Graph Foreman

graph-foreman

**A foreman for your task graph**: executes an approved plan as a DAG — dispatching parallel subagent executors, refusing to sign off any task a fresh reviewer has not inspected, and watching the whole site on a live dashboard.

An agent skill for [Claude Code](https://claude.com/claude-code) and compatible agents, installable via [skills.sh](https://skills.sh/):

npx skills add JrSantiaggo/graph-foreman


What it does

• Parallel executors, capped — independent tasks run at once; up to 3 executors out of 4 total slots, so a finished task never waits for capacity to be reviewed.
• Author ≠ verifier, enforced — every task is validated by a FRESH reviewer agent that never saw the code being written. done refuses a self-reported validation; review refuses the task's own author. Evidence is recorded per verdict.
• Collision-proof plans — init rejects dependency cycles (naming the loop) and two parallel tasks that declare overlapping touches paths, before any agent starts.
• Live dashboard — a read-only server on :4949 renders the DAG as phase swimlanes with animated dep edges, per-task state, retries and an event log. Killing it never affects a run.
• Zero dependencies — three files, Node 18+, no npm install. State is plain JSON + append-only NDJSON under .specs/graph/ (gitignored scratch).
• Tokens are spent only by agents — the engine and the dashboard are plain Node processes making no model calls. Executors and reviewers (subagents) are the cost; the orchestrator adds a small constant overhead; watching the dashboard costs nothing.

What you need before running it

The skill executes, it never plans:

1. An approved plan in the engine's format (.specs/graph/plans/.plan.json — tasks, deps, vali