dsh-orbit

Orbit addresses a recurring pain in long AI-driven sessions: when a model freely loops through multi-step work, it can drift off-plan, spin in circles, or lose its place mid-run. Orbit fixes this by splitting the task into a Commander (planning, step review, and final acceptance), an Executor (one step at a time with real tool calls and verification), and a Smart Watchdog (anomaly diagnosis and recovery). Flow control stays with the supervisor; the model never decides what happens next on its own.

Every run is durable and bounded. State is persisted to disk for cold recovery. Loop budget, correction rounds, and Watchdog diagnoses all carry explicit ceilings. Once a run starts, the assigned models and reasoning levels are frozen, so mid-run UI changes only affect the next run. A workspace mutation fence ensures only one legal Orbit driver at a time, Commander and Watchdog receive read-only tool allowlists, and the Executor unlocks only the tools its current step explicitly declares. For high-uncertainty steps, an optional MoA mode dispatches two to four independent candidates, a Judge selects the best, and Orbit deterministically promotes the winner before the Executor runs real tests and the Commander gives final sign-off. SINGLE mode works out of the box; MoA and Browser are opt-in add-ons.

Best suited for DeepSeek Harness users who run multi-step engineering tasks and want auditability, cold-resume capability, and hard boundaries around model autonomy. Install Orbit alone for full SINGLE-mode capability, then layer in MoA or Browser only when a task calls for multi-candidate competition or real browser interaction.