Are you running a self-hosted queue of autonomous Claude Code sessions?
retained text
[](https://www.reddit.com/r/BuildWithClaude/?f=flair_name%3A%22Project%22)Hi everybody, I'm writing this post to search for fellow developers who are building something similar. For a few months I’ve been building a personal “dev pipeline” around Claude Code, and I’d like to compare notes with anyone doing something similar. What it does today: * **Queue of work entries:** each entry is a self-contained prompt with an estimate, a priority, dependencies, the resources it occupies (file paths or folders, so non-overlapping entries run in parallel) and a machine-checkable acceptance command. * **Worker:** a daemon picks ready entries, opens a git worktree per entry, launches a Claude Code session in tmux, and relaunches it from a saved hand-off prompt when context or limits run out. * **Session boundary:** the session only commits. The worker rebases, runs the gates (tests on the touched slices, a secret scan, a “ratchet” on bundle size / SQL query counts / complexity) and pushes by fast-forward. No PRs. * **Release:** automatic deploy to a small home server, decided by a post-release probe. No human approval. * **Dashboard:** shows the queue, the worker loop, per-entry token spend and a progress diary. An MCP server lets me add, read and reorder entries from the Claude mobile app. * **Machine-wide rules, skills and hooks:** everything is kept at a bare minimum and loaded accordingly to single task needs (realized huge tokens savings here). **Where it needs improvements** Gates vs autonomy: strict gates (the ratchet especially) kept refusing pushes and burning whole sessions on relaunches. I’m now moving the ratchet to “advisory, auto-tighten on improvement, open a follow-up entry on regression”. State consistency: worker, DB, API and dashboard sometimes disagree about an entry’s state (probably some unchecked race conditions here) **What I’m looking for** People running unattended or long-running Claude Code (or other agent) loops on their own repos: how do you queue, gate, retry and stop? Tools I might be reinventing: orchestrators, agent task queues, merge queues for agents, cost tracking, ratchet tools like Betterer or ESLint bulk suppressions. Happy to share more detail on any part (worker design, entry schema, gate setup). Not selling anything, it’s a personal project and I'm writing cause I've already improved it a lot and I'm not finding any new approach or best practices worth adding. Thanks
document fetched 2026-10-31 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xjb3j0/are_you_running_a_selfhosted_queue_of_autonomous