Developer Index skill documents four id prefixes; a fifth (web:) is the majority of unscoped results

[prettycolor]

The firecrawl-developer-index skill (and docs.firecrawl.dev/features/developer) both state
that the artifact kind is the id prefix, and enumerate four: doc:, issue:, pull_request:,
readme:.

In practice a fifth prefix, web:, is returned, and on unscoped queries it is frequently the
majority of results. It does not appear in the skill file, the feature docs, or the API reference.

Reproduction (keyless, no API key needed)

curl -s -X POST https://api.firecrawl.dev/v2/search/developer \
  -H "Content-Type: application/json" \
  -d '{"query":"drizzle orm mysql connection pool exhausted","k":8}' \
| python3 -c "import sys,json,collections; d=json.load(sys.stdin); print(collections.Counter(r['id'].split(':')[0] for r in d['results']))"

Measured 2026-08-28 against api.firecrawl.dev/v2:

Request	Prefix distribution
unscoped, k=8	{'issue': 3, 'doc': 2, 'web': 3}
unscoped, k=8, different query	{'issue': 3, 'doc': 2, 'web': 3}
same query + types:["doc","issue","pull_request","readme"]	{'issue': 4, 'doc': 3, 'pull_request': 1} — no web:
types:["web"]	400 {"code":"BAD_REQUEST","error":"unknown type: web (expected doc, issue, pull_request, readme)"}

So web: can be excluded, but only by naming all four valid types explicitly. It cannot be
requested, and it is not documented.

Why this matters for agent behaviour

Three things in the shipped guidance combine badly:

1. The skill tells the agent the kind is one of four prefixes.
2. The skill's stated bar is "quote the passage, cite the url" as a primary source.
3. The skill says "Scope last, not first — search the whole index, then narrow."

Following (3) produces the unscoped call, which returns majority web: — ordinary web pages.
The agent then satisfies (2) by quoting one, believing per (1) that it is a curated artifact.
The net effect is that the recommended first move is the one most likely to return
non-primary sources, with nothing in the guidance to flag it.

Also worth noting: the CLI surface cannot mitigate this. On firecrawl-cli@1.23.3,
firecrawl developer exposes only --limit, -o, --json, --pretty, -k, and its help says
to put scoping intent in the query text. There is no --types, so a CLI user cannot exclude
web: at all. A live CLI run returned {issue: 3, doc: 2, web: 3}.

Suggested fix (any one of these would close it)

• Document web: as a possible prefix in the skill file, the feature page, and the API reference,
  and say plainly that it is an open web page rather than a curated index artifact; or
• have the skill recommend sending all four types by default, rather than "scope last"; or
• add a --types flag to the CLI developer command so the CLI surface can exclude it too; or
• exclude web: from /search/developer and leave it to /search.

Happy to send the raw responses if useful.

---

One more control, added after drafting

The distribution is identical authenticated and keyless, so this is not an auth-tier artifact:

surface	prefixes
authenticated CLI, unscoped	{issue: 3, doc: 2, web: 3}
keyless HTTP, same query, unscoped	{issue: 3, doc: 2, web: 3}
keyless HTTP, same query, + types	{issue: 4, doc: 3, pull_request: 1}

The variable is types, not the key. A different query returned {issue: 4, doc: 4} on both
surfaces, so web: is query-dependent rather than always present.