by mosonlab
You write the specs. It clears the board: coding agent chains plan, review, implement, verify and merge unattended on your own machine, on the Codex and Claude subscriptions you already have.
# Add to your Claude Code skills
git clone https://github.com/mosonlab/annealanneal is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by mosonlab. You write the specs. It clears the board: coding agent chains plan, review, implement, verify and merge unattended on your own machine, on the Codex and Claude subscriptions you already have. It has 52 GitHub stars.
anneal's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/mosonlab/anneal" and add it to your Claude Code skills directory (see the Installation section above).
anneal is primarily written in TypeScript. It is open-source under mosonlab on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh anneal against similar tools.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
You write the specs. It clears the board.
Anneal runs chains of coding agents on your own machine. Queue tasks on the board, and each one is planned, reviewed, implemented, verified and merged unattended, on the Codex and Claude subscriptions you are already signed in to.
Install · How it works · Status · 简体中文
Live capture: Anneal clearing its own task board. This repository is built by Anneal itself.
You do one thing: write specs. You queue them on the board, Anneal breaks them down and orchestrates the work, and you walk away. A chain picks up each task and takes it the whole way through plan, plan review, implementation, two independent code reviews, fix application, regression verification, and the merge itself. When you come back, the board is clear: you read the pull requests that matter, open an agent window on the ones you care about, and iterate until you are satisfied.
In between, nothing needs you. A chain only stops when a human decision is needed: an agent asks a question through the web Inbox, a step you marked as gated waits for your call, or a run escalates. Once you answer in the Inbox, the board moves on.
The pull requests in this repository are specified, planned, reviewed,
implemented and merged by Anneal's own chains, running on one local
machine.
Chain-delivered commits carry Co-Authored-By: Anneal Chain and
X-Anneal-Run / X-Anneal-Step trailers, so you can check in the git
log which commits the chains produced.
A chain instantiates a template of steps. Each step binds an agent role: a prompt, a model, a reasoning effort and the runner CLI it executes on. The flagship Full Assurance template covers delivery in twelve steps.
The template is data, not code. Roles, prompts, models and gates are all editable, so you can define your own agent roles and reshape the chain into a workflow for your own team of agents.
Agents: a role, a prompt, a model and effort, and the runner it goes to.
| # | Step | Agent role | What it does | Runner | Model · effort |
|---|---|---|---|---|---|
| 1 | Write a spec | spec |
Turns the task into the specification of record | Claude | Claude Opus 5 · high |
| 2 | Plan | plan |
Cuts the spec into parallel vertical tracer-bullet slices | Claude | Claude Fable 5 · medium |
| 3 | Plan review | review-coordinator |
Reviews every slice against the spec and the frozen base | Pi | GPT-5.6 Sol · xhigh |
| 4 | Revise plan | plan-reviser |
Edits the slice set against the findings, in a fresh session | Codex | GPT-5.6 Sol · high |
| 5 | Implementation | implementation-plan-executioner |
Executes the slice set from the live dependency frontier and opens the pull request | Codex | GPT-5.6 Sol · high, with GPT-5.6 Luna · max subagents |
| 6 | Code review | review-coordinator-sol |
Reviews the integrated diff at the pinned base and head | Pi | GPT-5.6 Sol · xhigh |
| 7 | Blind code review | review-coordinator-opus |
Reviews the same diff again, blind to step 6's findings | Claude | Claude Opus 5 · high |
| 8 | Apply review fixes | senior-dev |
Dispositions every finding from both reviews and applies the adopted ones | Codex | GPT-5.6 Sol · high |
| 9 | Documentation | librarian |
Updates internal documentation to match the delivered code | Pi | GPT-5.6 Luna · xhigh |
| 10 | Regression verification | regression-verifier |
Refreshes onto the target branch and reruns the regressions | Codex | GPT-5.6 Luna · xhigh |
| 11 | Merge readiness | — | Recomputes the head, requires every open review to clear, emits an exact-head authorization | — | mechanical, no model run |
| 12 | Merge execution | merge-integrator |
Re-verifies every precondition against the live pull request, then merges | — | mechanical, no model run |
A finished chain: twelve steps done, including four autonomous merge-tail repairs during regression verification.
The design rule behind the steps: whoever wrote the code never judges
it. Authorship, review and verification run in separate sessions, the
two code reviews cannot see each other's findings, and step 8
adjudicates both. Each session starts clean, with a purpose-built
prompt and an environment the runner constructs itself: your global
agent config and skills never leak in as noise. Role bindings live in
agents/templates/compound-engineer-workflow/;
board column semantics in the
task-routing contract.
You need:
^20.19.0 || ^22.13.0 || >=24; use 22.17.0 from
.nvmrc, with npm 10.9.2+git clone https://github.com/mosonlab/anneal.git
cd anneal
git checkout v0.6.0
npm ci
npm run setup:local
npm run build
docker compose up -d --wait --wait-timeout 60 postgres
npm run db:migrate:release -- --fresh
Then start npm run dev:api, npm run dev:runner and npm run dev:web
in three terminals, in that order, and open http://127.0.0.1:5173.
The full sequence with its preflights is in
docs/release/developer-preview.md.
Developer Preview 6 (v0.6.0): interfaces and stored data shapes may change between previews, and the only upgrade path is a fresh install. macOS on Apple Silicon only.
Read before pointing this at anything you care about: Anneal
launches coding CLIs with non-interactive permission bypass, as your
macOS user, outside a sandbox. Use a disposable repository and a
machine you are willing to let an agent modify. Details in
docs/release/security.md.
Provider CLIs, their authentication and their plan terms stay between
you and the provider; the authoritative support statement is
docs/release/support-matrix.md.
Architecture · Install · Security · Migration and recovery · Release notes · Contributing · Support
Five skills from
mattpocock/skills (MIT,
Copyright (c) 2026 Matt Pocock) supply working text for the chain's
prompts; see THIRD_PARTY_NOTICES.md. This
snapshot is licensed under the MIT License, with its
boundary defined by public-snapshot.json.
Community link: LINUX DO — sincere, friendly, united, and professional.