@izkac/forgekit

A disciplined development workflow for AI coding agents — portable across Claude Code, Cursor, Codex, Copilot, Gemini, Windsurf, and opencode.

AI agents are fast but undisciplined: they claim success on stubs, skip verification, over-build, and quietly leave the human behind. Forgekit installs a gated workflow that holds an agent to the standards a senior engineer would — and gives you, the operator, the visibility and control to stay in charge of a whole fleet of them.

One package, three command-line tools:

| Bin | What it does | |-----|--------------| | forgekit | Install / update the Forge skills into any supported agent | | forge | Run the Forge workflow — sessions, integrity gates, fleet control | | review | Standalone multi-pass thorough code review |

Why Forge (the workflow)

Forge moves every substantial change through explicit phases — triage → brainstorm → plan → implement → verify → review → done — and enforces them with executable gates, not honor-system promises:

• Operator brief — before any code is written, the agent produces a plain-language HTML explainer of what's about to be built (with diagrams). It opens in your browser and gates implementation: you approve what you actually understand, not a wall of spec markdown.
• Runtime spine — every change declares how each capability wires up: library → runtime owner → what it writes → the evidence it works. No orphaned code that's "done" but never called.
• Executable product loop — for jobs, workers, and pipelines the end-to-end acceptance is a command that must exit green, not a paragraph claiming it would.
• Registered deferrals — "wire it up later" is only allowed when tracked; the finish gate refuses unresolved ones.
• Session scorecard — every session leaves a graded, measurable trail (A–F).

The result: agents that can't fake completion, and a workflow you can trust to run with less babysitting.

Fleet control — command every session from one terminal

Running agents across three IDEs and two terminals? forge fleet is a single control surface over all of them, across every project and engine:

forge fleet list       # every session: phase, task progress, engine, activity
forge fleet watch      # live-refreshing dashboard (active sessions only; --all shows everything)
forge fleet view <s>   # detail + live transcript tail (Claude Code)
forge fleet send <s> "pause and report"   # message any session; --all broadcasts

Sessions auto-register the moment they touch Forge — including ones running inside Claude Desktop or Cursor. You see where everything stands at a glance and steer any of them without switching windows.

Portable — one workflow, every agent

The same skills and the same discipline, installed into each agent's native skills directory:

Claude Code · Cursor · Codex CLI · GitHub Copilot · Gemini CLI · Windsurf · opencode

Planning is engine-flexible too: use the OpenSpec CLI if you have it, or the built-in specs engine (plain markdown, same layout) if you don't. Optional ADRs archive each finished change into an Architecture Decision Record.

Install

npm i -g @izkac/forgekit
forgekit install            # interactive: pick skills + agents (space toggles, `a` = all)

Non-interactive:

forgekit install --skills forge,thorough-code-review --agents claude,cursor --force

Then wire a project and start working:

cd your-project
forge init --claude --cursor      # slash commands, rules, hooks, .forge/

# in the agent chat:
/forge add a health endpoint that returns { ok: true }

The agent triages, brainstorms with you, plans a tracked change, writes the operator brief for your approval, then implements task-by-task with TDD and per-task review — refusing to call itself done until the gates are green.

Docs

• How to use Forgekit — full tutorial: install, project wiring, worked examples, integrity gates, fleet, cheat sheet
• Full Forge reference — phases, pace matrix, integrity rules (also installed at ~/.claude/skills/forge/docs/forge.md)
• Repository

License

MIT. The Forge skill and several of its workflows are adapted from Superpowers (MIT).