prekuter/dryforge
Bounded-autonomy plugin harness for agents. Intent to implementation: ready, then go.
Dryforge – Turning your intent into working code
What it is – Dryforge is a lightweight plugin that sits on top of existing AI‑code‑generation agents (Claude Code, Codex, Grok Build, GitHub Copilot, Antigravity). It lets you express what you want to build, then lets the agent work autonomously inside that boundary. The tool records the intent, verifies the result, and writes plain‑Markdown documentation that stays in your repository.
Why it exists – Most current “agent” tools force you to fit your work into pre‑written workflows, test‑first rituals, or swarms of sub‑agents. When a new model arrives those scripts break, and the user ends up supervising every step or letting the agent guess silently. Dryforge flips the relationship: your intent decides the method, not the other way around.
Key ideas
- Bounded autonomy – The agent can do anything it needs, but it cannot change anything that belongs to you without an explicit approval.
- Authority model – Every decision is owned either by you (the human) or the agent. The tool surfaces only the decisions that belong to you.
- Intent‑first loop – Two commands drive the workflow:
ready– ingest any description (one‑line idea, spec, PDF, existing docs), ask only the questions that require human judgment, and write a draft intent under.dryforge/.go– execute the intent, run real verification (tests, checks, builds), and report the outcome. If the work would diverge from the approved intent, it pauses for you.
- Migration – One‑time command that reads an existing codebase, extracts what can be inferred, and asks you to clarify the missing business rules, security policies, etc., turning a legacy repo into a Dryforge‑managed project.
- Plain‑text docs – All intent and decisions are stored as ordinary Markdown files (
AGENTS.md,docs/*.md). They remain usable even after the plugin is removed, and any future agent can read them.
Supported agents – Claude Code, Codex, Grok Build, GitHub Copilot CLI, Antigravity CLI. Adding a new model only requires packaging the same “skill source”.
Installation – Each supported platform provides a tiny marketplace command, e.g.:
/plugin marketplace add prekuter/dryforge # Claude Code
codex plugin marketplace add prekuter/dryforge # Codex
Updates are handled through the same marketplace commands; auto‑update can be toggled per platform.
Typical workflow
- Start –
ready <your‑idea>(or$readyin Codex). Dryforge reads your repo, asks you any missing questions, and writes.dryforge/intent.md. - Approve – Review the generated intent and confirm.
- Build – Run
go. The agent implements the intent, runs real checks, and either merges the changes (new project) or creates a branch and asks how you want to integrate. - Iterate – Repeat
ready→gofor new features or changes. The intent files are updated, keeping a living record of why the code is the way it is.
Requirements – Git must be installed. Before go runs the working tree must be clean (no uncommitted or untracked files outside .dryforge/), and the main branch must be up‑to‑date with its remote.
Community & support – Discord server, GitHub issues, and a small website (dryforge.dev). The project is Apache‑2.0 licensed.
TL;DR – Dryforge is a plugin that lets you give AI code‑generation agents a clear, documented intent, runs the agent autonomously within that intent, and keeps the decisions in plain Markdown so the codebase stays human‑controlled and model‑agnostic.
Related
- Project
- Project
- Project
- Project
- Project