Agent Harness

The background, agent-agnostic watcher: a debounced scan-on-change loop with a privacy-safe event stream. Paths, rule-ids, and counts only, never source or finding prose.

How Claude Code's native hooks close the loop: the agent makes an edit, the PostToolUse hook scans it, a finding is handed back (additionalContext in feedback mode, stderr and exit 2 in block mode), the agent fixes its own mistake, the re-scan clears, and the Stop hook gates before the agent declares done.

Install the harness into an agent with shipmoor harness install, inspect it with status --json, and pick a mode: observe, feedback, or block. What each mode does to the agent loop.

Wire Shipmoor into Cursor with the native .cursor/hooks.json adapter, the supported single-repo path, plus an advanced pattern for multi-root workspaces.

Wire Shipmoor into Claude Code on three levels: the scan engine, always-on Skills, and Harness lifecycle hooks that make the scan happen and gate the result.

The four harness adapters, Claude Code, Cursor, Codex, and Aider: what each install command writes, the files it manages, and how the scan decision is delivered back to the agent.

The local feedback loop for coding agents: the harness watches what your agent edits, runs shipmoor scan against those edits, and routes findings back into the agent's context so it fixes its own mistakes during the session. Bundled into the shipmoor binary as shipmoor harness. (IC feature.)