// the find
context-machine-lab/sleepless-agent
🤖 24/7 AI agent that maximizes Claude Code Pro usage via Slack. Auto-processes tasks, manages isolated workspaces, creates Git commits/PRs, and optimizes day/night usage thresholds.
Sleepless Agent is a Python daemon that wraps Claude Code CLI to run tasks continuously via a Slack interface, targeting developers with a Claude Code Pro subscription who want to queue work and come back to finished PRs. It uses SQLite for task state, isolated directories per task, and Git automation to commit results and open pull requests. Essentially a lightweight personal CI that uses Claude as the worker.
The workspace isolation design is solid — each task gets its own directory with access only to its workspace and a shared folder, which prevents tasks from stomping on each other. Making Slack optional and keeping the CLI as a first-class interface was the right call; it's usable without configuring a Slack app. The three-agent pipeline (planner → worker → evaluator) is a reasonable pattern for catching bad output before it lands in a PR. SQLite for task persistence is appropriately boring — no Postgres dependency just to track a queue.
The usage threshold logic (96% at night, 95% during the day) is nearly meaningless — a 1% difference won't save your quota from a runaway task, and the agent has no way to estimate token cost before starting. The auto-task-generation feature is a footgun: an idle agent inventing its own tasks and auto-committing to your repo is a bad default for anyone running this against a real project. There's no sandboxing beyond directory isolation — Claude Code can still run arbitrary shell commands, install packages, or make network requests from within a task workspace, which matters if you're pointing it at code that calls external services. At v0.1.0 the error recovery story is thin; if the Claude Code CLI process dies mid-task, the task stays in_progress until manual intervention.