Enju

Coding agents produce output that looks correct until it isn’t — and without a structured checkpoint, the failure surfaces at review, not at generation. Orbit addresses this by wrapping any JSON-speaking coding agent (Claude, Codex, Cursor, or a custom CLI) in a validation loop: it selects a task from a dependency-ordered backlog, runs the agent, executes tests and lint as gates, and records the outcome in four structured artifacts before the orbit is allowed to close. Nothing advances unless the checks pass.

The artifact trail is Orbit’s sharpest differentiating feature. Every run produces an agent-result.json (structured status, changed files, raw output), an evaluation.json (rubric scoring across task focus, completion, diff signal, and validation), a review.json (accept, iterate, or stop recommendation), and a progress.md (human-readable mission log). When you are comparing two coding agents against the same task definition, you are comparing structured JSON artifacts — not reading through commit diffs and forming opinions.

Orbit fits teams who are evaluating which coding agent to standardize on, or who want deterministic gates before agent-generated code reaches a merge queue. The ‘self-healing repo’ pattern described in the docs — feed Orbit failing tests and require proof of a passing run before closing the task — is a concrete use case that works without API cost, since the replay demo runs in mock mode locally. Where Orbit does not fit: it has no scheduler, no retry loop, and no autonomous backlog driver. The harness advances one orbit at a time; the outer orchestration is yours to build.

Orbit is MIT licensed, self-hostable, and intentionally scoped small. The vendor page describes it as ‘intentionally small,’ and the contribution guidance reflects that — adapters, demo missions, and mission templates are the extension surface, not a plugin ecosystem. Teams who need to connect it to an existing CI system write an adapter; the repo structure and starter issues support that path.