foreman
Software Factory Foreman: an agent supervisor that uses Jev decisions to keep coding agents on task.
Overview
foreman (github.com/thruwire/foreman, thruwire.ai) is Josh Rosen's (@JoshARosen) native Python asyncio runtime for semantic supervision. Workers in the coding agent loop produce factory evidence; Foreman runs a parallel assess loop that feeds that evidence to TypeSafe Jev and reads calibrated scores across responsibility buckets: completion, verification, worker health, repository instruction drift, and human escalation. Routing then picks continue, steer, stop, retry, verify, or finish. The README frames Foreman as an architectural experiment, not a claim that it already beats every conventional harness.
Problem: Coding agents can churn for hours while humans guess whether work is done, tests are enough, or someone should intervene.
Built for: Teams experimenting with software factory layouts where Codex or OpenCode workers implement tickets and a separate supervisor must steer without rewriting every line.
First indexed on Jev Directory: 2026-09-17
Creator and team
- Name
- Josh Rosen
- Handle
- @JoshARosen
“Generative models work. Foreman watches the work.”
How Jev is used
- Role in the product flow
- Parallel supervision loop scoring factory responsibilities from worker evidence
- Primitives
- ScoreNoul
- State in
- Factory events from Codex or OpenCode runs: implementation progress, test output, requirement signals, and worker telemetry described in Foreman docs.
- Decision out
- Calibrated responsibility scores (for example implementation_complete, requirements_satisfied, needs_verification, worker_stuck, needs_human) that drive steering actions.
- Worker loop: reason, tool, observe while emitting factory events
- Foreman loop ingests evidence on each assess tick
- Jev evaluates configured responsibilities with typed scores
- Router maps responsibility pattern to continue, steer, stop, retry, verify, or finish
Foreman does not ask Jev to write patches. The coding agent does software engineering; Jev answers fast structured questions about whether the factory is healthy. README diagrams show responsibility groups with example probabilities (implementation_complete near 0.91 while ready_to_finish stays low until verification catches up). That is Score and Noul shaped supervision: thresholds become policy without parsing chat prose. Operators configure responsibilities and routing in repo docs (theory, runtime, steering, workers). Bring a TypeSafe key for live assess calls. Pair Foreman with ordinary agent harnesses when you want a second opinion that only watches evidence.
Sourced performance claims
- Public GitHub repo thruwire/foreman had about five hundred six stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Dual asyncio loops: coding agent plus Foreman assess
- Responsibility buckets for completion, verification, health, escalation
- Codex and OpenCode worker backends documented
- Open Python repo and thruwire.ai product site
Stack
- Python
- asyncio
- Codex
- OpenCode
- TypeSafe System One
Pricing: Open source runtime; TypeSafe usage bills per assess call configured in your factory.
Links
FAQ
- Does Foreman replace Codex or OpenCode?
- No. README positions workers as the implementers and Foreman as the watcher that scores responsibilities and steers the factory.
- Which Jev primitives appear in the loop?
- Responsibility outputs are calibrated scores over completion, verification, health, and escalation questions. Treat them as Score and Noul style gates in routing docs, not free-form chat.
- Where is the announce thread?
- Josh Rosen's X post at x.com/JoshARosen/status/2100573432089866717 introduces the factory foreman concept; deep wiring lives in github.com/thruwire/foreman docs.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.