fix(librarian): use terse tool-phase prompt so gemma4 calls tools

The scholarly persona prompt reproduced the exact pathology
TATLOCK_ORCHESTRATION_PROMPT fixed for the butler: gemma4 answered in
character ('please provide your request') without calling a single tool.
The research phase now uses a tool-discipline prompt; Tatlock's synthesis
supplies the voice. Anti-fabrication rules kept verbatim.

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

src/agents/librarian/agent.py

@@ -138,6 +138,30 @@ If a tool fails or you cannot access a data source:
 - It is better to return no information than to return fabricated information
 """
 
+
+# Tool-phase prompt actually used by the agent. The scholarly persona prompt
+# above suppresses tool calling on small local models (gemma4 answers in
+# character - "please provide your request" - without ever calling a tool),
+# the same pathology TATLOCK_ORCHESTRATION_PROMPT fixed for the butler.
+# Tatlock's synthesis phase supplies the user-facing voice, so the research
+# phase only needs tool discipline. Kept: the anti-fabrication rule.
+LIBRARIAN_TASK_PROMPT = """You are The Librarian, the research executor of the \
+Tatlock household. Your only job is to gather accurate findings by calling the \
+provided tools.
+
+- ALWAYS use tools - never answer a research task from memory alone.
+- Research or wiki questions: call hybrid_search first; then search_wiki and \
+get_wiki_page to read specific pages BEFORE summarizing them.
+- Current or external information (weather, news, live facts): call search_web; \
+call read_url when given a specific URL.
+- Wiki writing: smart_create_wiki_page when asked for a page about a topic; \
+create_wiki_page only for user-provided verbatim content; update_wiki_page for \
+edits (search_wiki, then get_wiki_page, then update).
+- Reply with a concise factual summary of what the tools returned, citing page \
+titles and URLs. A later step writes the polished answer, so no personality.
+- NEVER fabricate. If a tool fails or returns nothing, state exactly what you \
+could not retrieve and stop."""
+
 # Lazy initialization to avoid connection issues during imports
 _librarian_agent: Agent[None, str] | None = None
 
@@ -151,7 +175,7 @@ def _create_librarian_agent() -> Agent[None, str]:
 
     agent: Agent[None, str] = Agent(
         model=model,
-        system_prompt=LIBRARIAN_SYSTEM_PROMPT,
+        system_prompt=LIBRARIAN_TASK_PROMPT,
         retries=2,
     )