fix(tools): stop web_search from swallowing requests that name a URL

Asked to open a specific page and summarise it, the model called
web_search with the URL as the query — and added a freshness filter of
its own, so the search returned nothing useful. It never fetched the
page it had just been handed.

The two descriptions are asymmetric. web_fetch already says "Use when
you already have a concrete URL/domain. NOT for open-ended searches (use
web_search)", so the model has a reason not to reach for it when
searching. web_search says nothing about URLs, so nothing argues against
using it for one. The gap shows up with smaller models, which lean on
the descriptions rather than on judgement.

Adds the reciprocal guidance. Measured with the real toolset on a local
model: the same request went from web_search to web_fetch.

Worth recording that tool count matters as much as wording here. The
same prompt resolved correctly with 3 tools in a clean context, chose
web_search with the 20 the agent loop offers, and chose web_fetch again
once the toolset was trimmed and this description fixed. The wording is
necessary but not sufficient — it is the cheap half of the fix.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/tool_schemas.py

@@ -64,7 +64,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "web_search",
-            "description": "Quick single web lookup for a fact or current event mid-task. NOT for 'research X' / 'do research on X' — those are deep-research jobs; use trigger_research instead.",
+            "description": "Quick single web lookup for a fact or current event mid-task, when you do NOT already know which page to read. NOT when the user hands you a URL — fetch that page with web_fetch instead of searching for it. NOT for 'research X' / 'do research on X' — those are deep-research jobs; use trigger_research instead.",
             "parameters": {
                 "type": "object",
                 "properties": {