🦄 Forge - a local-first, multi-agent software-engineering runtime that runs Claude Code / Codex-style agentic workflows entirely on your own machine via Ollama, llama.cpp, vLLM, and LM Studio (cloud models optional). Ships 18 sandboxed tools, 6 model providers, and a full REPL + UI dashboard in a single Node CLI - no telemetry, no lock-in.
Polishes .beads/beads.jsonl in isolated context. Runs the full polish loop (self-containment, dependencies, test obligations, coverage vs plan, duplicates, Forge invariants) and returns a focused report. Use when you want to burn a polish round without filling the main conversation.
Reviews Forge TypeScript changes for correctness, security, sandbox/permission coverage, state-machine integrity, and test adequacy. Invoke for any non-trivial diff before merge.
Audits Forge documentation for drift against the code. Checks that README counts, ARCHITECTURE hot paths, SETUP env vars, and INSTALL instructions still match reality. Read-only.
Reads multiple competing markdown plans from .flywheel/plans/ and synthesizes a best-of-all-worlds hybrid. Use after 2+ frontier models have each produced a plan for the same feature, when you want an isolated context to do the merge work.
Meta-skill agent. Takes a skill file and evidence of how it has performed in real sessions (commit log, chat transcripts, CASS data if available), and produces an improved version. Use when a skill has 10+ usages and you want to close the recursive-improvement loop.
Runs the Forge test suite and analyses failures. Isolated context so full vitest output does not pollute the main conversation. Returns a focused summary.