Claude Sprint Agent

Claude Sprint

A wrapper script that runs an unattended Claude Code "sprint" in headless mode, monitors the output for a `DONE` signal, and automatically tears down the screen session when work is complete.

Built for the workflow: *work with Claude Code interactively, then hand off a scoped set of tasks for Claude to complete autonomously while you're away.*

---

What it does

1. Runs claude in headless print mode (-p) with a scoped permission set tuned for dev work.
2. Optionally resumes a prior Claude Code session so Claude keeps the full context of what you were working on.
3. Streams all output to a timestamped, project-tagged log file (~/claude-sprint---.log) so every run is preserved, with ~/claude-sprint-.log as a symlink to the most recent run on that project.
4. Watches the log for DONE or NO TASKS — Claude's signals for "batch done" and "backlog empty", respectively.
5. Writes a timestamped status file alongside the log, also symlinked at ~/claude-sprint-.status.
6. Uses per-project screen session names (claude-sprint-) so multiple sprints can run in parallel without collision.
7. Kills the wrapping screen session automatically once the sprint ends.

---

Requirements

• Claude Code installed and authenticated (claude auth status should return logged-in).
• GNU screen (sudo apt install screen on Debian/Ubuntu).
• bash — the script uses bash-specific features (PIPESTATUS, [[ ... ]]).
• A project directory where Claude has prior session history if you plan to resume (-c / --resume).
• The tools in the allowlist installed where relevant to your project (git, docker, language runtimes, etc.). Tools not installed just won't get used — they won't break the script.

---

Installation

1. Download the script and make it executable (wget saves files without the execute bit, so chmod +x is required):

   wget https://raw.