harbor-clerk

name: harbor-clerk description: Search and read documents from a local Harbor Clerk knowledge base. Use when the user references their personal documents, contracts, notes, emails, or asks "what did I store about X?"

Harbor Clerk - extended memory for agents

Harbor Clerk is a local document corpus with hybrid FTS+vector search and citation-ready reads. This skill exposes it via the harbor-clerk CLI.

First step: discover the surface

Run harbor-clerk --help for the full command list, then harbor-clerk <cmd> --help for any specific command. The help is comprehensive: JSON return shapes, examples, and common mistakes are all there.

Four patterns you'll use most

1. Search → expand: harbor-clerk search "..." then harbor-clerk expand-context <chunk_id> -n 3 on the best hit.
2. Find every matching document: harbor-clerk find-all "..." --text-contains "literal phrase" --presentation full when the task asks to list or enumerate.
3. Read a known document: harbor-clerk read-document <doc_id> for full text with pagination.
4. Check ingest status before searching for new content: harbor-clerk ingest-status <doc_id> returns per-stage state.

What you can trust

• Search returns top-level .hits[]; Find All returns top-level .results[]. Result-bearing commands include source, citation, and chunk_id when available. Use those fields as the citation trail, then call read-passages when you need exact surrounding text.
• possible_conflict: true means top hits span multiple similarly scored documents; inspect the relevant sources before answering.
• ingest-status, read-document, and entity commands may require a Full-tier API key. If access is denied, continue with search and passage reads, and tell the user the key may be scoped too narrowly for that command.
• The CLI exits non-zero on failure. Exit code 3 specifically means an admin has disabled CLI access; tell the user.