Skip to content
LutherwavesPublic

About

Operational tools for running AI coding agents — usage-limit projection, session/context load, token-efficiency watchers

Resources

Code of conduct

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

builder

Release Validate License: MIT Conventional Commits

builder makes your coding agent answer to you. Every piece of work ends with a short report: is it done, what you have to do, what isn't proven yet, what left your machine, and every call the agent made on its own, with the reason. Nothing leaves the machine (no push, no PR, no message) without your yes.

It's a plugin for Claude Code, Codex CLI, Cursor CLI, Grok Build and opencode. It takes a request through spec, plan, build, review and ship, with security in every stage. Skills and one session hook; no server, no account, no network call at session start.

▶ Needs you: a customer who moved can now change their email at checkout, in one step.

▶ You (1): push feat/change-email and open the PR. I ask below.

⚠ Not proven: the confirmation email in a real inbox.

✓ Proven: go test ./... → 14 passed (example).

↗ Outside: nothing sent.

Calls: the old address gets a notice, because a stolen account changes the email first.

The marks mean one thing each: ✓ done or proven, ✗ failed, ▶ you act, ⚠ not proven, ↗ left this machine. After the report the agent asks each ▶ You item as a choice, its recommendation first.

Why builder

builder is one framework from a request to a release. Five stages hand the work on, and each stage stops at a gate before the next one starts.

Stage Skill You get Gate before the next stage
Spec spec A spec: the person, the moment, the non-goals and when it is done You approve the spec
Plan plan Tasks that each start with a failing test and end in one commit Each done-when line of the spec maps to a task
Build build The code, one task at a time, test first Every test passes, with the output quoted
Review review Each way the change hurts a person, proven, with a fix and a test No open Critical or Important finding
Ship ship A pull request, then a merged and checked release Your yes before anything leaves the machine

Each stage ends by loading the next one, so you do not name the skills yourself. On Claude Code they are /builder:spec, /builder:plan, /builder:build, /builder:review and /builder:ship. A new session finds where the work stands from the spec, the plan, the task commits and the saved review.

A small change skips the spec and the plan. It still gets a failing test or a named check, a review and the report.

Security is a step of each stage, not a pass someone must remember. The spec names each sensitive surface the work touches. A task on such a surface writes its hostile-input test first. The review runs a security pass by itself, and ship stops on an open check. security audits the whole repository and remembers what it covered.

Two more things keep the work honest. When something fails and the cause is not known, debug finds the root cause and pins it with a test before any fix. Before the last test run, build removes what its own change did not need: a comment that restates the code, a guard for a case that cannot happen, a helper with one caller. review lists the same things under "Slop".

builder your own CLAUDE.md
Starts from the person and the non-goals every piece of work if you wrote it
Each solo call recorded with its reason every report, under Calls if you wrote it
Unproven work marked as unproven every report if you wrote it
Security in each stage, not on request spec, plan, build, review and ship if you wrote it

Install

builder runs on macOS and Linux. It needs Bash, and it offers to install one small helper binary in your first session.

Claude Code

/plugin marketplace add Lutherwaves/builder
/plugin install builder@builder

Codex CLI. Trust the plugin's hook when Codex asks.

codex plugin marketplace add Lutherwaves/builder
codex plugin add builder@builder

Grok Build. Grok Build reads Claude Code plugins. Add this repository as a marketplace in its /plugins screen, and install builder.

Cursor CLI and opencode: see docs/install.md.

The first session asks to install the helper binary. It runs the session-start hook and the security score. Answer yes, not now (quiet for 7 days) or never. The installer downloads the binary of your plugin's version from builder's GitHub release, checks the checksum and, with a signed-in gh, the build provenance. It installs to ~/.config/builder/bin. To install it yourself, run bash <plugin root>/scripts/install.sh.

Harness support

A harness counts as supported once every eval fixture passes on it.

Harness Version Eval date Status
Claude Code 2.1.308 (Claude Code) 2026-10-22 experimental: spec 2/6; 11 not graded

Skills

Skill What you get
builder The thirteen tenets, the five stages and their gates, and the report.
spec A request becomes a spec. It reads the repository, asks one question at a time with its recommendation, shows two or three ways to build it, and waits for your approval.
plan An approved spec becomes tasks that each start with a failing test, reviewed as an engineer and as a designer. Work on an API, a service boundary, a data store or an agent tool also gets an architecture review.
build The plan runs one task at a time, test first, one commit per task. It stops when a step is wrong and proposes the change. It removes what its own change did not need. It answers each review comment with one outcome and one reply.
review A review that proves each harm with evidence, a fix and a test, with a security pass and the running product. Posts nothing without your yes.
ship Reviewed work becomes a checked release. It stops on an open finding and waits for your yes before each push, pull request, merge, tag and release.
security A whole-repository audit: every layer, standards-mapped findings with CVSS scores, and what changed since last time.
debug A failure becomes a root cause in one sentence, a test that pins it, and a fix at the cause. It stops after three fixes that did not hold.

Ask for a skill by name, or just describe the job. The agent loads the right one.

Works with

builder runs on its own. When the remote is on GitHub, ship checks that gh is present and signed in. It also checks that commit signing is on. When a check fails, ship stops and names the fix.

  • gh: ship opens and merges the pull request through it.

Contributing

Every change lands through a pull request that passes bash scripts/validate.sh. See CONTRIBUTING.md, and SECURITY.md to report a vulnerability.

Credits

builder takes ideas, and no text, from four MIT-licensed projects: obra/superpowers (the spec, plan and build loop), garrytan/gstack (the reviews of a plan, the use of the running product, shipping), cloudflare/security-audit-skill (the audit method) and PostHog/posthog, MIT outside its ee/ folder (the bar a review finding must pass, the last sweep, the handling of review comments).

License

MIT. The ideas builder learned from other projects, and their licenses, are in Credits.

About

Operational tools for running AI coding agents — usage-limit projection, session/context load, token-efficiency watchers

Resources

Code of conduct

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages