MemPalace

Most coding agent runs end the same way: the agent reports success, the diff looks plausible, and you spend the next hour figuring out what actually changed. Orbit is a self-hosted, MIT-licensed harness that turns agent runs into bounded, auditable loops. It selects a single task from a dependency-ordered backlog, routes it to your agent via a JSON-speaking CLI contract, runs your existing tests, lint, and type checks, and only closes the loop if validation passes. Four artifacts survive every run: a structured result file capturing what the agent returned, an evaluation scoring task focus and diff signal, a review recommendation, and a human-readable progress log.

The differentiating feature is the validation gate. The vendor states the design principle directly: if the agent cannot prove it, the orbit does not close. That means failing tests are not a soft signal — they are a hard stop. For teams running self-healing workflows where failing tests are the baseline requirement, this is the mechanism that makes agent output trustworthy rather than aspirational.

Orbit fits teams that already have strong test suites and want to run agents against a backlog without losing auditability. It also fits teams evaluating multiple coding agents — swap the adapter, run the same task, compare the JSON artifacts instead of comparing demos. Where it breaks: repositories with thin or no test coverage get no benefit from the validation layer, because there is nothing to enforce. The harness is also intentionally narrow — it does not manage CI pipelines, handle pull request creation, or coordinate agents working in parallel. Teams who need those layers are building them on top of Orbit or reaching for a broader platform.

The repo ships a deterministic replay demo that requires no API key, which the docs describe as running with a single shell command after a standard Python venv setup. The adapter contract is JSON-based, so any CLI-driven coding agent can be connected without modifying the core harness.