How RADAR works.

01 The loop

One loop for every task.

Every task goes around the same loop, and each step runs only if the step before it passed.

Open

A task is opened with its goal and a re-runnable acceptance command. The command is checked when the task is opened, without being run: a criterion that cannot run as written is refused then, not discovered at the end.

Lease

One owner holds a task at a time, through a lease and a fence token. A write from a stale owner is refused.

Work

The agent works in its own git worktree, on its own branch.

Claim

The agent's report that it is done is recorded as a claim. A claim does not close the task.

Verify

RADAR runs the acceptance command itself, on the current revision of the work. The agent's own report is read, but it is not proof.

Close or reopen

A pass lets the task close, with the verify record kept as its proof. A failure reopens the task with the failure attached.

Limits

Three worker attempts, then three takeovers by another agent. After that the ledger refuses more attempts and the task waits for you. Switching model, session or machine does not reset the count, and a retry has to say what changed.

02 The ledger

One record every agent writes to.

Every agent reads and writes the same task ledger: a local SQLite file. It is the single record of who holds each task, what was claimed and what the test showed.

Tasks

The goal and the acceptance command, written before the work starts.

Ownership

The current owner of each task, with its lease and fence token.

Attempts

Counted per task, across agents, models and machines.

Claims

An agent's report that the work is done, kept apart from results.

Verify records

Each acceptance run, bound to the commit and the working tree it ran on.

Changes

A changed acceptance command is recorded with the old and the new version; an override is recorded too.

03 Isolation

Each attempt in its own copy.

• Each attempt works in its own git worktree, on its own branch. Two workers never share a tree.
• A worker cannot rewrite its own acceptance command. Only the manager can change it, and only before the task closes; every change keeps the old and the new version.
• Work reaches your main branch only after verify passes, and the acceptance command runs once more after the merge.
• Force-pushes are refused for every role.
• These rules run as code, before and after a model is called. They are not a prompt, so an agent cannot talk its way past them.

04 Agents

The agents you already use.

RADAR runs Claude Code, Codex CLI, OMP and Hermes Agent, on your computer or on Linux servers of your own. It does not replace them: they still do the work.

One contract

The same rules, the same ledger, the same verify gate and the same attempt limits apply to all four.

Manager

The agent you talk to manages: it opens tasks, hands them to workers and closes a task only with a passing verify record.

Workers

Workers do the work in their own worktrees. They cannot close tasks and cannot change their acceptance command.

05 Local first

Local first.

RADAR needs no account and no cloud.

On your machine

The ledger, the memory notes, the worktrees and every verify run, on your computer or on Linux servers of your own.

Needs a network

Your agents, which talk to their own providers as they do today. Every other outside service is optional and off until you turn it on.

Requirements

Windows or Linux (macOS is not supported yet), Python 3.12 or newer, and Git. RADAR uses the Python standard library only.

Cost

No bill of its own. The agents keep using the accounts you sign them in with. A paid API is used only if you name the provider and a budget.

Install

Give your coding agent the repository link and say “install this”. It checks your computer, asks before installing anything missing, installs RADAR into a virtual environment and reports what radar init and radar doctor say.

06 Commands

A short list of commands.

The commands you will use first. Placeholders are in angle brackets.

07 What it does not do

What RADAR does not do.

• It does not replace your agents. They still write the code; RADAR gives them one record, one check they cannot skip and one memory.
• It does not write your tests. A check is only as good as its test; RADAR makes the test explicit before the work starts, so a person can read it.
• It does not close work that has no test. A task without a runnable acceptance command is refused when it is opened.
• It does not keep your agents offline. They still talk to their own providers, as they do today.
• The open edition is for individual developers, not teams; a company edition is in development.

08 Status