@voidwire/llm-summarize 3.4.0 → 3.6.0

package/index.ts

@@ -30,15 +30,17 @@ export interface SessionInsights {
   should_search?: boolean;
   extractions?: Extraction[];
   // Insights mode fields
+  current_focus?: string;
+  next_steps?: string[];
   decisions?: string[];
   patterns_used?: string[];
   preferences_expressed?: string[];
   problems_solved?: string[];
-  tools_heavy?: string[];
 }
 
 export interface SummarizeResult {
   insights?: SessionInsights;
+  rawText?: string;
   error?: string;
   model?: string;
   tokens_used?: number;
@@ -58,6 +60,8 @@ export interface SummarizeOptions {
   mode?: "quick" | "insights";
   /** User name to include in summary (e.g., "Rudy") */
   userName?: string;
+  /** Override the system prompt (bypasses mode-based prompt selection) */
+  systemPrompt?: string;
 }
 
 export type ProviderType = "anthropic" | "openai" | "ollama";
@@ -141,53 +145,52 @@ Output valid JSON only. No markdown, no explanation.`;
  * Note: userName param kept for API compatibility but not used in insights mode
  */
 function buildInsightsPrompt(_userName?: string): string {
-  return `You are a senior engineering manager extracting reusable insights from development sessions.
+  return `You are an engineering knowledge extractor. Given a development session transcript, extract reusable insights as structured JSON.
 
-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)
+Transcripts use role markers:
+- "User Asked:" = the human (directs, decides, provides context)
+- "Assistant Response:" = the AI (implements, builds, debugs)
 
-Your job: extract what's worth remembering for future sessions.
+<output_format>
+Return a JSON object with these fields. Include a field ONLY when the transcript provides clear evidence. Omit empty arrays entirely.
 
-<output_schema>
 {
-  "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"]
+  "summary": "One sentence: what was accomplished and the key outcome",
+  "current_focus": "The specific task, feature, or problem actively being worked on (omit if exploratory)",
+  "next_steps": ["Concrete action to take when work resumes — name the actual task"],
+  "decisions": ["Decision made — rationale and what alternatives were considered"],
+  "patterns_used": ["Technique or approach applied — why it was chosen over alternatives"],
+  "preferences_expressed": ["User preference revealed through direction, correction, or explicit statement"],
+  "problems_solved": ["Problem encountered — root cause identified and specific fix applied"]
 }
-</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>
-- Include a field ONLY when the transcript provides clear evidence
-- Omit empty arrays entirely
-- 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.`;
+</output_format>
+
+<quality_rules>
+Every value MUST be a complete sentence with context. Never output bare nouns, short phrases, or sentence fragments.
+
+BAD (will be rejected):
+- "SQLite"
+- "detached worker"
+- "Fixed bug"
+- "Continue working"
+
+GOOD (specific, contextual, reusable):
+- "Chose SQLite over Postgres for single-user CLI — no server dependency needed"
+- "Used detached worker pattern to avoid blocking the stop hook during LLM calls"
+- "Fixed state file writing to wrong directory — was using read-only data path instead of persistent home"
+- "Wire up the webhook endpoint to the event processor and verify with integration test"
+
+For next_steps specifically: never say "Continue from current position" or "Resume work" — name the actual task to be done.
+</quality_rules>
+
+<attribution>
+Users direct and decide. Assistants implement and execute.
+- User: requested, approved, directed, chose, preferred, corrected
+- Assistant: implemented, built, debugged, refactored, created, fixed
+- Never say "User implemented" or "User built"
+</attribution>
+
+Output valid JSON only. No markdown, no code blocks, no explanation.`;
 }
 
 /**
@@ -421,12 +424,14 @@ async function callAnthropic(
 
     if (!insights) {
       return {
+        rawText: content,
         error: `Failed to parse response as JSON: ${content.slice(0, 200)}`,
       };
     }
 
     return {
       insights,
+      rawText: content,
       model,
       tokens_used: result.usage?.output_tokens,
     };
@@ -487,12 +492,14 @@ async function callOpenAI(
 
     if (!insights) {
       return {
+        rawText: content,
         error: `Failed to parse response as JSON: ${content.slice(0, 200)}`,
       };
     }
 
     return {
       insights,
+      rawText: content,
       model,
       tokens_used: result.usage?.completion_tokens,
     };
@@ -554,12 +561,14 @@ async function callOllama(
 
     if (!insights) {
       return {
+        rawText: content,
         error: `Failed to parse response as JSON: ${content.slice(0, 200)}`,
       };
     }
 
     return {
       insights,
+      rawText: content,
       model,
       tokens_used: result.eval_count,
     };
@@ -597,7 +606,8 @@ export async function summarize(
   const apiKey = config.apiKey;
   const mode: SummarizeMode = options?.mode || "insights";
   const userName = options?.userName;
-  const systemPrompt = getPromptForMode(mode, userName);
+  const systemPrompt =
+    options?.systemPrompt || getPromptForMode(mode, userName);
 
   // Validate config
   if (!provider) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voidwire/llm-summarize",
-  "version": "3.4.0",
+  "version": "3.6.0",
   "description": "Structured session insight extraction for knowledge systems",
   "type": "module",
   "main": "./index.ts",
@@ -18,9 +18,6 @@
     "README.md",
     "LICENSE"
   ],
-  "scripts": {
-    "test": "bun test"
-  },
   "keywords": [
     "llm",
     "summarize",
@@ -42,5 +39,8 @@
   },
   "engines": {
     "bun": ">=1.0.0"
+  },
+  "scripts": {
+    "test": "bun test"
   }
-}
+}