Short loops. Real progress.
Not sprint planning. Not Jira. Sprint is a CLI for running AI work in focused loops.
sprint loop # Runs all tasks, one clean commit each
AI coding agents (Claude Code, Codex) are powerful but chaotic:
- They wander off-topic
- Progress is hard to track
- Git history becomes a mess
- You lose time re-explaining context
sprint gives you:
# TASKS.md
### T1: Set up project @codex
**Commit:** `chore: initialize project`
### T2: Add user authentication @claude
**Depends:** T1
**Commit:** `feat: add auth`
### T3: Write tests @codex
**Depends:** T2
**Commit:** `test: add auth tests`Then:
sprint loopEach task → right agent → clean commit → next task. Walk away and come back to a working project with reviewable history.
# Install
git clone https://github.com/vonwao/sprint-cli.git ~/.sprint-cli
export PATH="$HOME/.sprint-cli/src:$PATH"
# Try the demo
git clone https://github.com/vonwao/sprint-cli-demo.git
cd sprint-cli-demo
sprint loop### T1: Task title @agent
**Depends:** T0 # Won't run until T0 is done
**Artifacts:** src/foo.ts # Expected output (informational)
**Commit:** `feat: foo` # Commit message
Task description. The agent sees this as its prompt.Agents:
@codex— Fast, good for well-defined tasks@claude— Better reasoning, good for complex tasks
TASKS.md → sprint picks ready task
→ launches assigned agent
→ agent does work
→ auto-commits with clean message
.agent/state.json → tracks what's done
LOG.md → append-only history
State lives in files. Restarts are safe. Git tracks everything.
Inspired by the Ralph Wiggum pattern:
sprint loop # Run until queue empty
sprint loop 5 # Max 5 iterations
sprint loop --push # Push after each commitKick it off and walk away. Come back to clean commits.
"Trello, but executable."
- More structured than pure agent loops
- Less overhead than Jira
- Git as the source of truth
Perfect when you want AI to do the work but stay in control of what gets built and in what order.
MIT