Shiptrail

Shiptrail runs a configurable pipeline of Claude Code, Codex, and OpenCode agents against your repository. Each job gets its own branch and worktree. The engine stores job state, streams logs to a local dashboard, and either merges a verified result or pauses with a clear request for you.

Requirements

• Python 3.11+ and uv.
• Git. A new folder is set up by shiptrail init; an existing repository can use any branch.
• Claude Code, Codex, or OpenCode CLI, signed in. Install the client you select in the pipeline.
• Optional: agent-browser for browser checks: npm install -g agent-browser && agent-browser install. shiptrail doctor reports it but does not require it.

Install

Install the engine once per machine:

uv tool install shiptrail

Release candidate, from TestPyPI, until the stable release is published:

uv tool install --default-index https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ shiptrail

If uv is not installed, install.py is an optional bootstrap. It prints uv's official installation instructions and asks before running the official installer, then installs the engine:

python install.py

Pass --target path/to/your-project to also run shiptrail init in that folder.

Quick start

Start from any folder, empty or not:

cd path/to/your-project
shiptrail init
shiptrail doctor
shiptrail start "build a todo app"

shiptrail init looks for Claude Code, Codex and OpenCode on your machine, shows the plan, and asks once before it changes anything. With one usable tool it uses that tool without asking. With several it asks which one the pipeline should start on (Enter takes the first):

Found Claude Code, Codex and OpenCode.
Which tool should the pipeline use by default?
  1) Claude Code
  2) Codex
  3) OpenCode
  4) Mixed: the template's choice for each stage
You can change any stage later on the Pipelines page. [1]:
Plan for C:\path\to\your-project:
  would create: git repository on main
  would write: AGENTS.md (starter; edit the TODO lines)
  would write: CLAUDE.md (imports AGENTS.md for Claude Code)
  would commit: AGENTS.md, CLAUDE.md as the first commit on main
  would write: shiptrail/config.json
  would write: shiptrail/backlog.md
  would write: .claude/skills/shiptrail/SKILL.md
  would write: .agents/skills/shiptrail/SKILL.md
  would write: .opencode/skills/shiptrail/SKILL.md
  would update: .gitignore
Continue? [Y/n]

The answer sets the tool every stage starts on; each stage can use a different tool afterwards, on the Pipelines page. "Mixed" keeps the template's tool for each stage and moves a stage to your first tool when its own is not installed. init installs the in-chat shortcut (/shiptrail or $shiptrail) for every tool it found. If none is installed it stops and says how to install one; --provider claude|codex|opencode|both skips detection, for scripts.

What init does:

• No git repository: creates one on main and makes a first commit with the starter instructions file. If git has no name or email configured, it stops and prints the two git config commands to run.
• Existing repository: uses its current branch as the base branch and records it in shiptrail/config.json. Nothing is renamed.
• No AGENTS.md or CLAUDE.md: writes a starter with TODO lines. With Claude Code alone that is CLAUDE.md. When other tools share the project it is AGENTS.md, plus a CLAUDE.md that imports it, so every tool reads the same instructions. Edit them to describe your install, dev server and checks. If the folder has a package.json with dev, test, build or lint scripts, those commands are filled in.

Then commit the shiptrail files (shiptrail/, the skill wrapper and .gitignore) before the first job. The engine never pushes to a remote.

Run shiptrail --help for the commands, or shiptrail <command> --help for one of them.

Use it

Start work with the skill command from your client:

/shiptrail Add a dark mode toggle to settings     (Claude Code)
$shiptrail Add a dark mode toggle to settings     (Codex)

Other commands:

You can also add work to shiptrail/backlog.md or the dashboard backlog, then start it with shiptrail next.

Define your pipeline

Edit the pipelines in shiptrail/config.json or use Dashboard Settings. Full runs a spec, builder, parallel reviewers, and approval. Fast uses the shorter implementation and approval flow. Configure roles, providers, model and effort per execution, retry limits, and merge policy. Jobs retain a snapshot of their pipeline and role instructions, so later settings changes affect new jobs only.

In the pipeline editor, Role, Model and Effort are dropdowns. Roles are the built-in ones plus any in shiptrail/roles/. Models and efforts are read from your installed tools, codex debug models and opencode models, which only list metadata and never call a model. The answers are cached for ten minutes. Claude Code has no such command, so it offers its aliases. If a tool is missing or does not answer, the dropdown falls back to built-in choices and says why. "Custom..." accepts any other value. A saved model that is not in a tool's list is kept and flagged.

When a job needs you

Jobs pause when an agent is blocked, a review requests changes beyond the retry limit, the approver escalates a product or risk decision, or the merge policy requires your review. shiptrail status and the dashboard show the reason and the reports to inspect. Resolve the request in the branch, then resume or rework the job. The engine never pushes to a remote.

The dashboard

Run shiptrail dashboard to open the dashboard for this repository, or shiptrail serve --open to start the service and open it. By default the URL is http://127.0.0.1:5099/. The dashboard shows jobs across registered repositories, pipeline steps, reports, timeline, logs and usage. It also supports backlog management, repository and machine settings, and job actions such as stop, resume, approve, rework and delete. The service listens on localhost only.

What a job leaves behind

The repository keeps its configuration and backlog in shiptrail/. shiptrail/backlog.md keeps its plain line format; backlog priority and status are committed beside it in the sidecar shiptrail/backlog.meta.json, which may be deleted safely (items fall back to normal priority and ready status). Job state, reports, attempt logs and archived outputs live under shiptrail/jobs/<job-id>/; worktrees are kept beside the repository until merged or deleted. The exact report set depends on the pipeline. The machine registry and service state live in the user data directory (or SHIPTRAIL_HOME).

Customising

Built-in roles ship with the engine. To customize a role for one repository, run shiptrail eject <role> and edit the resulting file in shiptrail/roles/. The dashboard Settings page edits pipeline configuration; machine settings control provider executable paths (cli.claude, cli.codex, cli.opencode), concurrency and available ports. Ejected roles remain repository-local and can be compared with the built-in version using shiptrail eject <role> --diff.

Working on this repo

Set up the development environment with uv sync --locked. The Python suite runs with uv run --locked pytest, or in parallel with uv run --locked pytest -n 4 (about a minute, against about six serially). Tests that touch the fixed sandbox port range are kept on one worker.

The dashboard lives in web/: run npm ci, then npm run build to rebuild the bundle the engine serves, and npm run typecheck, npm run lint, npm test and npm run e2e to check it.

License

See LICENSE.