starlight

Motita is an AI agent that requires a deterministic check to declare a task complete. The agent splits every task into three layers: a reasoning engine that proposes actions, a sandbox that runs those actions under real limits, and an anchor that runs a user-defined check. The model can propose a PASS, but only the anchor can declare one. If the check fails, the agent receives the failing command, its real output, and the structured verdict, then retries. If it cannot pass, it escalates and says so, never faking a PASS. The agent is written in pure Go with zero external dependencies and is distributed as a single static binary. It supports multiple model providers, including OpenAI, Anthropic, Gemini, Copilot, Qwen, and Ollama, and can run on low-resource hardware. The anchor by default detects the project's existing gate from files like Makefile, go test, package.json, cargo test, or pytest, or it can run a custom command and check exit code and output pattern. The sandbox runs actions with its own HOME, temporary files, and installed toolchains kept outside the repository, and it reports what it could not apply. The agent counts tokens per provider and per agent, and can split work across background agents on separate git branches. It exits with 0 when every task passed its check, 1 when any failed, and 2 for a bad configuration, making it suitable for cron or systemd.