ReleaseWitness

The practical path

From a light goal
to your first handover.

ReleaseWitness is easiest when every runtime sees the same repository. This guide makes the first ten minutes concrete.

Codex, Claude Code and Antigravity sharing one ReleaseWitness assignment and report protocol.
Different runtimes can work together because the assignment, scope and report contract stay provider-neutral.
Before you begin

You need Python 3.11+, a prepared project repository and at least one real worker runtime with local file and shell access.

Quickstart

Place the ReleaseWitness control plane in the target repository, then initialize and validate it:

relwit init
relwit validate
python examples/multi-agent-demo/run_demo.py

Runtime roster

Register each real session with a unique id and a non-overlapping source scope. Codex, Claude Code and Antigravity are execution surfaces; supervisor, worker and reviewer are workflow roles.

relwit agent register
--id claude-web --role worker
--scope src/frontend --scope tests/frontend
--capability web --max-active 1

First cycle

  1. Prompt the supervisor.

    Give it the goal, the real worker ids and the constraints. It will create the roadmap and scoped tasks.

  2. Dispatch and pull.

    Run relwit supervisor cycle. Each worker opens its generated outbox prompt and runs worker pull --agent <id>.

  3. Report evidence.

    The worker changes only its claimed scope, runs requested checks and submits task report.

  4. Return to the supervisor.

    Run the next cycle. It ingests reports, runs QA and records the next action in work/SUPERVISOR_REPORT.md.

Automatic worker intake

If your runtime has a local CLI or adapter, configure one argv runner once and let a bounded worker run pull the assignment, invoke it and verify the report.

relwit agent register
--id codex-api --role worker
--scope src/backend
--runner-arg=python
--runner-arg=tools/codex_worker_adapter.py
--runner-arg=--assignment
--runner-arg={assignment_path}
relwit worker run --agent codex-api --max-tasks 1

The default runner uses argv without a shell, stores only bounded sanitized evidence plus a local diagnostic spool reference, and turns a missing report into a failed report. An optional preflight must return an explicit ready state before pull; manual worker pull remains the default for runtimes without an adapter.

Hướng dẫn nhanh bằng tiếng Việt

Đặt control plane trong cùng repository với application. Đăng ký từng runtime thật bằng id duy nhất, giao scope không chồng lấn, rồi gửi cho supervisor một goal ngắn. Prompt đầy đủ cho worker nằm trong work/outbox/; worker chỉ cần pull, sửa trong scope và report bằng CLI.

Continue to operations