indigoai-us/core/packages/hq-pack-cowork/skills/hq-cowork-search

/hq-cowork-search — Search HQ from a sandboxed agent

Searches indexed HQ content via the host-side qmd binary, surfaced as an MCP tool so a sandboxed agent (Cowork) can use it even though qmd isn't installed inside its Linux VM.

Args: $ARGUMENTS — query plus optional flags.

| Arg | Meaning | |---|---| | <query> (positional, required) | Natural-language search query. | | -c <collection> / --collection <collection> | Scope to one collection (e.g. hq-infra, hq-knowledge, indigo). | | -n <N> / --limit <N> | Number of results (default 10, max 50). | | --format json\|md\|files | Output format. Omit for the default human snippet. |

Collections worth knowing about (varies per HQ install — run qmd ls on the host to enumerate):

• hq-infra — skills, policies, hooks
• hq-knowledge — core/knowledge/public/hq-core/
• hq-workers
• hq-projects
• {company-slug} — per-company knowledge, one collection per company

When to use this instead of /search

• You're in Cowork or another sandboxed plugin host — the regular /search skill (and the qmd charter rule that says "qmd first") relies on the qmd binary on PATH, which isn't true inside the sandbox.
• You want JSON results for downstream programmatic use — pass --format json and parse the response.

On a host-side session, prefer the unprefixed /search or qmd query directly — fewer hops, same results.

What you do

Step 1 — Parse the query

Extract <query> (required) and optional flags. Without a query, ask what the user is looking for. If the query is vague ("find docs"), ask for a more specific phrase — qmd hybrid is good but not magic.

Step 2 — Call the tool

{
  "query": "<query>",
  "collection": "<collection>",   // omit to search all
  "limit": <N>,                   // omit to use default 10
  "format": "cli|json|md|files"   // omit to use default cli
}

Call mcp__hq__hq_search.

Step 3 — Surface results

For cli / md format, render the tool output as-is (already human-readable). For json / files, summarize: top 5 paths + scores, then offer to read the most relevant via the host MCP if the user wants the body.

If no results, suggest broadening — drop the collection filter, try a synonym, or fall back to qmd vsearch (pure semantic) via a follow-up call with a different query phrasing.

Why this skill exists

The HQ charter mandates "qmd first" for HQ search — but only a host-side session can actually shell out to the qmd binary. Inside Cowork's sandbox, neither qmd nor the index DB it reads (~/.qmd/) is reachable. The hq-pack-cowork MCP server runs qmd query (hybrid: expansion + RRF + rerank) on the host and returns the formatted result. This skill is the in-session adapter.