Self-healing repos
Give Orbit failing tests or lint issues and require proof before a task is marked complete.
Mission control for AI coding agents.
Orbit turns agent work into bounded, validated, auditable loops: one task, real checks, durable artifacts, and a progress trail humans can review.
Bounded tasks
Dependency-aware backlog selection keeps each orbit focused.
Validation gates
Tests, lint, and type checks decide whether work advances.
Review artifacts
Every run leaves inspectable JSON and progress evidence.
Agent-neutral
Use Claude, Codex, Cursor, or any JSON-speaking CLI.
Run a local orbit in minutes.
The deterministic replay demo needs no API key. It shows Orbit selecting a task, running an agent path, validating the result, and recording the evidence.
git clone https://github.com/human-again/orbit
cd orbit
python -m venv .venv
.venv/bin/pip install pytest pillow
MOCK=1 ./replay.sh auth-rescueIf the agent cannot prove it, the orbit does not close.
Orbit records the evidence that matters: what the agent returned, what validation proved, how the run scored, and what changed next.
agent-result.json
Structured status, changed files, notes, and raw agent output.
evaluation.json
Rubric scoring for task focus, completion, diff signal, and validation.
review.json
Accept, iterate, or stop recommendation for the completed orbit.
progress.md
Human-readable mission log with validation status and next steps.
Built for the messy middle of agentic development.
Backlog execution
Break work into dependency-ordered tasks and let the harness advance one verified orbit at a time.
Adapter experiments
Swap coding agents behind the same contract and compare artifacts instead of anecdotes.
Contribute an adapter, demo, or mission template.
Orbit is intentionally small. The best contributions make the harness easier to verify, easier to replay, or easier to connect to another coding-agent workflow.