thirdshift-0.17.4 is not a library.

thirdshift

A factory that turns GitHub issues into ready-for-review pull requests with unattended coding agents. The humans are the day shift; the agents work the third shift while you do something else.

Use Claude Code, Codex, Antigravity CLI, Grok Build, Muse Code or OpenCode, logged in as you. Runs use the selected agent's plan or account; thirdshift itself is free and open source.

Each Ticket run implements the issue, reviews the change against the issue and your repository's standards, and keeps the pull request mergeable and passing CI. It leaves the pull request for your review, or merges it when asked. If a run fails, it reports what work was preserved and how to continue.

Read the user manual here on GitHub or on thirdshift.app. For a visual introduction, see How thirdshift works. Coding agents can start from thirdshift.app/llms.txt, a Markdown index of the manual, the site and the source.

Install

curl -LsSf https://github.com/JacobStephens2/thirdshift/releases/latest/download/thirdshift-installer.sh | sh
thirdshift setup

The installer puts a single binary in ~/.local/bin; no Rust toolchain or checkout of this repo is needed. Supported: x86_64 Linux (Rocky, Ubuntu and WSL2) and Apple Silicon macOS. See supported versions and other installation methods.

Have git, authenticated gh, and one of the supported coding agent CLIs installed. Set git's user.name and user.email and make sure it can push. Prerequisites and permissions covers setup and access: sessions run unattended as your user, with access to that account's files, credentials and tools.

Use thirdshift update to update a shell-installer copy.

Usage

From a clone of the issue's repository, on the branch you want the pull request to target:

cd ~/repos/widgets
thirdshift https://github.com/acme/widgets/issues/7

You can also use the issue number from that clone, choose a Harness, or ask for a merge:

thirdshift 7                # leave the pull request ready for review
thirdshift 7 harness codex   # use Codex instead of the default Claude Code
thirdshift merge 7           # also merge the pull request when it is ready

These examples use the default settings. Command options override your User config; --no-merge leaves a pull request for review even if your defaults request merging.

The Ticket run works in its own git worktree. Your checkout's uncommitted changes stay out of it; push any Base branch commits first. thirdshift watches CI and starts Repair sessions for merge conflicts or failing checks. Run the same issue again to continue existing work.

Use thirdshift help for the command reference, or read Ticket run behavior, command options and failure recovery.

Workflows

Start with	Command	What happens
One GitHub issue	`thirdshift 7`	A Ticket run delivers one pull request, with optional Self-merge.
An issue with sub-issues	`thirdshift 42`	A Spec run works through Tickets in dependency order, in parallel where possible, and delivers one Spec PR.
Issues marked `ready-for-agent`	`thirdshift pickup`	A Pickup run takes the lowest-numbered Ready issue.
Architecture improvements	`thirdshift architect`	An Architect run reviews the codebase, publishes a plan and builds it, or files an idea for triage.
A security audit	`thirdshift secure`	A Security run records findings privately; fixing a reproduced finding is optional.
Merged factory work	`thirdshift score`	A Scoring run samples and grades it, and proposes Bug origins; `thirdshift scores --json` lists samples, five criterion Scores, calibration and proposals by week; `thirdshift scores <pull request> --mine` records yours before revealing the model's.
One skill	`thirdshift skill <skill> [issue <issue>]`	A standalone skill runs headlessly and prints its final message.
A Custom process	`thirdshift process <name> [issue <issue>]`	An experimental Custom process runs skill, command or catalogue Steps from `~/.thirdshift/processes/<name>.toml`, in named throwaway worktrees when the file declares them. The catalogue Step `run-published-issue` starts a Ticket run or a Spec run on the issue the session Step before it published. An optional Step can fail without stopping the process. `issue <issue>` gives every session Step that issue. One Issue-branch worktree plus a `delivery` Step publishes a ready-for-review pull request from that Issue branch and then removes the Claim. A `self-merge` Step directly after it merges when `merge` or the User config asks, and a file can only tighten limits. With no issue, and no Issue-branch worktree, it is a Pass and skips while another Pass on the repository is running. Status, the Console and the Activity log list it as Process `<name>`. A failed one is in Needs you, and it sends a Run notification under the same email rules as a Ticket run. A skipped Process sends none. Each Step may name its Harness, Model and Effort, otherwise inheriting the Command's.

A Custom process can use implement → delivery on an Issue-branch worktree for the Ticket run's implementation and bug diagnosis. An unreproduced bug ends successfully with needs-info, skipping later Steps and cleaning up the Claim and worktree.

Each Architect, Pickup or Security run makes one pass. Use cron or another operating system scheduler to repeat it. See scheduling, Architect runs every few minutes and Security runs on a schedule.

Monitor and manage work

From another terminal, inspect this machine's work across repositories:

thirdshift status
thirdshift show 'widgets#7'
thirdshift logs 'widgets#7' --follow

status groups Commands by what needs you, what is ready for review, what is landing and what is still working. show includes stages and child Ticket runs. Session traces retain agent events and timed work; you can inspect the span tree or export a Command's whole Trace.

For a live browser view of the same Inbox, with stages, logs and Session traces:

thirdshift console --no-sign-in

Open the localhost address it prints. The Console never starts work. Use thirdshift stop <command> to stop a live Command and wait for cleanup; add --requeue to return its issue to Pickup. A signed-in Console account can do the same.

Console on a server

The Console and status show the serving OS account's Commands by default. To add readable Commands from other local accounts, use thirdshift console other-accounts readable (or a comma-separated list of account names). Other-account Commands show their owner and are read-only. See account visibility for filesystem access requirements.

The Console can run as a persistent service behind HTTPS and sign-in. See server installation for configuration, passwords, the generated Linux installer and Caddy, Apache or nginx setup.

Documentation

Configuration, Harness, Model and Effort, and email notifications.
Machine status, showing Commands and reading logs, and stopping a Command.
Console installation and Console Themes and browser tests.
Host resources: thirdshift resources --json and the Console's Resources page show RAM, CPU load and configurable limits that delay new sessions.
Building and testing and releasing.
Exact prompts and Factory skills, also readable on the site.
Built-in processes are compiled into the binary. thirdshift process list lists Processes and marks Built-in processes; thirdshift process show <name> prints a Process file exactly so you can copy it under a new Custom process name.
Domain vocabulary and architectural decisions.
Open issues for known gaps and planned work.

Credits and license

Released under the MIT License. Factory skills include work by Matt Pocock, Dex Horthy and Cloudflare. See full credits and license notices.

These section links keep older bookmarks useful. The complete explanations now live in the user manual.

thirdshift 0.17.4