Forge is a Claude Code + Codex workflow control plane.
Prompt packs tell agents what they should do. Forge checks what agents are actually doing through Claude Code hooks, Codex-side installed assets, routing logs, health checks, and smoke tests.
This repository is the extracted, public-friendly source layout for a local Forge installation. It currently targets Claude Code and Codex on Windows/PowerShell first.
- A single
forgeCLI plus Claude Code slash-command assets and Codex-installed workflow assets. - A Forge skill that explains routing, levels, risk, and workflow rules.
- Runtime hooks for write-time guardrails and session/stop audits.
- PowerShell health and smoke checks for docs, workspace manifest, M1 compliance, live route freshness, and session state.
- Optional adapter points for external workflow tools such as Superpowers, GSD, Task Master, gstack, GitNexus, UI review, and lint tools.
Most daily use should fit three commands:
forge doctor
forge workflows
forge task new -Title "Describe the task"
forge verifyUse them as:
- doctor: check whether Forge can safely run in this repo.
- workflows: show whether BMAD, Superpowers, gstack, Compound Engineering, and GSD are active, staged, vendor-only, or missing.
- task new: create a tracked task artifact before agent work.
- verify: produce a fast local release-readiness verdict before PR or merge.
Forge keeps deeper adapter, baseline, and smoke details behind those commands. If a check fails, the verdict should name the failing layer and the next fix command.
forge verify defaults to a local Lite health check plus minimal smoke so it stays fast and does not depend on upstream network access. Use forge verify -Full before final release, forge smoke -Quick for extended local smoke, and forge version -FixDrift to refresh a source-linked install when forge_source_drift=true.
Forge treats workflow kits as routed capabilities, not always-on prompt bulk. BMAD is the planning source, Superpowers is the execution discipline layer, gstack is a manual gate, and Compound Engineering/GSD stay manual-approval unless explicitly enabled for learnings or state handoff.
commands/ Claude Code slash-command markdown files
skills/forge/ Forge skill
docs/ Protocol, boundary, and schema documents
scripts/ PowerShell health, smoke, routing, and reset helpers
hooks/ Claude Code hook entrypoints; Codex consumes the shared scripts/docs/skills without Claude hooks
examples/ Example project settings and manifest snapshots
From this repository root:
pwsh -NoLogo -NoProfile -ExecutionPolicy Bypass -File .\scripts\Install-ForgeLocal.ps1 -RepoPath "<repo>"The installer copies commands, skill, docs, scripts, and hooks into the current user's .claude and .codex Forge roots, and installs project hooks under the target project's .claude/hooks directory when applicable. It also writes forge.cmd to %USERPROFILE%\.local\bin, so forge doctor, forge task new, and forge verify work once that directory is on PATH.
After installation:
forge doctor -RepoPath "<repo>"For deeper validation:
forge verify -RepoPath "<repo>" -Full- Contribution guide: CONTRIBUTING.md
- Architecture boundaries: docs/architecture/boundaries.md
- Release checklist: docs/release-checklist.md
- License: MIT
- Minimal example: examples/minimal-project
Forge is not a replacement for Markdown-first workflow kits such as flow-kit. It is the runtime enforcement layer that can sit beside them:
- Markdown artifacts define intent, scope, tasks, and review expectations.
- Forge hooks and checks enforce selected rules at runtime.
- Optional adapters can call external tools when installed and fall back to built-in checks when missing.
- Windows/PowerShell paths are still first-class; cross-platform packaging is not done.
- Some hooks remain Claude Code-specific by design; shared scripts/docs/skills are installed for both Claude Code and Codex.
- Public adapter contracts are not fully formalized yet.
- Existing local installations may need session state and live route coverage refreshed before all health checks pass.