DevChestra

Conduct your coding agents like an orchestra.

DevChestra is a local workflow builder for headless coding agents. Draw a workflow on the canvas or have your coding agent create one through the CLI, then watch Claude, Codex, OpenCode, or Pi do the work — with branches, loops, and human check-ins where you want them.

It's not another agent to babysit. It's the stage your agents play on.

Above: a real DevChestra workflow that routes a PR review by risk, with a separate follow-up path. Runs, transcripts, and replay live in the same window.

Why it's nice

Pasting the same instructions into chat every time or using 100 diffrent SKILL.md and hoping the agent does it the same way? That's the problem. Chat is vibes: same prompt, different tools, diffrent model, different order, different day.

DevChestra makes the boring parts deterministic. You build the process once — who runs, what tools and MCPs they get, what they're allowed to touch, what they must hand back, what happens next — and it runs that way every time. The agents still improvise inside their box (they're LLMs, after all), but the orchestration around them doesn't.

Research, then rank. Draft, then review. Write code, run tests, fix what broke. No babysitting prompts, no pasting output from one chat into the next.

What can you build?

We built a PR review workflow that routes changes by risk and sends follow-up work down its own path. That's one way to put the pieces together. A few others:

• Bug triage to PR — reproduce the issue, investigate, implement a fix, run tests, and ask for a review before opening a PR.
• Ticket to PR — turn a Linear or GitHub ticket into a plan, implementation, tests, and a PR, with a human check-in along the way.
• User flow testing — walk through a product flow, check the UI, capture what broke, and send failures back for a fix and another test.
• Lead research — find potential leads, check them against your criteria, and hand a structured list to a person for review.
• Content pipeline — research a topic, draft, edit, check against a brief, and get approval before publishing.

The workflow decides the order, handoffs, branches, and review points. The agents do the work inside each step, using the tools and access you give them.

A starter workflow: Fix failing tests

Every new project is seeded with a small workflow you can run right away, "Fix failing tests":

1. Run tests with your project's test command (detected from package.json, Cargo.toml, go.mod or pyproject.toml).
2. Tests green? If they already pass, the run ends as "Already green".
3. Otherwise a loop hands the failing output to an agent, which makes the smallest fix it can, then reruns the tests. It repeats until they pass, for up to 3 rounds. If they still fail, the run ends as "Still failing" and keeps the last test output.
4. Keep these changes? A person reviews the result. Approve and the run ends as "Fixed". Reject and it ends as "Declined".

It's a good one to open first: it uses a shell step, an agent, a gate, a loop and an approval, so you can see how the pieces connect. It's best run in a fresh worktree so a rejected fix is easy to throw away.

Beyond code: an AI UGC video pipeline

Workflows aren't limited to code. This one turns product images into a finished UGC-style short and posts it:

1. Start with the product images as a parameter.
2. Write the script. An agent drafts the UGC script, broken into shots.
3. Cast influencers. An agent proposes a few influencer options that fit the script, and Pick best influencer chooses one.
4. Generate shots. A for-each step goes through the script's shots one at a time. Each round runs an agent connected to the Higgsfield MCP to generate that shot's video.
5. Edit with HyperFrames. An agent assembles the shots, adds subtitles and finishes the cut.
6. Review video. A reviewer agent checks the result against the script. If it finds defects, a Fix pass agent repairs them. If it finds none, the video moves straight on.
7. Post to social. A join merges the "fixed" and "clean" paths, and a shell step posts the video.

The agents do the creative work. The workflow decides the order, which tools each one gets, and when a defect sends the video back for another pass.

What you build with

The canvas and CLI use the same workflow vocabulary:

Agent — run an agent step. Pick your CLI and model, write the prompt, done.

Shell — run a command. Tests, linters, scripts, whatever you'd run yourself.

Gate — route by condition. First matching branch wins, otherwise falls through.

Join — merge branches back together. Wait for the first one, or all of them.

Approval — pause for a person. Nothing moves until someone says go.

Decision — ask a decision model when you need a judgment call, not just code.

Loop — repeat until a condition is met. Each round keeps its own history.

For each — process every item in a list, one at a time.

Run workflow — call a reusable workflow and wait for it to finish.

End — finish the run.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗