wastech-orchestrator

Turns a written task into a pull request for your review.

  • Open source
  • Actively developed
  • Apache License 2.0

The agents do the editing, and the orchestrator runs the flow and all the Git work around them, from the branch to the pull request. Each step is checkpointed in a local SQLite database, so every run leaves a record you can trace afterwards.

View repositoryRead the docs

Why it’s here

This is the agent pipeline I work with, and the one I would set up in your team’s repository. Its code and docs are public, so you can read how it works before deciding anything.

Problem

An agent’s edit needs several steps around it before a reviewer can follow the change. When nobody owns those steps, an agent’s pull request is hard to review and its result is hard to repeat. It is one of the reasons agent work stalls, next to the quality of the context the agent reads.

Solution

The orchestrator takes the process and the Git work on itself, and the agents are not asked to commit or push. A flow describes the steps, and tasks are written against it. The pull request opens once the checks you configured and the agent review have passed, with a written summary of what changed.

  • A task goes to an agent and comes back as a pull request once the checks you configured for your repository have passed.
  • The steps are written down as a flow, so every run follows the stages you defined for it.
  • An optional audit commit puts the task file and its summary next to the code.

Key features

  • From task to pull requestA Markdown task becomes a branch, an edit, your repository’s checks and a pull request with a written summary.
  • Two agents, one interfaceOpenAI Codex and Anthropic Claude Code are interchangeable, and with both set up, a run falls back to the other on an infrastructure failure.
  • Git stays with the orchestratorAgents are not asked to commit or push. With strict isolation on and a host that can sandbox them, they cannot write to .git, and preflight warns when that protection is missing.
  • Resumable runsEvery step is checkpointed, so a restart resumes the in-flight task without committing twice or reopening a pull request.

Pipeline

  1. TaskA Markdown file in your repositoryYou
  2. BranchCreated from your base branchOrchestrator
  3. Flow stepsPlanning and editing, one step at a timeCoding agents
  4. ChecksYour own test and lint commands, after every editOrchestrator
  5. ReviewReads the change and sends findings back for fixesCoding agents
  6. CommitA scoped commit, then the pushOrchestrator
  7. Pull requestOpened with a summary for your reviewOrchestrator

Who it is for

  • People and teams who want coding agents working in their own repository.
  • Tech leads who want every agent change to go through review and leave a record.

Stack

Core
  • Python 3.12+
Agents
  • OpenAI Codex CLI
  • Anthropic Claude Code CLI
State
  • SQLite (checkpointed state, audit tables)
Git
  • git
  • GitHub CLI (gh)

Install

pipx install "git+https://github.com/VladimirMakarevich/wastech-orchestrator.git"cd path/to/your-repoworc install .worc preflight
Requires Python 3.12+, git and the codex or claude CLI on your PATH. The GitHub CLI is only needed if pull requests should open automatically. The install wizard sets up the repository, and preflight checks the agents and the isolation policy without running a task.

Limitations

  • Runs one task at a time, end to end.
  • Opens pull requests on GitHub only, through the gh CLI.
  • The checks step gates a task only after you list your test and lint commands in the config; a fresh install lists none.
  • A fresh install leaves strict isolation off, and a host without an OS sandbox, such as native Windows, cannot block writes to .git.

Want your team working this way?

Write to me