← Frank Zhang

Open-source project · Agent runtime

Ambu

A small command-line supervisor for coding agents. It checks the work with your verification command and stops when the same failure repeats.

01

What it is

Ambu (agentctl) is a small command-line tool for bounded, check-driven repair with OpenCode. It runs a verification command that you supply, limits repair attempts, stops on repeated failures, and records each run.

02

The problem

The README names two problems with autonomous coding agents in a repository.

Runaway loops: an agent that hits an environment bug or an impossible constraint retries and produces the same error again.

Missing verification: an agent reports a task as done without running the project's tests or verification commands.

03

How it works

  1. RouteRoute the request.agentctl do classifies the intent of a prompt and routes it to a station such as oracle, fixer or coder. An ambiguous prompt falls back to read-only.
  2. DispatchDispatch the task.Ambu wraps opencode run and sends the task to the agent.
  3. CheckRun the declared check.It runs the verification command you declared. The supervisor records the checks. A worker's own claim cannot pass them.
  4. CompareCompare failures.It normalizes the failure output, removing timings, timestamps, memory addresses and process IDs, then hashes it. If the identical blocker recurs, it halts early.
  5. GateGate high-risk actions.High-risk actions such as git_push, deploy and secrets are blocked unless they are passed with --approve.

04

Evidence and limits

05