kohaku4yz/pipeline-bus — explained in plain English
Analysis updated 2026-05-18
Automate a small dev workflow where one AI model implements a ticket and another reviews it before a human merges.
Coordinate two machines running different AI models without setting up any message queue or webhook infrastructure.
Use git history as a built in audit trail for every task, review comment, and status change.
| kohaku4yz/pipeline-bus | abhayparolkar/zboy | dmaynard51/ultima1-ios | |
|---|---|---|---|
| Stars | 6 | 6 | 6 |
| Language | Shell | Shell | Shell |
| Setup difficulty | moderate | hard | moderate |
| Complexity | 3/5 | 4/5 | 3/5 |
| Audience | ops devops | developer | general |
Figures from each repo's GitHub metadata at analysis time.
Requires cron access on both an implementer and reviewer machine, each with its own AI coding agent installed.
pipeline-bus is a git only continuous integration and delivery pipeline that connects separate machines and AI coding agents into a shared, automated workflow without any dedicated backend server. An owner writes a task file describing what needs to be built along with acceptance criteria that can be checked by running a command, then pushes it to this repository. A cron job on an implementation machine, running a cheaper, high throughput AI model, notices the new task, claims it, writes the code into its own working branch, and pushes that branch back. A separate cron job on a review machine, running a mid tier model, then reviews the change against the acceptance criteria and leaves comments. The human owner does the final review and merges the finished work into the real target repository. The only communication layer in the whole system is GitHub itself: there is no webhook, no message queue, and no process talking directly to another process. Each side simply polls its own state and pushes commits, and git's own atomic push and commit history double as both the coordination mechanism and the audit trail. The review step only ever returns one of two verdicts, approved or needs changes, so the state machine that drives the whole pipeline can be a single line of logic, and any needs changes verdict must include specific direction on what to change so tasks do not stall on vague feedback. To limit runaway costs, a task gets at most two rounds of review before it is marked stuck and a human is notified to look at it directly, rather than letting the two models argue back and forth indefinitely. The project estimates a typical task costs well under a dollar in total model usage across both rounds. It is intended for small, well scoped tasks such as one patch, one document, or one small tool, not for large open ended engineering work. The project is written mainly in Shell script, is released under the MIT license, and is not meant to be opened up as a public marketplace for untrusted task submitters, since anyone who can push to the repository can effectively make the implementation agent run arbitrary commands.
A pure git based CI/CD pipeline that uses GitHub itself as the message bus to route tasks between an AI implementer, an AI reviewer, and a human owner.
Mainly Shell. The stack also includes Shell, Git, Cron.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice (MIT license).
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly ops devops.
This repo across BitVibe Labs
double-check against the repo, no cap.