@voidwire/llm-summarize 3.2.0 → 3.4.0

package/index.ts

@@ -18,8 +18,18 @@ import { join } from "path";
 // Types
 // ============================================================================
 
+export interface Extraction {
+  term: string;
+  type: "project" | "topic" | "tool" | "person";
+  confidence: "high" | "medium";
+}
+
 export interface SessionInsights {
   summary: string;
+  // Quick mode extraction fields
+  should_search?: boolean;
+  extractions?: Extraction[];
+  // Insights mode fields
   decisions?: string[];
   patterns_used?: string[];
   preferences_expressed?: string[];
@@ -59,44 +69,122 @@ export type SummarizeMode = "quick" | "insights";
 
 /**
  * Build quick mode prompt with optional user name
+ * Now includes context extraction for knowledge retrieval
  */
 function buildQuickPrompt(userName?: string): string {
-  const nameInstruction = userName ? `Start with "${userName}".` : "";
+  const name = userName || "User";
+
+  return `You are a context classifier for knowledge retrieval. Analyze conversation context to determine what prior knowledge would be valuable.
+
+Input format:
+Project: <project name>
+Previous Assistant: <last assistant message>
+User Prompt: <current user message>
+
+Produce JSON with:
+1. summary: Brief description (1-2 sentences) of what the user is doing/asking. Start with "${name}".
+2. should_search: Whether to search the knowledge base
+3. extractions: Terms worth searching for
+
+should_search = true when:
+- References past work, decisions, discussions
+- Mentions project, tool, or person by name
+- Asks "what was...", "how did we...", "remember when..."
+- Technical domain benefits from prior learnings
+
+should_search = false when:
+- Greetings, acknowledgments ("ready", "thanks", "ok")
+- Simple commands ("run tests", "commit this")
+- Continuation signals ("yes", "do it", "go ahead")
+
+Extraction types:
+- project: Named codebase, repo, system (sable, lore, momentum)
+- topic: Domain, concept, technical area (hooks, authentication, Tier 2)
+- tool: Library, CLI, framework (llm-summarize, SQLite, Bun)
+- person: Named individual
+
+Confidence:
+- high: Explicitly stated
+- medium: Strongly implied
+
+Skip generic words. Only extract terms that yield useful knowledge results.
 
-  return `Summarize what the user is asking or doing in one sentence.
-${nameInstruction}
-Output JSON only: {"summary": "One sentence summary"}`;
+<example>
+Project: sable
+Previous Assistant: I'll update the UserPromptSubmit hook to call llm-summarize.
+User Prompt: What does Lore return for project queries?
+
+{"summary": "${name} is asking about Lore's return format for project queries", "should_search": true, "extractions": [{"term": "Lore", "type": "project", "confidence": "high"}, {"term": "project queries", "type": "topic", "confidence": "high"}]}
+</example>
+
+<example>
+Project: sable
+Previous Assistant: The extraction prompt is ready. Should I add it?
+User Prompt: yes do it
+
+{"summary": "${name} is confirming to proceed with the extraction prompt", "should_search": false, "extractions": []}
+</example>
+
+<example>
+Project: sable
+Previous Assistant: Starting new session.
+User Prompt: What was the issue we hit with the stop hook last time?
+
+{"summary": "${name} is asking about a previous issue with the stop hook", "should_search": true, "extractions": [{"term": "stop hook", "type": "topic", "confidence": "high"}, {"term": "sable", "type": "project", "confidence": "medium"}]}
+</example>
+
+Output valid JSON only. No markdown, no explanation.`;
 }
 
 /**
- * Build insights mode prompt with optional user name
+ * Build insights mode prompt for session insight extraction
+ * Note: userName param kept for API compatibility but not used in insights mode
  */
-function buildInsightsPrompt(userName?: string): string {
-  const nameInstruction = userName
-    ? `Start the summary with "${userName}".`
-    : "";
+function buildInsightsPrompt(_userName?: string): string {
+  return `You are a senior engineering manager extracting reusable insights from development sessions.
 
-  return `You are an experienced engineering manager reviewing session transcripts to extract actionable insights.
+You receive transcripts with clear role markers:
+- "User Asked:" = the human directing work (requests, approves, provides context)
+- "Assistant Response:" = the AI executing work (implements, builds, debugs, explains)
 
-Analyze the development session and extract structured observations.
+Your job: extract what's worth remembering for future sessions.
 
 <output_schema>
 {
-  "summary": "One sentence: what was accomplished or decided",
-  "decisions": ["Specific decision and its reasoning"],
-  "patterns_used": ["Development pattern or approach observed"],
-  "preferences_expressed": ["Preference revealed through actions - DO NOT include user name"],
-  "problems_solved": ["Problem addressed and how - DO NOT include user name"],
-  "tools_heavy": ["Tool used repeatedly or notably"]
+  "summary": "One sentence capturing what was accomplished and how",
+  "decisions": ["Decision made with reasoning and trade-offs considered"],
+  "patterns_used": ["Development pattern or approach, with context on why it was chosen"],
+  "preferences_expressed": ["Preference revealed through direction or feedback"],
+  "problems_solved": ["Problem encountered and the specific solution applied"],
+  "tools_heavy": ["Tool used repeatedly or for critical work"]
 }
 </output_schema>
 
+<attribution_rules>
+- User actions: requested, approved, directed, provided, chose, preferred
+- Assistant actions: implemented, built, debugged, refactored, created, fixed
+- Never say "User implemented" or "User built" — users direct, assistants execute
+</attribution_rules>
+
+<quality_guidance>
+Extract specifics with context, not bare facts:
+
+SPARSE (avoid):
+- "Made a database decision"
+- "Fixed a bug"
+- "Used TypeScript"
+
+RICH (prefer):
+- "Chose SQLite over Postgres for single-user CLI tool — avoids server dependency"
+- "Fixed race condition in webhook handler by adding mutex lock — was causing duplicate events"
+- "Used Zod for runtime validation at API boundary — catches malformed input before it hits business logic"
+</quality_guidance>
+
 <rules>
-- ${nameInstruction || "Write summary in third person."}
-- Include a field ONLY when the conversation provides clear evidence
-- Extract specifics: "Chose SQLite over Postgres for single-user simplicity" not "Made a database decision"
+- Include a field ONLY when the transcript provides clear evidence
 - Omit empty arrays entirely
-- IMPORTANT: Only use user name in the summary field, nowhere else
+- Capture the "why" when present — reasoning is more valuable than the decision alone
+- Technical specifics (library names, patterns, trade-offs) make insights reusable
 </rules>
 
 Output valid JSON only. No markdown code blocks, no explanation.`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voidwire/llm-summarize",
-  "version": "3.2.0",
+  "version": "3.4.0",
   "description": "Structured session insight extraction for knowledge systems",
   "type": "module",
   "main": "./index.ts",