Agent Harness

The standalone debounced watcher: shipmoor-harness watch --emit json emits one privacy-safe event per change — paths, rule-ids, and counts only, never source or finding prose.

How the PostToolUse hook closes the loop: the agent makes an edit, the hook scans it, a finding blocks the tool (exit 2) and is handed back, the agent self-corrects, and the re-scan clears. Includes the Claude Code hook-response shape used in feedback mode.

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

Which agents the harness supports and how each is wired — Claude Code lifecycle hooks, Cursor (soft), and the Aider lint-cmd integration.

The local agent feedback loop: the harness watches what your coding agent edits, runs shipmoor scan against those edits, and routes findings back into the agent's context. The skills-vs-harness-vs-CLI model and the scanner/orchestrator architecture. (IC feature.)