anneal README

The project's README from github.com/mosonlab/anneal. As one markdown file: README.md.

Anneal

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.

status license platform node

Install · How it works · Status · 简体中文

Multiple tasks running in parallel across the board

Live capture: Anneal clearing its own task board. This repository is built by Anneal itself.

The workflow it is built for

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.

What you get

Anneal is built with Anneal

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.

How it works

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 view: each agent's role, model, reasoning effort and runner

Agents: a role, a prompt, a model and effort, and the runner it goes to.

The twelve steps in full: role, runner, model and effort for each

An agent's title names the role only — Planner, Code Reviewer, Senior Dev — while its slug spells out role-model-effort, so two agents that do the same job on different models stay distinguishable at a glance.

# Step Agent role What it does Runner Model · effort
1 Write a spec spec-opus-high Turns the task into the specification of record Claude Claude Opus 5 · high
2 Plan plan-fable-medium Cuts the spec into parallel vertical tracer-bullet slices Claude Claude Fable 5 · medium
3 Plan review review-coordinator-astra-medium Reviews every slice against the spec and the frozen base Codex GPT-6 Astra · medium
4 Revise plan plan-reviser-opus-medium Edits the slice set against the findings, in a fresh session Claude Claude Opus 5 · medium
5 Implementation plan-executor-astra-low Executes the slice set from the live dependency frontier and opens the pull request Codex GPT-6 Astra · low, with GPT-5.6 Luna · max subagents
6 Code review code-reviewer-sol-high Reviews the integrated diff at the pinned base and head Codex GPT-5.6 Sol · high
7 Blind code review code-reviewer-opus-medium Reviews the same diff again, blind to step 6's findings Claude Claude Opus 5 · medium
8 Apply review fixes senior-dev-astra-low Dispositions every finding from both reviews and applies the adopted ones Codex GPT-6 Astra · low
9 Documentation librarian-luna-xhigh Updates internal documentation to match the delivered code Codex GPT-5.6 Luna · xhigh
10 Regression verification regression-verifier-luna-max Refreshes onto the target branch and reruns the regressions Codex GPT-5.6 Luna · max
11 Merge readiness — Recomputes the head, requires head-bound regression PASS evidence and a server-side ancestry check, 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
Task detail: a completed twelve-step chain with each step's role and status, the run cost, and the merge-tail repair timeline

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.

Quick start

The verified release path needs:

macOS on Intel is expected to work but is not yet release-verified; Windows is unsupported.

git clone https://github.com/mosonlab/anneal.git
cd anneal
git checkout v0.8.0
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.

Status

Developer Preview 8 (v0.8.0): interfaces and stored data shapes may change between previews, and the only upgrade path is a fresh install. The verified targets are macOS on Apple Silicon and Linux (Ubuntu 24.04 LTS, x86_64). macOS on Intel is expected to work but is not yet release-verified; Windows is unsupported.

For the pull-request workflow, follow Add a project.

Read before pointing this at anything you care about: Anneal launches coding CLIs with non-interactive permission bypass, as your own user account, 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.

Documentation

Architecture · Install · Security · Migration and recovery · Release notes · Contributing · Support

Credits and license

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.

anneal (mosonlab/anneal)

Searching Database…

185 ·GitHub
Anneal
What is this?Anneal is a local orchestration system that runs chains of coding agents on your own machine: you write a spec, queue it on a board, and it is planned, reviewed, implemented, verified and merged unattended using your own Codex and Claude Code subscriptions.
Is this a CLI tool or a local service with a UI?It's a local multi-process service: an API, a runner and a web app you start separately, with a browser board where you queue tasks and a web Inbox for the questions agents escalate to you.
How do I get it running on my machine?Clone the repo and run npm run dev:api, npm run dev:runner and npm run dev:web in three separate terminals, in that order, then open the web board.
Do I need API keys for Codex and Claude, or do I just sign in?No API keys needed: Anneal launches the Codex CLI, Claude Code and Pi you already have installed and signed in to on your subscription, and holds no credentials or proxy of its own. It also needs Node (^20.19/^22.13/>=24), Docker Compose and Git.
Is this stable enough to point at a project I care about?Not yet — it's Developer Preview 8, interfaces and stored data can change between previews with only a fresh-install upgrade path, and it runs CLIs with permission bypass outside a sandbox, so the README recommends a disposable repo and a machine you're fine letting an agent modify.
Does it run on Windows, or only macOS and Linux?Only Apple Silicon Macs and Linux (Ubuntu 24.04 LTS) are release-verified; Intel Macs are expected to work but unverified, and Windows is unsupported.
Is it safe to let this touch my real codebase?The README explicitly warns it is not, for now: agents run with non-interactive permission bypass under your own user account with no sandbox, so a test-only repo is recommended.
What does the Full Assurance template actually do?It's a 12-step chain covering spec writing, planning, plan review, revision, implementation, two blind code reviews, fix application, documentation, regression verification, and a final mechanical merge-readiness and merge-execution step.
Why are there two separate code reviews?Anneal's design rule is that whoever wrote the code never judges it: the two code reviews run blind to each other's findings, and a later step adjudicates both sets of findings before applying fixes.
Can I customize the agent roles and chain steps?Yes, the template is data, not code: roles, prompts, models and gates live in editable files under agents/templates/, so you can define your own agent roles and reshape the workflow.
What happens when a chain can't proceed without me?It stops and waits: an agent can ask a question through the web Inbox, a step you marked as gated pauses for your decision, or a run escalates, and the chain resumes once you answer.
What are similar projects?orca, gastown, wicked-studio, oh-my-codex, cc-wf-studio, buzz, claude-code, oh-my-subagents
Which topics is it in?code-review, gpt, workflow-automation, agent-orchestration, git, docker-compose, multi-agent, local-first, claude, codex, agent, claude-code
Creator
mosonlab
Size
30.5 MB
Stars
185
Updated
CommitSep 26
Age
1mo
Commits · 90 Days