lai-harness

lai harness is a compact, auditable, local-first coding harness for OpenAI-compatible local model servers. It provides deterministic structure for a small model to inspect a repository, plan changes, edit bounded files, run validation, and expose work through an authenticated loopback control plane without treating model output as authority. The public command is lai. It is not a general-purpose agent platform, hosted service, or security sandbox for untrusted code.

The harness narrows coding work by providing deterministic repository context before model reasoning, small mode-specific tool schemas, explicit policy decisions for writes, shell, Git, network, credentials, and external actions, safe workspaces and hash-bound promotion instead of direct source checkout mutation, auditable run records, trajectories, budgets, and validation evidence, and Gateway-compatible local-chat APIs for future UI integration. It has one authoritative runtime boundary in src/local-agent; the src/lai wrapper dispatches user-facing commands. A companion Gateway can talk to the authenticated loopback control plane, but Gateway does not become the authority for filesystem, shell, credentials, browser, MCP or Git effects.

Requirements include Linux or WSL2-first workflow, Python 3.11 or newer, Git, Bash for project scripts, Node.js only for VS Code extension/package validation, an OpenAI-compatible local model endpoint for model-backed modes, and Docker only for verified remote work-runs and sandboxed fixture execution. The model server is not bundled. Security notes state that the model server or control API should not be exposed outside loopback without a separate reviewed boundary, and real credentials and authenticated external accounts are disabled in the current package.