@voidwire/llm-summarize 3.6.0 → 3.8.0

package/cli.ts

@@ -2,27 +2,15 @@
 /**
  * llm-summarize CLI
  *
- * Philosophy:
- * - Structured session insight extraction for knowledge systems
- * - Multi-provider support (Anthropic, OpenAI, Ollama)
- * - Deterministic JSON output for tooling integration
- * - Config-driven - no hardcoded defaults
+ * Structured session insight extraction for knowledge systems.
+ * Uses @voidwire/llm-core for LLM transport — services configured
+ * via ~/.config/llm-core/services.toml, API keys via apiconf.
  *
  * Usage:
  *   llm-summarize <text>
  *   llm-summarize --stdin
  *   echo "text" | llm-summarize --stdin
  *
- * Config: ~/.config/llm/config.toml
- *   [llm]
- *   provider = "ollama"
- *   model = "Qwen2.5:3b"
- *   api_base = "https://ollama.example.com"
- *   max_tokens = 1024
- *
- * Secrets: ~/.config/llm/.env
- *   ANTHROPIC_API_KEY=sk-ant-...
- *
  * Exit codes:
  *   0 - Success
  *   1 - API error (rate limit, auth, network)
@@ -56,18 +44,13 @@ function printUsage(): void {
   console.error(`
 llm-summarize - Extract structured insights from session transcripts
 
-Philosophy:
-  Structured session insight extraction for knowledge systems.
-  Config-driven - specify exact provider/model.
-  JSON output for tooling integration.
-
 Usage: llm-summarize [options] <text>
        llm-summarize --stdin
 
 Options:
   --mode <mode>         Summarization mode: quick or insights (default: insights)
-  --model <name>        Override model from config
-  --max-tokens <n>      Max output tokens (default: from config or 1024)
+  --model <name>        Override model (default: claude-3-5-haiku-20241022)
+  --max-tokens <n>      Max output tokens (default: 1024)
   --stdin               Read text from stdin
   -h, --help            Show this help
 
@@ -75,26 +58,10 @@ Modes:
   quick     - Fast one-liner summary (for user prompts)
   insights  - Full SessionInsights extraction (for responses)
 
-Config file: ~/.config/llm/config.toml
-  [llm]
-  provider = "ollama"
-  model = "Qwen2.5:3b"
-  api_base = "https://ollama.example.com"
-  max_tokens = 1024
-
-Secrets file: ~/.config/llm/.env
-  ANTHROPIC_API_KEY=sk-ant-...
-  OPENAI_API_KEY=sk-...
-
-Environment overrides:
-  LLM_PROVIDER          Override provider
-  LLM_MODEL             Override model
-  LLM_API_KEY           Override API key
-
-Supported providers:
-  anthropic - Claude models (claude-3-5-haiku-latest, claude-sonnet-4-20250514)
-  openai    - GPT models (gpt-4.1-mini, gpt-4o)
-  ollama    - Local models (Qwen2.5:3b, llama3.2:3b, etc.) - no API key needed
+Configuration:
+  LLM transport is handled by @voidwire/llm-core.
+  Services: ~/.config/llm-core/services.toml
+  API keys: managed via apiconf (see @voidwire/apiconf)
 
 Output format:
   {
@@ -103,18 +70,14 @@ Output format:
       "decisions": ["Specific decisions with reasoning"],
       "patterns_used": ["Development patterns observed"],
       "preferences_expressed": ["User preferences revealed"],
-      "problems_solved": ["Problems addressed and how"],
-      "tools_heavy": ["Tools used notably"]
+      "problems_solved": ["Problems addressed and how"]
     },
-    "model": "qwen2.5:3b",
+    "model": "claude-3-5-haiku-20241022",
     "tokens_used": 150
   }
 
 Examples:
-  # Extract insights from session transcript
   cat session.txt | llm-summarize --stdin
-
-  # From clipboard
   pbpaste | llm-summarize --stdin
 `);
 }

package/index.ts

@@ -11,8 +11,7 @@
  *   // result.insights.summary, result.insights.decisions, etc.
  */
 
-import { readFileSync, existsSync } from "fs";
-import { join } from "path";
+import { complete } from "@voidwire/llm-core";
 
 // ============================================================================
 // Types
@@ -46,12 +45,10 @@ export interface SummarizeResult {
   tokens_used?: number;
 }
 
-export interface LLMConfig {
-  provider: string | null;
-  model: string | null;
-  apiKey: string | null;
-  apiBase: string | null;
-  maxTokens: number;
+export interface SummarizeConfig {
+  service?: string; // Named service from services.toml (optional, uses default_service)
+  model: string; // Model name — required by complete()
+  maxTokens: number; // Max output tokens
 }
 
 export interface SummarizeOptions {
@@ -64,7 +61,6 @@ export interface SummarizeOptions {
   systemPrompt?: string;
 }
 
-export type ProviderType = "anthropic" | "openai" | "ollama";
 export type SummarizeMode = "quick" | "insights";
 
 // ============================================================================
@@ -145,52 +141,57 @@ 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 an engineering knowledge extractor. Given a development session transcript, extract reusable insights as structured JSON.
-
-Transcripts use role markers:
-- "User Asked:" = the human (directs, decides, provides context)
-- "Assistant Response:" = the AI (implements, builds, debugs)
-
-<output_format>
-Return a JSON object with these fields. Include a field ONLY when the transcript provides clear evidence. Omit empty arrays entirely.
-
-{
-  "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_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.`;
+  return `You are a session state extractor. Given a development conversation, produce a JSON snapshot of the session's current state.
+
+<instructions>
+1. Read the conversation in the <transcript> section
+2. Ignore the <previous_state> section — it is background context only, not part of this session
+3. Extract ONLY what happened in the transcript
+4. Produce a JSON object with the fields described below
+</instructions>
+
+<fields>
+- summary: One sentence describing what was accomplished this session
+- current_focus: The specific task or feature being worked on (omit if exploratory)
+- next_steps: Array of concrete next actions. Name the specific task.
+- decisions: Array of decisions made this session, each with rationale
+- patterns_used: Array of techniques or approaches applied, each with context
+- preferences_expressed: Array of user preferences revealed through direction or correction
+- problems_solved: Array of problems encountered with root cause and fix
+</fields>
+
+Include a field only when the transcript contains clear evidence. Omit empty arrays. Every value must be a complete sentence.
+
+<example>
+<input>
+<previous_state>Focus: Building authentication system</previous_state>
+<transcript>
+User Asked: Let's use JWT instead of sessions for auth
+Assistant Response: Switched from express-session to jsonwebtoken. JWTs are stateless so we don't need Redis for session storage anymore. Updated the middleware to verify tokens on each request.
+User Asked: Make sure the tokens expire after 24 hours
+Assistant Response: Set expiresIn to 24h in the sign options. Also added a refresh token flow so users don't get logged out mid-work.
+</transcript>
+</input>
+<output>
+{"summary":"Implemented JWT-based authentication replacing session-based auth, with 24-hour token expiry and refresh token flow","current_focus":"Authentication system implementation","next_steps":["Test the refresh token flow with expired tokens","Add token revocation for logout"],"decisions":["Chose JWT over sessions — eliminates Redis dependency since tokens are stateless","Set 24-hour token expiry with refresh flow — balances security with user convenience"],"preferences_expressed":["User directed specific token expiry of 24 hours"]}
+</output>
+</example>
+
+<example>
+<input>
+<previous_state>Focus: Investigating test failures</previous_state>
+<transcript>
+User Asked: The CI is failing on the webhook tests
+Assistant Response: Found the issue — the test was using a hardcoded timestamp that expired. Changed it to use a relative timestamp. Also found that the webhook handler had a race condition where two events could arrive simultaneously and both pass the idempotency check. Added a mutex lock.
+User Asked: Good catch on the race condition
+</transcript>
+</input>
+<output>
+{"summary":"Fixed CI test failure caused by hardcoded timestamp and discovered a race condition in the webhook handler","current_focus":"Webhook test failures and handler reliability","problems_solved":["Fixed expired hardcoded timestamp in webhook tests — replaced with relative timestamp calculation","Fixed race condition in webhook handler where simultaneous events bypassed idempotency check — added mutex lock"],"next_steps":["Verify CI passes with the timestamp and mutex fixes"]}
+</output>
+</example>
+
+Output valid JSON only.`;
 }
 
 /**
@@ -254,329 +255,22 @@ function extractJson(raw: string): SessionInsights | null {
 // ============================================================================
 
 /**
- * Load environment variables from .env file
- */
-function loadEnvFile(envPath: string): Record<string, string> {
-  const env: Record<string, string> = {};
-
-  if (!existsSync(envPath)) {
-    return env;
-  }
-
-  try {
-    const content = readFileSync(envPath, "utf-8");
-    for (const line of content.split("\n")) {
-      const trimmed = line.trim();
-      if (!trimmed || trimmed.startsWith("#")) continue;
-
-      const eqIdx = trimmed.indexOf("=");
-      if (eqIdx === -1) continue;
-
-      const key = trimmed.slice(0, eqIdx).trim();
-      let value = trimmed.slice(eqIdx + 1).trim();
-
-      // Remove quotes if present
-      if (
-        (value.startsWith('"') && value.endsWith('"')) ||
-        (value.startsWith("'") && value.endsWith("'"))
-      ) {
-        value = value.slice(1, -1);
-      }
-
-      env[key] = value;
-    }
-  } catch {
-    // Ignore parse errors
-  }
-
-  return env;
-}
-
-/**
- * Resolve env: references in config values
- */
-function resolveEnvRef(
-  value: string,
-  envVars: Record<string, string>,
-): string | null {
-  if (value.startsWith("env:")) {
-    const varName = value.slice(4);
-    return envVars[varName] || process.env[varName] || null;
-  }
-  return value;
-}
-
-/**
- * Load configuration from config.toml with env file support
- * Config: ~/.config/llm/config.toml
- * Secrets: ~/.config/llm/.env
+ * Load configuration for llm-summarize.
+ *
+ * Returns defaults suitable for llm-core's complete() function.
+ * Service resolution (API keys, endpoints) is handled by llm-core
+ * via ~/.config/llm-core/services.toml and apiconf.
+ *
+ * To configure:
+ *   1. Set up apiconf: ~/.config/apiconf/config.toml
+ *   2. Set up services: ~/.config/llm-core/services.toml
+ *   3. Optionally override model/maxTokens via SummarizeOptions
  */
-export function loadConfig(): LLMConfig {
-  const configDir = join(process.env.HOME!, ".config", "llm");
-  const configPath = join(configDir, "config.toml");
-  const envPath = join(configDir, ".env");
-
-  // Load .env file first
-  const envVars = loadEnvFile(envPath);
-
-  // No defaults - config required
-  const config: LLMConfig = {
-    provider: null,
-    model: null,
-    apiKey: null,
-    apiBase: null,
+export function loadConfig(): SummarizeConfig {
+  return {
+    model: "claude-3-5-haiku-20241022",
     maxTokens: 1024,
   };
-
-  if (!existsSync(configPath)) {
-    return config;
-  }
-
-  try {
-    const content = readFileSync(configPath, "utf-8");
-
-    // Parse [llm] section
-    const providerMatch = content.match(/^\s*provider\s*=\s*"([^"]+)"/m);
-    if (providerMatch) {
-      config.provider = providerMatch[1];
-    }
-
-    const modelMatch = content.match(/^\s*model\s*=\s*"([^"]+)"/m);
-    if (modelMatch) {
-      config.model = modelMatch[1];
-    }
-
-    const apiKeyMatch = content.match(/^\s*api_key\s*=\s*"([^"]+)"/m);
-    if (apiKeyMatch) {
-      config.apiKey = resolveEnvRef(apiKeyMatch[1], envVars);
-    }
-
-    const apiBaseMatch = content.match(/^\s*api_base\s*=\s*"([^"]+)"/m);
-    if (apiBaseMatch) {
-      config.apiBase = apiBaseMatch[1];
-    }
-
-    const maxTokensMatch = content.match(/^\s*max_tokens\s*=\s*(\d+)/m);
-    if (maxTokensMatch) {
-      config.maxTokens = parseInt(maxTokensMatch[1], 10);
-    }
-  } catch {
-    // Ignore parse errors
-  }
-
-  // Environment variables override config
-  if (process.env.LLM_PROVIDER) config.provider = process.env.LLM_PROVIDER;
-  if (process.env.LLM_MODEL) config.model = process.env.LLM_MODEL;
-  if (process.env.LLM_API_KEY) config.apiKey = process.env.LLM_API_KEY;
-
-  return config;
-}
-
-// ============================================================================
-// Provider Implementations
-// ============================================================================
-
-/**
- * Call Anthropic API
- */
-async function callAnthropic(
-  text: string,
-  model: string,
-  maxTokens: number,
-  apiKey: string,
-  systemPrompt: string,
-  apiBase?: string,
-): Promise<SummarizeResult> {
-  const endpoint = apiBase || "https://api.anthropic.com/v1/messages";
-
-  try {
-    const response = await fetch(endpoint, {
-      method: "POST",
-      headers: {
-        "x-api-key": apiKey,
-        "anthropic-version": "2023-06-01",
-        "content-type": "application/json",
-      },
-      body: JSON.stringify({
-        model,
-        max_tokens: maxTokens,
-        temperature: 0.3,
-        system: systemPrompt,
-        messages: [
-          {
-            role: "user",
-            content: text,
-          },
-        ],
-      }),
-    });
-
-    if (!response.ok) {
-      const errorText = await response.text();
-      return {
-        error: `Anthropic API error: ${response.status} ${errorText}`,
-      };
-    }
-
-    const result = await response.json();
-    const content = result.content?.[0]?.text || "";
-    const insights = extractJson(content);
-
-    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,
-    };
-  } catch (error) {
-    return {
-      error: `Anthropic request failed: ${String(error)}`,
-    };
-  }
-}
-
-/**
- * Call OpenAI API
- */
-async function callOpenAI(
-  text: string,
-  model: string,
-  maxTokens: number,
-  apiKey: string,
-  systemPrompt: string,
-  apiBase?: string,
-): Promise<SummarizeResult> {
-  const endpoint = apiBase || "https://api.openai.com/v1/chat/completions";
-
-  try {
-    const response = await fetch(endpoint, {
-      method: "POST",
-      headers: {
-        Authorization: `Bearer ${apiKey}`,
-        "Content-Type": "application/json",
-      },
-      body: JSON.stringify({
-        model,
-        max_tokens: maxTokens,
-        temperature: 0.3,
-        messages: [
-          {
-            role: "system",
-            content: systemPrompt,
-          },
-          {
-            role: "user",
-            content: text,
-          },
-        ],
-      }),
-    });
-
-    if (!response.ok) {
-      const errorText = await response.text();
-      return {
-        error: `OpenAI API error: ${response.status} ${errorText}`,
-      };
-    }
-
-    const result = await response.json();
-    const content = result.choices?.[0]?.message?.content || "";
-    const insights = extractJson(content);
-
-    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,
-    };
-  } catch (error) {
-    return {
-      error: `OpenAI request failed: ${String(error)}`,
-    };
-  }
-}
-
-/**
- * Call Ollama API (chat endpoint for system prompt support)
- */
-async function callOllama(
-  text: string,
-  model: string,
-  maxTokens: number,
-  apiBase: string,
-  systemPrompt: string,
-): Promise<SummarizeResult> {
-  const endpoint = `${apiBase}/api/chat`;
-
-  try {
-    const response = await fetch(endpoint, {
-      method: "POST",
-      headers: {
-        "Content-Type": "application/json",
-      },
-      body: JSON.stringify({
-        model,
-        messages: [
-          {
-            role: "system",
-            content: systemPrompt,
-          },
-          {
-            role: "user",
-            content: text,
-          },
-        ],
-        stream: false,
-        options: {
-          num_predict: maxTokens,
-          temperature: 0.3,
-        },
-      }),
-    });
-
-    if (!response.ok) {
-      const errorText = await response.text();
-      return {
-        error: `Ollama API error: ${response.status} ${errorText}`,
-      };
-    }
-
-    const result = await response.json();
-    const content = result.message?.content || "";
-    const insights = extractJson(content);
-
-    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,
-    };
-  } catch (error) {
-    return {
-      error: `Ollama request failed: ${String(error)}`,
-    };
-  }
 }
 
 // ============================================================================
@@ -587,7 +281,7 @@ async function callOllama(
  * Summarize text using configured LLM
  *
  * @param text - Text to summarize
- * @param config - LLM configuration (from loadConfig())
+ * @param config - Summarize configuration (from loadConfig())
  * @param options - Optional overrides for model, maxTokens, and mode
  * @returns SummarizeResult with insights or error
  *
@@ -597,67 +291,50 @@ async function callOllama(
  */
 export async function summarize(
   text: string,
-  config: LLMConfig,
+  config: SummarizeConfig,
   options?: SummarizeOptions,
 ): Promise<SummarizeResult> {
-  const provider = config.provider;
-  const model = options?.model || config.model;
-  const maxTokens = options?.maxTokens || config.maxTokens;
-  const apiKey = config.apiKey;
-  const mode: SummarizeMode = options?.mode || "insights";
-  const userName = options?.userName;
-  const systemPrompt =
-    options?.systemPrompt || getPromptForMode(mode, userName);
-
-  // Validate config
-  if (!provider) {
-    return {
-      error: `No provider configured. Set provider in ~/.config/llm/config.toml`,
-    };
-  }
-
-  if (!model) {
-    return {
-      error: `No model configured. Set model in ~/.config/llm/config.toml`,
-    };
-  }
-
-  // API key required for cloud providers
-  if (!apiKey && provider !== "ollama") {
-    return {
-      error: `No API key configured. Set api_key = "env:VAR_NAME" in ~/.config/llm/config.toml`,
-    };
-  }
+  try {
+    const mode: SummarizeMode = options?.mode || "insights";
+    const userName = options?.userName;
+    const systemPrompt =
+      options?.systemPrompt || getPromptForMode(mode, userName);
+    const model = options?.model || config.model;
+
+    // Validate model before calling complete() (which throws on empty model)
+    if (!model) {
+      return {
+        error:
+          "No model configured. Set model in loadConfig() or pass via options.model",
+      };
+    }
 
-  // Call appropriate provider
-  if (provider === "anthropic") {
-    return callAnthropic(
-      text,
-      model,
-      maxTokens,
-      apiKey!,
-      systemPrompt,
-      config.apiBase || undefined,
-    );
-  } else if (provider === "openai") {
-    return callOpenAI(
-      text,
+    const result = await complete({
+      service: config.service,
       model,
-      maxTokens,
-      apiKey!,
+      prompt: text,
       systemPrompt,
-      config.apiBase || undefined,
-    );
-  } else if (provider === "ollama") {
-    if (!config.apiBase) {
+      maxTokens: options?.maxTokens || config.maxTokens,
+      temperature: 0.3,
+    });
+
+    const insights = extractJson(result.text);
+
+    if (!insights) {
       return {
-        error: `No api_base configured for ollama. Set api_base in ~/.config/llm/config.toml`,
+        error: "Failed to parse insights from response",
+        rawText: result.text,
       };
     }
-    return callOllama(text, model, maxTokens, config.apiBase, systemPrompt);
-  } else {
+
     return {
-      error: `Unknown provider: ${provider}. Supported: anthropic, openai, ollama`,
+      insights,
+      rawText: result.text,
+      model: result.model,
+      tokens_used: result.tokens.output,
     };
+  } catch (err) {
+    const error = err instanceof Error ? err.message : String(err);
+    return { error };
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voidwire/llm-summarize",
-  "version": "3.6.0",
+  "version": "3.8.0",
   "description": "Structured session insight extraction for knowledge systems",
   "type": "module",
   "main": "./index.ts",
@@ -40,6 +40,9 @@
   "engines": {
     "bun": ">=1.0.0"
   },
+  "dependencies": {
+    "@voidwire/llm-core": "0.2.0"
+  },
   "scripts": {
     "test": "bun test"
   }