offline simulation and tests, but describes the architecture as an experiment rather than proof that it outperforms a conventional coding-agent harness.
- Language
- Python
- License
- MIT
- Author
- thruwire
Sources
Related projects
fast-jev-compaction
Prune stale coding-agent tool history without rewriting the context that remains.
What Jev decides
Whether each old tool call and its result should remain available.
Jev Review
A code-review pipeline assembled from small typed judgments instead of one long reviewer prompt.
What Jev decides
What is risky, which files and evidence matter, what could fail, how severe it is, and who should review it.
Distill
Route coding-agent work, reasoning effort, and context retention through bounded Jev decisions.
What Jev decides
Which configured model and effort should handle a call, what evidence deserves attention, whether another attempt is useful, and what context should be retained.