modelstat 0.0.24 → 0.0.26

package/dist/cli.mjs

@@ -1,4 +1,13 @@
 #!/usr/bin/env node
+{
+  const [__msMaj, __msMin] = process.versions.node.split('.').map(Number);
+  if (__msMaj < 20 || (__msMaj === 20 && __msMin < 18)) {
+    process.stderr.write(`modelstat requires Node \u2265 20.18 (you have ${process.version}).\n`);
+    process.stderr.write('Install Node 20+: https://nodejs.org\n');
+    process.stderr.write('Debian/Ubuntu: curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash - && sudo apt-get install -y nodejs\n');
+    process.exit(1);
+  }
+}
 import { createRequire as __modelstatCR } from "node:module";
 const require = __modelstatCR(import.meta.url);
 var __create = Object.create;
@@ -21122,8 +21131,8 @@ var require_snapshot_utils = __commonJS({
 var require_snapshot_recorder = __commonJS({
   "../../node_modules/.pnpm/undici@7.25.0/node_modules/undici/lib/mock/snapshot-recorder.js"(exports, module) {
     "use strict";
-    var { writeFile, readFile, mkdir } = __require("fs/promises");
-    var { dirname: dirname7, resolve: resolve6 } = __require("path");
+    var { writeFile, readFile, mkdir: mkdir2 } = __require("fs/promises");
+    var { dirname: dirname8, resolve: resolve6 } = __require("path");
     var { setTimeout: setTimeout2, clearTimeout: clearTimeout2 } = __require("timers");
     var { InvalidArgumentError, UndiciError } = require_errors();
     var { hashId, isUrlExcludedFactory, normalizeHeaders, createHeaderFilters } = require_snapshot_utils();
@@ -21354,7 +21363,7 @@ var require_snapshot_recorder = __commonJS({
           throw new InvalidArgumentError("Snapshot path is required");
         }
         const resolvedPath = resolve6(path5);
-        await mkdir(dirname7(resolvedPath), { recursive: true });
+        await mkdir2(dirname8(resolvedPath), { recursive: true });
         const data = Array.from(this.#snapshots.entries()).map(([hash, snapshot]) => ({
           hash,
           snapshot
@@ -44231,14 +44240,15 @@ var init_queue = __esm({
 });
 
 // ../../packages/companion-core/src/pipeline/prompts.ts
-var OLLAMA_CHAT_MODEL, OLLAMA_EMBED_MODEL, SUMMARISER_SYSTEM_PROMPT, SUMMARISER_MAX_TOKENS, SUMMARISER_TEMPERATURE, QWEN_CHARS_PER_TOKEN;
+var OLLAMA_CHAT_MODEL, OLLAMA_EMBED_MODEL, SUMMARISER_SYSTEM_PROMPT, SUMMARISER_MAX_TOKENS, ABSTRACT_OUTPUT_MAX_CHARS, SUMMARISER_TEMPERATURE, QWEN_CHARS_PER_TOKEN;
 var init_prompts = __esm({
   "../../packages/companion-core/src/pipeline/prompts.ts"() {
     "use strict";
     OLLAMA_CHAT_MODEL = "qwen3.5:0.8b";
     OLLAMA_EMBED_MODEL = "bge-small-en-v1.5";
-    SUMMARISER_SYSTEM_PROMPT = "You summarise an AI coding session in ONE sentence, \u2264 240 characters. If the user message includes sampled conversation excerpts, base your summary on what the developer was actually working on (the substance \u2014 what was being built, debugged, refactored, or designed). If only metadata is given, paraphrase the metadata. Never quote the excerpts verbatim. No PII, no code literals, no file paths, no API keys. Reply with only the sentence.";
-    SUMMARISER_MAX_TOKENS = 120;
+    SUMMARISER_SYSTEM_PROMPT = "Write the SHORTEST paragraph (1-3 sentences) that captures EXACTLY what was ACHIEVED in this coding segment, packed with the concrete domain keywords the developer used. Lead with an outcome verb \u2014 shipped, fixed, migrated, ramped, wired, diagnosed, refactored, designed, deployed, reverted, instrumented. Name the specific things touched: feature names, frameworks, components, bug classes, decisions. Density beats length: a 50-char sentence that names the actual feature beats 200 chars of filler. Skip narration ('the developer'), skip vague verbs ('worked on', 'explored', 'looked into'), skip preamble. If only metadata is given, name the project + tool + visible action concisely. Never quote excerpts verbatim. No PII, no API keys, no file paths, no code literals. Reply with ONLY the paragraph.";
+    SUMMARISER_MAX_TOKENS = 160;
+    ABSTRACT_OUTPUT_MAX_CHARS = 400;
     SUMMARISER_TEMPERATURE = 0.2;
     QWEN_CHARS_PER_TOKEN = 3.3;
   }
@@ -44278,6 +44288,7 @@ async function buildForOneSession(sessionId, events, adapters2) {
   const boundaries = [];
   let runStart = 0;
   let runStartMs = Date.parse(sorted[0].ts);
+  let runChars = (sorted[0].content_excerpt ?? "").length;
   for (let i = 1; i < sorted.length; i++) {
     const prev2 = sorted[i - 1];
     const cur = sorted[i];
@@ -44288,6 +44299,7 @@ async function buildForOneSession(sessionId, events, adapters2) {
     if (gap >= SEGMENT_TIME_GAP_MS) split = true;
     else if (turnsInRun >= SEGMENT_MAX_TURNS) split = true;
     else if (runMs >= SEGMENT_MAX_DURATION_MS) split = true;
+    else if (runChars >= SEGMENT_MAX_CONTENT_CHARS) split = true;
     else if (turnEmbeddings[i - 1].length > 0 && turnEmbeddings[i].length > 0 && cosineDistance(turnEmbeddings[i - 1], turnEmbeddings[i]) > SEGMENT_TOPIC_THRESHOLD) {
       split = true;
     }
@@ -44295,6 +44307,9 @@ async function buildForOneSession(sessionId, events, adapters2) {
       boundaries.push(i);
       runStart = i;
       runStartMs = Date.parse(cur.ts);
+      runChars = (cur.content_excerpt ?? "").length;
+    } else {
+      runChars += (cur.content_excerpt ?? "").length;
     }
   }
   boundaries.push(sorted.length);
@@ -44355,13 +44370,16 @@ async function summariseSlice(sessionId, slice, adapters2) {
 Sampled excerpts from the conversation (already redacted of PII and secrets):
 ${excerptBlock}
 
-Write ONE sentence (\u2264240 chars) describing what the human was working on. Focus on the substance \u2014 what was being built, debugged, or designed. No quotes, no PII, no code literals, no file paths.` : `Session context: ${promptFacts || "generic coding session"}.
-Write one sentence describing what the human was doing.`;
-  let rawAbstract;
-  try {
-    rawAbstract = await adapters2.summarize({ prompt, maxTokens: 80 });
-  } catch {
-    rawAbstract = promptFacts || `${first.tool} session with ${slice.length} turns`;
+Write the SHORTEST keyword-dense paragraph (1-3 sentences, \u2264${ABSTRACT_OUTPUT_MAX_CHARS} chars) naming exactly what was achieved. Lead with an outcome verb. Pack with concrete domain keywords (frameworks, features, components, decisions). Skip narration and filler.` : `Session context: ${promptFacts || "generic coding session"}.
+Write the shortest possible paragraph naming the project + tool + visible action. Skip filler.`;
+  const rawAbstract = await adapters2.summarize({
+    prompt,
+    maxTokens: SUMMARISER_MAX_TOKENS
+  });
+  if (!rawAbstract || rawAbstract.trim().length === 0) {
+    throw new Error(
+      `summariser returned empty abstract for session ${sessionId} (${slice.length} turns) \u2014 check that the configured summariser is healthy`
+    );
   }
   const regexPass = redact(rawAbstract);
   let abstractText = regexPass.text;
@@ -44414,7 +44432,11 @@ Write one sentence describing what the human was doing.`;
     tool: first.tool,
     started_at: first.ts,
     ended_at: last.ts,
-    abstract: redacted.text.slice(0, ABSTRACT_MAX_CHARS),
+    // Slice to the user-visible cap (ABSTRACT_OUTPUT_MAX_CHARS, default
+    // 400) — well below the 512 storage cap. Models occasionally
+    // overshoot the prompt's "≤N chars" instruction; the slice is the
+    // hard guarantee the dashboard relies on.
+    abstract: redacted.text.slice(0, ABSTRACT_OUTPUT_MAX_CHARS),
     tokens,
     tags,
     // counts is `Record<string, number>` after the optional model
@@ -44478,18 +44500,20 @@ function inferEnvironment(branch) {
   if (b === "dev" || b === "develop" || b.startsWith("dev/")) return "Dev";
   return null;
 }
-var SEGMENT_TIME_GAP_MS, SEGMENT_TOPIC_THRESHOLD, SEGMENT_MAX_TURNS, SEGMENT_MAX_DURATION_MS, ABSTRACT_MAX_CHARS;
+var SEGMENT_TIME_GAP_MS, SEGMENT_TOPIC_THRESHOLD, SEGMENT_MAX_TURNS, SEGMENT_MAX_DURATION_MS, SEGMENT_MAX_CONTENT_CHARS, ABSTRACT_MAX_CHARS;
 var init_pipeline = __esm({
   "../../packages/companion-core/src/pipeline/index.ts"() {
     "use strict";
     init_redact();
     init_ids();
+    init_prompts();
     init_redact();
     init_prompts();
     SEGMENT_TIME_GAP_MS = 15 * 6e4;
     SEGMENT_TOPIC_THRESHOLD = 0.35;
     SEGMENT_MAX_TURNS = 100;
     SEGMENT_MAX_DURATION_MS = 30 * 6e4;
+    SEGMENT_MAX_CONTENT_CHARS = 12e3;
     ABSTRACT_MAX_CHARS = 512;
   }
 });
@@ -44511,11 +44535,123 @@ var init_src3 = __esm({
 // ../../packages/companion-core/src/node/file-queue-store.ts
 import { promises as fs3 } from "fs";
 import { dirname as dirname3 } from "path";
-var SENT_TTL_MS;
+var SENT_TTL_MS, FileQueueStore;
 var init_file_queue_store = __esm({
   "../../packages/companion-core/src/node/file-queue-store.ts"() {
     "use strict";
     SENT_TTL_MS = 24 * 60 * 60 * 1e3;
+    FileQueueStore = class {
+      filePath;
+      items = /* @__PURE__ */ new Map();
+      loaded = false;
+      writing = null;
+      constructor(filePath) {
+        this.filePath = filePath;
+      }
+      /** Lazily load the snapshot on first access so construction stays
+       * synchronous (matches the old SqliteQueueStore contract). */
+      async ensureLoaded() {
+        if (this.loaded) return;
+        try {
+          const raw = await fs3.readFile(this.filePath, "utf8");
+          const parsed = JSON.parse(raw);
+          if (parsed.version === 1 && parsed.items) {
+            const now = Date.now();
+            for (const [k, v] of Object.entries(parsed.items)) {
+              if (v.synced && now - v.last_event_ts_ms > SENT_TTL_MS) continue;
+              this.items.set(k, v);
+            }
+          }
+        } catch (err) {
+          if (err.code !== "ENOENT") {
+            try {
+              const bak = await fs3.readFile(`${this.filePath}.bak`, "utf8");
+              const parsed = JSON.parse(bak);
+              if (parsed.version === 1 && parsed.items) {
+                for (const [k, v] of Object.entries(parsed.items)) {
+                  this.items.set(k, v);
+                }
+              }
+            } catch {
+            }
+          }
+        }
+        this.loaded = true;
+      }
+      /** Serialize writes so overlapping mutations don't race on the
+       * temp-file / rename swap. writing is a single-slot promise; each
+       * caller awaits the previous write and then enqueues its own. */
+      async persist() {
+        const prior = this.writing;
+        const run = async () => {
+          await prior?.catch(() => void 0);
+          const snap = {
+            version: 1,
+            items: Object.fromEntries(this.items.entries())
+          };
+          const tmp = `${this.filePath}.tmp`;
+          await fs3.mkdir(dirname3(this.filePath), { recursive: true });
+          await fs3.writeFile(tmp, JSON.stringify(snap), "utf8");
+          try {
+            await fs3.rename(this.filePath, `${this.filePath}.bak`);
+          } catch (err) {
+            if (err.code !== "ENOENT") throw err;
+          }
+          await fs3.rename(tmp, this.filePath);
+        };
+        this.writing = run();
+        try {
+          await this.writing;
+        } finally {
+          this.writing = null;
+        }
+      }
+      async put(item) {
+        await this.ensureLoaded();
+        if (this.items.has(item.source_event_id)) return;
+        this.items.set(item.source_event_id, item);
+        await this.persist();
+      }
+      async has(sourceEventId2) {
+        await this.ensureLoaded();
+        return this.items.has(sourceEventId2);
+      }
+      async listUnsentBySession() {
+        await this.ensureLoaded();
+        const out = /* @__PURE__ */ new Map();
+        const sorted = [...this.items.values()].filter((i) => !i.synced).sort((a, b) => {
+          if (a.session_id !== b.session_id)
+            return a.session_id < b.session_id ? -1 : 1;
+          return a.last_event_ts_ms - b.last_event_ts_ms;
+        });
+        for (const it of sorted) {
+          const arr = out.get(it.session_id) ?? [];
+          arr.push(it);
+          out.set(it.session_id, arr);
+        }
+        return out;
+      }
+      async markSent(sourceEventIds, batchId2) {
+        if (sourceEventIds.length === 0) return;
+        await this.ensureLoaded();
+        for (const id of sourceEventIds) {
+          const it = this.items.get(id);
+          if (!it) continue;
+          it.synced = true;
+          it.sent_batch_id = batchId2;
+        }
+        await this.persist();
+      }
+      async countUnsent() {
+        await this.ensureLoaded();
+        let n = 0;
+        for (const it of this.items.values()) if (!it.synced) n++;
+        return n;
+      }
+      /** Compat stub — Sqlite store exposed close(); nothing to release here. */
+      close() {
+      }
+    };
   }
 });
 
@@ -44581,17 +44717,192 @@ var init_ollama = __esm({
   }
 });
 
+// ../../packages/companion-core/src/node/llama.ts
+import { mkdir } from "fs/promises";
+import { existsSync as existsSync5 } from "fs";
+import { homedir as homedir4 } from "os";
+import { dirname as dirname4, join as join3 } from "path";
+function defaultLlamaConfig() {
+  const env2 = globalThis.process?.env ?? {};
+  const modelsDir = env2.MODELSTAT_MODELS_DIR ?? join3(homedir4(), ".modelstat", "models");
+  const modelUrl = env2.MODELSTAT_LLAMA_MODEL_URL ?? DEFAULT_LLAMA_MODEL_URL;
+  const modelPath = env2.MODELSTAT_LLAMA_MODEL_PATH ?? join3(modelsDir, basenameFromUrl(modelUrl));
+  return {
+    modelPath,
+    modelUrl,
+    modelsDir,
+    // Qwen3.5-4B has a 128K native context; we only need enough to
+    // fit our prompt (~2 KB) plus thinking budget plus answer.
+    // 4096 is plenty and keeps memory footprint reasonable on
+    // older machines.
+    contextSize: Number(env2.MODELSTAT_LLAMA_CONTEXT ?? 4096)
+  };
+}
+function stripThinking(text) {
+  return text.replace(/<think>[\s\S]*?<\/think>/gi, "").replace(/<think>[\s\S]*$/i, "").trim();
+}
+function basenameFromUrl(url) {
+  const clean = url.split("?")[0].split("#")[0];
+  const parts = clean.split("/");
+  return parts[parts.length - 1] || "model.gguf";
+}
+async function ensureLlamaModel(cfg = defaultLlamaConfig()) {
+  if (existsSync5(cfg.modelPath)) return cfg.modelPath;
+  await mkdir(dirname4(cfg.modelPath), { recursive: true });
+  const res = await fetch(cfg.modelUrl);
+  if (!res.ok || !res.body) {
+    throw new Error(
+      `model download failed: ${res.status} ${res.statusText} (${cfg.modelUrl})`
+    );
+  }
+  const total = Number(res.headers.get("content-length") ?? 0);
+  const totalMb = total > 0 ? (total / (1024 * 1024)).toFixed(0) : "?";
+  console.log(
+    `[modelstat] downloading summariser model (~${totalMb} MB) \u2192 ${cfg.modelPath}`
+  );
+  const tmp = `${cfg.modelPath}.partial`;
+  const { createWriteStream: createWriteStream2 } = await import("fs");
+  const { Readable: Readable3 } = await import("stream");
+  const { rename } = await import("fs/promises");
+  const out = createWriteStream2(tmp);
+  const isTty = Boolean(
+    process.stdout.isTTY
+  );
+  let received = 0;
+  let lastLog = 0;
+  let lastBytes = 0;
+  let lastTimeForRate = Date.now();
+  const startTime = Date.now();
+  const renderProgress = (final = false) => {
+    const now = Date.now();
+    const dt = (now - lastTimeForRate) / 1e3;
+    const dBytes = received - lastBytes;
+    const rateMbps = dt > 0 ? dBytes / 1024 / 1024 / dt : 0;
+    lastBytes = received;
+    lastTimeForRate = now;
+    const mb = (received / (1024 * 1024)).toFixed(1);
+    const pct = total > 0 ? (received / total * 100).toFixed(1) : "?";
+    const eta = total > 0 && rateMbps > 0 ? Math.max(0, Math.round((total - received) / 1024 / 1024 / rateMbps)) : null;
+    const etaStr = eta != null ? ` \xB7 ETA ${eta}s` : "";
+    const elapsed = ((now - startTime) / 1e3).toFixed(0);
+    const line = `[modelstat]   ${mb} / ${totalMb} MB (${pct}%) \xB7 ${rateMbps.toFixed(1)} MB/s${etaStr} \xB7 ${elapsed}s`;
+    if (isTty && !final) {
+      process.stdout.write(`\r${line}\x1B[K`);
+    } else {
+      console.log(line);
+    }
+  };
+  const nodeStream = Readable3.fromWeb(
+    res.body
+  );
+  nodeStream.on("data", (chunk) => {
+    received += chunk.length;
+    const now = Date.now();
+    if (now - lastLog > (isTty ? 200 : 2e3)) {
+      renderProgress();
+      lastLog = now;
+    }
+  });
+  await new Promise((resolve6, reject) => {
+    nodeStream.pipe(out);
+    out.on("finish", () => resolve6());
+    out.on("error", reject);
+    nodeStream.on("error", reject);
+  });
+  renderProgress(true);
+  if (isTty) process.stdout.write("\n");
+  await rename(tmp, cfg.modelPath);
+  console.log(`[modelstat] summariser model ready (${cfg.modelPath})`);
+  return cfg.modelPath;
+}
+async function loadOnce(cfg) {
+  if (loaded) return loaded;
+  if (loadPromise) return loadPromise;
+  loadPromise = (async () => {
+    const llamaMod = await import("node-llama-cpp");
+    const modelPath = await ensureLlamaModel(cfg);
+    const llama = await llamaMod.getLlama();
+    const model = await llama.loadModel({ modelPath });
+    const context = await model.createContext({ contextSize: cfg.contextSize });
+    const session = new llamaMod.LlamaChatSession({
+      contextSequence: context.getSequence(),
+      systemPrompt: SUMMARISER_SYSTEM_PROMPT
+    });
+    loaded = { session };
+    return loaded;
+  })();
+  try {
+    return await loadPromise;
+  } catch (err) {
+    loadPromise = null;
+    throw err;
+  }
+}
+function llamaSummarize(cfg = defaultLlamaConfig()) {
+  return async ({ prompt, maxTokens }) => {
+    const { session } = await loadOnce(cfg);
+    const run = inflight.then(async () => {
+      session.resetChatHistory();
+      void maxTokens;
+      const raw = await session.prompt(prompt, {
+        temperature: SUMMARISER_TEMPERATURE,
+        maxTokens: LLAMA_MAX_TOKENS
+      });
+      const stripped = stripThinking(raw ?? "");
+      if (stripped.length === 0) {
+        throw new Error(
+          `bundled summariser produced no answer text after stripping <think> blocks (raw length=${(raw ?? "").length}). The thinking budget may be too low or the model template is misconfigured.`
+        );
+      }
+      return stripped.slice(0, 240);
+    });
+    inflight = run.catch(() => void 0);
+    return run;
+  };
+}
+var DEFAULT_LLAMA_MODEL_URL, LLAMA_MAX_TOKENS, loaded, loadPromise, inflight;
+var init_llama = __esm({
+  "../../packages/companion-core/src/node/llama.ts"() {
+    "use strict";
+    init_prompts();
+    DEFAULT_LLAMA_MODEL_URL = "https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf";
+    LLAMA_MAX_TOKENS = 1024;
+    loaded = null;
+    loadPromise = null;
+    inflight = Promise.resolve();
+  }
+});
+
 // ../../packages/companion-core/src/node/index.ts
+var node_exports = {};
+__export(node_exports, {
+  DEFAULT_LLAMA_MODEL_URL: () => DEFAULT_LLAMA_MODEL_URL,
+  FileQueueStore: () => FileQueueStore,
+  SqliteQueueStore: () => FileQueueStore,
+  defaultLlamaConfig: () => defaultLlamaConfig,
+  defaultOllamaConfig: () => defaultOllamaConfig,
+  ensureLlamaModel: () => ensureLlamaModel,
+  llamaSummarize: () => llamaSummarize,
+  ollamaEmbed: () => ollamaEmbed,
+  ollamaSummarize: () => ollamaSummarize,
+  ollamaTokenize: () => ollamaTokenize
+});
 var init_node2 = __esm({
   "../../packages/companion-core/src/node/index.ts"() {
     "use strict";
     init_file_queue_store();
     init_file_queue_store();
     init_ollama();
+    init_llama();
   }
 });
 
 // src/pipeline.ts
+var pipeline_exports = {};
+__export(pipeline_exports, {
+  buildSegments: () => buildSegments,
+  preflightSummariser: () => preflightSummariser
+});
 async function probeOllama(baseUrl) {
   try {
     const ctrl = new AbortController();
@@ -44606,40 +44917,58 @@ async function probeOllama(baseUrl) {
     return false;
   }
 }
-function fallbackAdapters() {
+function bundledAdapters() {
   return {
     embed: async () => [],
-    // Throw → companion-core/summariseSlice's try/catch falls back to
-    // the `promptFacts` string which is what we want here.
-    summarize: async () => {
-      throw new Error("ollama_unavailable");
-    },
+    summarize: llamaSummarize(defaultLlamaConfig()),
     tokenize: (text) => Math.max(1, Math.ceil(text.length / 4))
   };
 }
 async function getAdapters() {
   if (adapters && probed) return adapters;
-  const cfg = defaultOllamaConfig();
-  const up = await probeOllama(cfg.baseUrl);
+  const ollamaCfg = defaultOllamaConfig();
+  const ollamaUp = await probeOllama(ollamaCfg.baseUrl);
   probed = true;
-  if (up) {
-    console.log(`[modelstat] ollama up at ${cfg.baseUrl} \u2014 using LLM pipeline`);
+  if (ollamaUp) {
+    console.log(
+      `[modelstat] ollama up at ${ollamaCfg.baseUrl} \u2014 using ${ollamaCfg.chatModel} for summarisation`
+    );
     adapters = {
-      embed: ollamaEmbed(cfg),
-      summarize: ollamaSummarize(cfg),
+      embed: ollamaEmbed(ollamaCfg),
+      summarize: ollamaSummarize(ollamaCfg),
       tokenize: ollamaTokenize()
     };
-  } else {
-    console.log(
-      `[modelstat] ollama not reachable at ${cfg.baseUrl} \u2014 using fallback pipeline (server does classification)`
+    return adapters;
+  }
+  try {
+    await import("node-llama-cpp");
+  } catch (err) {
+    throw new Error(
+      `modelstat agent can't start: ollama is not reachable AND the bundled summariser (node-llama-cpp) failed to load. Either start ollama (\`ollama serve\`) with \`ollama pull ${ollamaCfg.chatModel}\`, or reinstall the agent so the native binding is rebuilt for this platform. Underlying error: ${err.message}`
     );
-    adapters = fallbackAdapters();
   }
+  console.log(
+    "[modelstat] using bundled local summariser (Qwen3.5-4B, runs on this machine)"
+  );
+  adapters = bundledAdapters();
   return adapters;
 }
 async function buildSegments(events) {
   return buildSegmentsForSession(events, await getAdapters());
 }
+async function preflightSummariser() {
+  const a = await getAdapters();
+  const out = await a.summarize({
+    prompt: 'Session context: smoke test. Sampled excerpts:\n  [turn 1] "hello world"\nWrite ONE sentence (\u2264240 chars) describing what the human was doing.',
+    maxTokens: 32
+  });
+  if (!out || out.trim().length === 0) {
+    throw new Error(
+      "summariser preflight returned empty output \u2014 the configured summariser is reachable but produced no text. Check the model is loaded."
+    );
+  }
+  return out.length > 60 ? `${out.slice(0, 57)}\u2026` : out;
+}
 var adapters, probed;
 var init_pipeline2 = __esm({
   "src/pipeline.ts"() {
@@ -44653,23 +44982,23 @@ var init_pipeline2 = __esm({
 
 // src/scan.ts
 import { readdir, stat as stat2 } from "fs/promises";
-import { homedir as homedir4 } from "os";
-import { join as join3 } from "path";
+import { homedir as homedir5 } from "os";
+import { join as join4 } from "path";
 async function scanAll(cb = {}) {
   const deviceId = state.deviceId;
   if (!deviceId) throw new Error("agent not enrolled \u2014 run `register` first");
   const jobs = [];
   try {
-    const base = join3(homedir4(), ".claude/projects");
+    const base = join4(homedir5(), ".claude/projects");
     const projects = await readdir(base).catch(() => []);
     for (const p of projects) {
-      const dir = join3(base, p);
+      const dir = join4(base, p);
       const ds = await stat2(dir).catch(() => null);
       if (!ds?.isDirectory()) continue;
       const files = await readdir(dir);
       for (const f of files) {
         if (!f.endsWith(".jsonl")) continue;
-        const full = join3(dir, f);
+        const full = join4(dir, f);
         jobs.push({
           path: full,
           parse: async () => {
@@ -44683,17 +45012,17 @@ async function scanAll(cb = {}) {
     console.warn("claude scan skipped:", e.message);
   }
   try {
-    const base = join3(homedir4(), ".codex/sessions");
+    const base = join4(homedir5(), ".codex/sessions");
     const years = await readdir(base).catch(() => []);
     for (const y of years) {
-      const months = await readdir(join3(base, y)).catch(() => []);
+      const months = await readdir(join4(base, y)).catch(() => []);
       for (const m of months) {
-        const days = await readdir(join3(base, y, m)).catch(() => []);
+        const days = await readdir(join4(base, y, m)).catch(() => []);
         for (const d of days) {
-          const files = await readdir(join3(base, y, m, d)).catch(() => []);
+          const files = await readdir(join4(base, y, m, d)).catch(() => []);
           for (const f of files) {
             if (!f.startsWith("rollout-") || !f.endsWith(".jsonl")) continue;
-            const full = join3(base, y, m, d, f);
+            const full = join4(base, y, m, d, f);
             jobs.push({
               path: full,
               parse: async () => {
@@ -44775,7 +45104,7 @@ var init_scan = __esm({
 // src/lock.ts
 import {
   closeSync,
-  existsSync as existsSync6,
+  existsSync as existsSync7,
   mkdirSync as mkdirSync3,
   openSync,
   readFileSync as readFileSync3,
@@ -44784,8 +45113,8 @@ import {
   writeFileSync as writeFileSync4,
   writeSync
 } from "fs";
-import { homedir as homedir6 } from "os";
-import { join as join5 } from "path";
+import { homedir as homedir7 } from "os";
+import { join as join6 } from "path";
 function isProcessAlive(pid) {
   if (!pid || pid <= 0) return false;
   try {
@@ -44886,8 +45215,8 @@ var LOCK_DIR, LOCK_FILE;
 var init_lock = __esm({
   "src/lock.ts"() {
     "use strict";
-    LOCK_DIR = join5(homedir6(), ".modelstat");
-    LOCK_FILE = join5(LOCK_DIR, "daemon.lock");
+    LOCK_DIR = join6(homedir7(), ".modelstat");
+    LOCK_FILE = join6(LOCK_DIR, "daemon.lock");
   }
 });
 
@@ -45624,9 +45953,9 @@ var init_handler = __esm({
         if (this.fsw.closed) {
           return;
         }
-        const dirname7 = sysPath.dirname(file);
+        const dirname8 = sysPath.dirname(file);
         const basename4 = sysPath.basename(file);
-        const parent = this.fsw._getWatchedDir(dirname7);
+        const parent = this.fsw._getWatchedDir(dirname8);
         let prevStats = stats;
         if (parent.has(basename4))
           return;
@@ -45653,7 +45982,7 @@ var init_handler = __esm({
                 prevStats = newStats2;
               }
             } catch (error) {
-              this.fsw._remove(dirname7, basename4);
+              this.fsw._remove(dirname8, basename4);
             }
           } else if (parent.has(basename4)) {
             const at = newStats.atimeMs;
@@ -46617,7 +46946,7 @@ __export(daemon_exports, {
   setProgress: () => setProgress,
   setQueue: () => setQueue
 });
-import { existsSync as existsSync7, statSync as statSync2 } from "fs";
+import { existsSync as existsSync8, statSync as statSync2 } from "fs";
 function setPhase(phase, message) {
   status.phase = phase;
   status.message = message ?? null;
@@ -46736,24 +47065,33 @@ async function runDaemon(opts = {}) {
   const hb = setInterval(() => void sendHeartbeat(), HEARTBEAT_INTERVAL_MS);
   hb.unref();
   void sendHeartbeat();
+  try {
+    setPhase("starting", "Preflight: summariser");
+    const { preflightSummariser: preflightSummariser2 } = await Promise.resolve().then(() => (init_pipeline2(), pipeline_exports));
+    const sample = await preflightSummariser2();
+    console.log(`[modelstat] summariser preflight ok: "${sample}"`);
+  } catch (err) {
+    setPhase("error", `summariser preflight failed: ${err.message}`);
+    throw err;
+  }
   await runDiscovery();
   await runScanCycle("startup");
   const chokidar = (await Promise.resolve().then(() => (init_esm2(), esm_exports))).default;
-  const { homedir: homedir8, platform: platform5 } = await import("os");
-  const { join: join9 } = await import("path");
-  const home2 = homedir8();
+  const { homedir: homedir9, platform: platform5 } = await import("os");
+  const { join: join10 } = await import("path");
+  const home2 = homedir9();
   const dirs = [
-    join9(home2, ".claude/projects"),
-    join9(home2, ".codex/sessions"),
-    join9(home2, ".cursor/ai-tracking"),
-    join9(home2, ".gemini"),
+    join10(home2, ".claude/projects"),
+    join10(home2, ".codex/sessions"),
+    join10(home2, ".cursor/ai-tracking"),
+    join10(home2, ".gemini"),
     ...platform5() === "darwin" ? [
-      join9(home2, "Library/Application Support/Cursor/User/workspaceStorage"),
-      join9(home2, "Library/Application Support/Claude")
+      join10(home2, "Library/Application Support/Cursor/User/workspaceStorage"),
+      join10(home2, "Library/Application Support/Claude")
     ] : [
-      join9(home2, ".config/Cursor/User/workspaceStorage")
+      join10(home2, ".config/Cursor/User/workspaceStorage")
     ]
-  ].filter((p) => existsSync7(p) && statSync2(p).isDirectory());
+  ].filter((p) => existsSync8(p) && statSync2(p).isDirectory());
   setPhase("watching", `Watching ${dirs.length} directories`);
   const watcher = chokidar.watch(dirs, {
     persistent: true,
@@ -46820,37 +47158,37 @@ var watch_exports = {};
 __export(watch_exports, {
   watchForever: () => watchForever
 });
-import { existsSync as existsSync8 } from "fs";
-import { homedir as homedir7, platform as platform3 } from "os";
-import { join as join8 } from "path";
+import { existsSync as existsSync9 } from "fs";
+import { homedir as homedir8, platform as platform3 } from "os";
+import { join as join9 } from "path";
 function resolveWatchDirs() {
-  const home2 = homedir7();
-  const xdgConfig = process.env.XDG_CONFIG_HOME ?? join8(home2, ".config");
-  const xdgData = process.env.XDG_DATA_HOME ?? join8(home2, ".local/share");
+  const home2 = homedir8();
+  const xdgConfig = process.env.XDG_CONFIG_HOME ?? join9(home2, ".config");
+  const xdgData = process.env.XDG_DATA_HOME ?? join9(home2, ".local/share");
   const candidates = [
     // universal (default HOME-rooted CLI data dirs)
-    join8(home2, ".claude/projects"),
-    join8(home2, ".codex/sessions"),
-    join8(home2, ".cursor/ai-tracking"),
-    join8(home2, ".gemini"),
-    join8(home2, ".aider"),
+    join9(home2, ".claude/projects"),
+    join9(home2, ".codex/sessions"),
+    join9(home2, ".cursor/ai-tracking"),
+    join9(home2, ".gemini"),
+    join9(home2, ".aider"),
     // XDG / Linux
-    join8(xdgConfig, "claude/projects"),
-    join8(xdgConfig, "codex/sessions"),
-    join8(xdgConfig, "Cursor/User/workspaceStorage"),
-    join8(xdgConfig, "Code/User/workspaceStorage"),
-    join8(xdgConfig, "Code - Insiders/User/workspaceStorage"),
-    join8(xdgData, "claude/projects"),
+    join9(xdgConfig, "claude/projects"),
+    join9(xdgConfig, "codex/sessions"),
+    join9(xdgConfig, "Cursor/User/workspaceStorage"),
+    join9(xdgConfig, "Code/User/workspaceStorage"),
+    join9(xdgConfig, "Code - Insiders/User/workspaceStorage"),
+    join9(xdgData, "claude/projects"),
     // macOS
     ...platform3() === "darwin" ? [
-      join8(home2, "Library/Application Support/Cursor/User/workspaceStorage"),
-      join8(home2, "Library/Application Support/Claude"),
-      join8(home2, "Library/Application Support/Code/User/workspaceStorage"),
-      join8(home2, "Library/Application Support/Windsurf/User/workspaceStorage"),
-      join8(home2, "Library/Application Support/Zed")
+      join9(home2, "Library/Application Support/Cursor/User/workspaceStorage"),
+      join9(home2, "Library/Application Support/Claude"),
+      join9(home2, "Library/Application Support/Code/User/workspaceStorage"),
+      join9(home2, "Library/Application Support/Windsurf/User/workspaceStorage"),
+      join9(home2, "Library/Application Support/Zed")
     ] : []
   ];
-  return Array.from(new Set(candidates)).filter((p) => existsSync8(p));
+  return Array.from(new Set(candidates)).filter((p) => existsSync9(p));
 }
 async function safeScan(reason) {
   if (scanning) {
@@ -46923,30 +47261,30 @@ import { createInterface as createInterface3 } from "readline";
 import { spawnSync } from "child_process";
 import {
   copyFileSync,
-  existsSync as existsSync5,
+  existsSync as existsSync6,
   mkdirSync as mkdirSync2,
   unlinkSync,
   writeFileSync as writeFileSync3
 } from "fs";
-import { homedir as homedir5, platform as platform2, userInfo } from "os";
-import { dirname as dirname4, join as join4 } from "path";
+import { homedir as homedir6, platform as platform2, userInfo } from "os";
+import { dirname as dirname5, join as join5 } from "path";
 import { fileURLToPath as fileURLToPath2 } from "url";
 var SERVICE_LABEL = "ai.modelstat.agent";
 var SYSTEMD_UNIT = "modelstat";
 function home() {
-  return homedir5();
+  return homedir6();
 }
 function stateDir() {
-  return join4(home(), ".modelstat");
+  return join5(home(), ".modelstat");
 }
 function binDir() {
-  return join4(stateDir(), "bin");
+  return join5(stateDir(), "bin");
 }
 function logDir() {
-  return join4(stateDir(), "logs");
+  return join5(stateDir(), "logs");
 }
 function installedCliPath() {
-  return join4(binDir(), "modelstat.mjs");
+  return join5(binDir(), "modelstat.mjs");
 }
 function runningCliPath() {
   return fileURLToPath2(import.meta.url).replace(/service\.(mjs|js|ts)$/, "cli.mjs");
@@ -46956,7 +47294,7 @@ function installBundle() {
   mkdirSync2(logDir(), { recursive: true });
   const src = runningCliPath();
   const dest = installedCliPath();
-  if (!existsSync5(src)) {
+  if (!existsSync6(src)) {
     throw new Error(
       `Can't find the CLI bundle to install from (${src}). Are you running a local dev build?`
     );
@@ -46968,21 +47306,21 @@ function nodeBinary() {
   return process.execPath;
 }
 function plistPath() {
-  return join4(home(), "Library", "LaunchAgents", `${SERVICE_LABEL}.plist`);
+  return join5(home(), "Library", "LaunchAgents", `${SERVICE_LABEL}.plist`);
 }
 function locateTrayExecutable() {
   const candidates = [
-    join4(home(), "Applications", "ModelstatTray.app", "Contents", "MacOS", "modelstat-tray"),
+    join5(home(), "Applications", "ModelstatTray.app", "Contents", "MacOS", "modelstat-tray"),
     "/Applications/ModelstatTray.app/Contents/MacOS/modelstat-tray"
   ];
   for (const p of candidates) {
-    if (existsSync5(p)) return p;
+    if (existsSync6(p)) return p;
   }
   return null;
 }
 function writePlist(cliPath) {
   const p = plistPath();
-  mkdirSync2(dirname4(p), { recursive: true });
+  mkdirSync2(dirname5(p), { recursive: true });
   const tray = locateTrayExecutable();
   const programArgs = tray ? `    <string>${tray}</string>` : [
     `    <string>${nodeBinary()}</string>`,
@@ -47002,8 +47340,8 @@ ${programArgs}
   <key>KeepAlive</key>
   <dict><key>SuccessfulExit</key><false/></dict>
   <key>ThrottleInterval</key><integer>30</integer>
-  <key>StandardOutPath</key><string>${join4(logDir(), "out.log")}</string>
-  <key>StandardErrorPath</key><string>${join4(logDir(), "err.log")}</string>
+  <key>StandardOutPath</key><string>${join5(logDir(), "out.log")}</string>
+  <key>StandardErrorPath</key><string>${join5(logDir(), "err.log")}</string>
   <key>EnvironmentVariables</key>
   <dict>
     <key>PATH</key><string>/usr/local/bin:/opt/homebrew/bin:/usr/bin:/bin</string>
@@ -47043,7 +47381,7 @@ function macUninstall() {
   const target = `gui/${uid}/${SERVICE_LABEL}`;
   launchctl(["bootout", target]);
   const plist = plistPath();
-  if (existsSync5(plist)) {
+  if (existsSync6(plist)) {
     try {
       unlinkSync(plist);
     } catch {
@@ -47056,12 +47394,12 @@ function macStatus() {
   return { running: r.ok, hint: r.ok ? "launchd managed" : "not installed" };
 }
 function systemdUnitPath() {
-  const xdg = process.env.XDG_CONFIG_HOME ?? join4(home(), ".config");
-  return join4(xdg, "systemd", "user", `${SYSTEMD_UNIT}.service`);
+  const xdg = process.env.XDG_CONFIG_HOME ?? join5(home(), ".config");
+  return join5(xdg, "systemd", "user", `${SYSTEMD_UNIT}.service`);
 }
 function writeSystemdUnit(cliPath) {
   const unitPath = systemdUnitPath();
-  mkdirSync2(dirname4(unitPath), { recursive: true });
+  mkdirSync2(dirname5(unitPath), { recursive: true });
   const unit = `[Unit]
 Description=modelstat agent
 Documentation=https://modelstat.ai
@@ -47076,8 +47414,8 @@ RestartSec=10
 # Don't restart-storm if the service is persistently unreachable.
 StartLimitIntervalSec=300
 StartLimitBurst=10
-StandardOutput=append:${join4(logDir(), "out.log")}
-StandardError=append:${join4(logDir(), "err.log")}
+StandardOutput=append:${join5(logDir(), "out.log")}
+StandardError=append:${join5(logDir(), "err.log")}
 
 [Install]
 WantedBy=default.target
@@ -47102,7 +47440,7 @@ function linuxInstall() {
 function linuxUninstall() {
   systemctl(["disable", "--now", `${SYSTEMD_UNIT}.service`]);
   const unit = systemdUnitPath();
-  if (existsSync5(unit)) {
+  if (existsSync6(unit)) {
     try {
       unlinkSync(unit);
     } catch {
@@ -47146,9 +47484,9 @@ function logsDir() {
 }
 function installTrayApp(sourceAppPath) {
   if (platform2() !== "darwin") return null;
-  if (!existsSync5(sourceAppPath)) return null;
-  const dest = join4(home(), "Applications", "ModelstatTray.app");
-  mkdirSync2(dirname4(dest), { recursive: true });
+  if (!existsSync6(sourceAppPath)) return null;
+  const dest = join5(home(), "Applications", "ModelstatTray.app");
+  mkdirSync2(dirname5(dest), { recursive: true });
   spawnSync("rm", ["-rf", dest]);
   const r = spawnSync("cp", ["-R", sourceAppPath, dest], { encoding: "utf8" });
   if (r.status !== 0) {
@@ -47158,28 +47496,28 @@ function installTrayApp(sourceAppPath) {
 }
 function bundledTrayAppPath() {
   if (platform2() !== "darwin") return null;
-  const here2 = dirname4(fileURLToPath2(import.meta.url));
+  const here2 = dirname5(fileURLToPath2(import.meta.url));
   const candidates = [
     // Pre-built .app — CI with codesigning drops one here.
-    join4(here2, "..", "vendor", "ModelstatTray.app"),
+    join5(here2, "..", "vendor", "ModelstatTray.app"),
     // Local dev layout: apps/agent-dev/src/service.ts → ../../tray-mac/build/ModelstatTray.app
-    join4(here2, "..", "..", "tray-mac", "build", "ModelstatTray.app")
+    join5(here2, "..", "..", "tray-mac", "build", "ModelstatTray.app")
   ];
   for (const c of candidates) {
-    if (existsSync5(c)) return c;
+    if (existsSync6(c)) return c;
   }
   const sourceDirs = [
-    join4(here2, "..", "vendor", "tray-mac"),
-    join4(here2, "..", "..", "tray-mac")
+    join5(here2, "..", "vendor", "tray-mac"),
+    join5(here2, "..", "..", "tray-mac")
   ];
   for (const src of sourceDirs) {
-    const build = join4(src, "build-app.sh");
-    if (!existsSync5(build)) continue;
+    const build = join5(src, "build-app.sh");
+    if (!existsSync6(build)) continue;
     if (!hasSwift()) return null;
     const r = spawnSync("bash", [build], { cwd: src, encoding: "utf8" });
     if (r.status === 0) {
-      const app = join4(src, "build", "ModelstatTray.app");
-      if (existsSync5(app)) return app;
+      const app = join5(src, "build", "ModelstatTray.app");
+      if (existsSync6(app)) return app;
     }
   }
   return null;
@@ -47411,12 +47749,33 @@ async function cmdConnect(opts) {
       warn(`tray install skipped: ${e.message}`);
     }
   }
+  step("Preparing local summariser (downloads on first run)");
+  let modelReady = false;
+  try {
+    const { ensureLlamaModel: ensureLlamaModel2, defaultLlamaConfig: defaultLlamaConfig2 } = await Promise.resolve().then(() => (init_node2(), node_exports));
+    await ensureLlamaModel2(defaultLlamaConfig2());
+    modelReady = true;
+    emitEvent(opts, "summariser_model_ready", {});
+    ok("summariser model on disk");
+  } catch (e) {
+    emitEvent(opts, "summariser_model_failed", {
+      error: e.message
+    });
+    warn(`couldn't prepare summariser model: ${e.message}`);
+    warn(
+      "the background service will retry the download on its first scan"
+    );
+  }
   step("Installing background service so the agent survives reboots");
   let serviceOk = false;
   try {
     const svc = installService();
     serviceOk = true;
-    emitEvent(opts, "service_installed", { path: svc.path, logs: svc.logs });
+    emitEvent(opts, "service_installed", {
+      path: svc.path,
+      logs: svc.logs,
+      summariser_ready: modelReady
+    });
     ok(`${platform4() === "darwin" ? "launchd" : "systemd --user"}: ${svc.path}`);
   } catch (e) {
     emitEvent(opts, "service_install_failed", { error: e.message });
@@ -47485,6 +47844,9 @@ async function cmdDiscover() {
   console.log("\u2713 reported to backend");
 }
 async function cmdScan() {
+  const { preflightSummariser: preflightSummariser2 } = await Promise.resolve().then(() => (init_pipeline2(), pipeline_exports));
+  const sample = await preflightSummariser2();
+  console.log(`[modelstat] summariser preflight ok: "${sample}"`);
   const r = await scanAll();
   console.log();
   console.log(