Recipe 03 — Keyword layer from Claude Code

Goal

Answer “where is CLAUDE_CODE_SYNC_SKILLS documented?” without an embedding call. The keyword layer is an FTS5 index (BM25 ranking) built beside each docset’s vector layer; a query in mode="keyword" is a sub-token match over the facts layer, returns the unit’s snippet and its url#anchor, and costs microseconds.

When not to use it

Steps

  1. Call hub_query_docset with the token and mode="keyword". docset is the store key, <host-slug>__<mirror-stem-slug> — for this mirror codeclaudecom__codeclaudecom, not the host name; a friendly name returns no such docset. The index is built on first use (docset_indexer keyword-index under the hood), so the first call on a docset is slower once.
  2. Read url and snippet off the top hit — those are the fields a keyword hit carries (score, url, seq, snippet); text, unit_type and origin come back from the semantic and hybrid legs instead. The URL carries the anchor.
  3. Open the page with hub_llms_full_read(key, page=<url>). That key is the llms-full catalog key from hub_llms_full_list (code.claude.com__docs), a different namespace from the docset key; page matches by exact source URL or a case-insensitive title substring.

From a Claude Code session with the global_ai_hub server connected:

hub_query_docset(docset="codeclaudecom__codeclaudecom", question="CLAUDE_CODE_SYNC_SKILLS", mode="keyword", top=3)

The reply is JSON. Verbatim, run against this hub on 2026-08-31 — nothing dropped:

{
  "docset": "codeclaudecom__codeclaudecom",
  "layer": "facts",
  "queried": "codeclaudecom__codeclaudecom__facts",
  "mode": "keyword",
  "results": [
    {
      "score": 15.2427,
      "url": "https://code.claude.com/docs/en/env-vars#variables",
      "seq": 8391,
      "snippet": " … Timeout in milliseconds for a mid-session skills resync when `[CLAUDE_CODE_SYNC_SKILLS]` is set (default: 30000). Bounds the download triggered when the … "
    },
    {
      "score": 14.651,
      "url": "https://code.claude.com/docs/en/env-vars#variables",
      "seq": 8392,
      "snippet": " … Timeout in milliseconds for the first query to wait for the initial skill list when `[CLAUDE_CODE_SYNC_SKILLS]` is set (default: 5000). When … "
    },
    {
      "score": 14.3472,
      "url": "https://code.claude.com/docs/en/env-vars#variables",
      "seq": 8390,
      "snippet": " … Before it runs the first query, Claude Code waits up to `[CLAUDE_CODE_SYNC_SKILLS]_WAIT_TIMEOUT_MS` for the list of your skills … "
    }
  ]
}

The square brackets are FTS5’s snippet() match markers, not part of the text; the leading and trailing … are its elision markers, so a snippet is a window around the match rather than the whole unit.

Read the ranking honestly: all three hits are on the same page, and the variable you asked for is third. Units 8391 and 8392 are CLAUDE_CODE_SYNC_SKILLS_INSTALL_TIMEOUT_MS and …_WAIT_TIMEOUT_MS; unit 8390 is CLAUDE_CODE_SYNC_SKILLS itself. BM25 rewards the token in a shorter field, so the two timeout units outscore the variable they refer to. That costs nothing here, because all three carry the same url#anchor — keyword tells you which page, and the anchor is identical whichever of the three you take.

Then, to open the page behind the top hit:

hub_llms_full_read(key="code.claude.com__docs", page="https://code.claude.com/docs/en/env-vars")

That returns an envelope, not raw markdown. The real fields on this call, with text cut here (it is 20,000 characters):

{
 "key": "code.claude.com__docs",
 "url": "https://code.claude.com/docs/llms-full.txt",
 "page_title": "Environment variables",
 "page_url": "https://code.claude.com/docs/en/env-vars",
 "total_chars": 475588,
 "truncated": true,
 "text": "Reference for environment variables that control Claude Code behavior.\n\nEnvironment variables can control Claude Code behavior such as model selection, authentication, request routing, and feature toggles. …  ← truncated here for the page; the tool returned 20,000 characters"
}

The # Title / Source: lines are the grammar the tool matches on inside llms-full.txt; they are lifted into page_title and page_url rather than left in text, so text starts at the page body. total_chars is the page’s full length and truncated says whether you got all of it — continue with offset, or raise limit.

The same two calls from the shell, for a script:

.venv/bin/python scripts/docset_indexer.py keyword codeclaudecom__codeclaudecom "CLAUDE_CODE_SYNC_SKILLS" --layer facts --mode phrase --top 3

--mode is any | all | phrase | raw; phrase is what an exact token wants.

Expected output

Every hit’s URL ends in the heading that documents the variable (/docs/en/env-vars#variables), and the reply’s layer names which layer answered (facts) while queried names the collection it actually read (…__facts). The corresponding line in the export’s llms-facts.txt is typed [parameter] — line 11,378 of that file begins - [parameter] `CLAUDE_CODE_SYNC_SKILLS`: Set to `1` … — but keyword hits do not carry the type, so ask for mode="hybrid" or "semantic" when you want unit_type on the hit. What a keyword hit does carry is exactly score, url, seq, snippet, as above.

The full-read then lands you on that page: page_title comes back as "Environment variables", page_url as the URL you asked for, so the anchor from the hit resolves to a heading you can see in text.

If layer says raw, the docset has no facts layer; the hit is still correct but is a text chunk without an anchor. That is the signal to run extract → export on it.

Cost

Measured on the hub: the FTS5 lookup is sub-millisecond after the index exists; the first call on a docset builds the index (seconds for this family’s 14,031 units). Zero model tokens, zero embeddings. The page read is bounded by limit, which defaults to 20,000 characters — about 5k tokens — and is capped at 200k by the tool. The env-vars page above is 475,588 characters, so the default read came back "truncated": true; a shorter reference page returns whole.

Runnable in step 4 (playground).