modelstat 0.0.42 → 0.0.43

package/dist/cli.mjs

@@ -4421,7 +4421,15 @@ var init_schemas = __esm({
           installation_id: external_exports.string(),
           identity_id: external_exports.string().nullable()
         })
-      ).optional()
+      ).optional(),
+      /** Optional per-session titles — session_id → short redacted title
+       * (≤120 chars) produced by the companion's local titler from the
+       * session's segment abstracts. Server-side this is last-write-wins
+       * per session; companions recompute it from the full session view on
+       * every upload, so the latest batch always carries the freshest
+       * title. Absent for runtimes without a titler (older agents, no-op
+       * browser summariser) — the server keeps whatever it has. */
+      session_titles: external_exports.record(external_exports.string(), external_exports.string().max(120)).optional()
     });
     HeartbeatPayload = external_exports.object({
       device_id: external_exports.string(),
@@ -4743,7 +4751,9 @@ async function parseClaudeCodeJsonl(ctx) {
     if (obj.type === "assistant") {
       const a = obj;
       const usage = a.message?.usage ?? {};
-      lastModel = a.message?.model ?? lastModel;
+      if (a.message?.model && a.message.model !== "<synthetic>") {
+        lastModel = a.message.model;
+      }
       if (!a.uuid || !sessionId) {
         skipped += 1;
         continue;
@@ -44387,6 +44397,89 @@ var init_cognition = __esm({
   }
 });
 
+// ../../packages/companion-core/src/pipeline/title.ts
+function buildTitleUserPrompt(input) {
+  const lines = input.abstracts.map(
+    (a, i) => `  [part ${i + 1}] ${a.replace(/\s+/g, " ").trim().slice(0, TITLER_ABSTRACT_SLICE_CHARS)}`
+  ).join("\n");
+  const facts = input.facts?.trim();
+  return `${facts ? `Session context: ${facts}.
+
+` : ""}Summaries of the session's parts (chronological):
+${lines}
+
+Write the title.`;
+}
+function stripCognitionSuffix(abstract) {
+  return abstract.replace(/\s*\[(?:Mood|Mind):[^\]]*\]/g, "").trim();
+}
+function sanitiseTitle(raw) {
+  if (!raw) return "";
+  const firstLine = raw.replace(/```[a-z]*\n?/gi, "").split(/[\n\r]/).map((l) => l.trim()).find((l) => l.length > 0) ?? "";
+  let t = firstLine.replace(/\s+/g, " ").replace(/^["'`“”]+/, "").replace(/["'`“”]+$/, "").replace(/[.!]+$/, "").trim();
+  if (!t) return "";
+  t = redact(t).text;
+  return t.slice(0, TITLE_MAX_CHARS).trim();
+}
+function fallbackTitle(abstracts) {
+  const first = abstracts.find((a) => a.trim().length > 0);
+  if (!first) return "";
+  const sentence = first.split(/(?<=[.!?])\s/, 1)[0] ?? first;
+  return sanitiseTitle(sentence);
+}
+function sampleAbstracts(abstracts, max = TITLER_MAX_ABSTRACTS) {
+  if (abstracts.length <= max) return abstracts;
+  const picks = /* @__PURE__ */ new Set([0, abstracts.length - 1]);
+  for (let i = 1; picks.size < max; i++) {
+    picks.add(Math.round(i * (abstracts.length - 1) / (max - 1)));
+  }
+  return [...picks].sort((a, b) => a - b).map((i) => abstracts[i]);
+}
+async function buildSessionTitles(segments, entitle) {
+  const bySession = /* @__PURE__ */ new Map();
+  for (const s of segments) {
+    const arr = bySession.get(s.session_id) ?? [];
+    arr.push(s);
+    bySession.set(s.session_id, arr);
+  }
+  const out = {};
+  for (const [sessionId, segs] of bySession) {
+    const sorted = [...segs].sort((a, b) => a.started_at.localeCompare(b.started_at));
+    const abstracts = sorted.map((s) => stripCognitionSuffix(s.abstract)).filter((a) => a.length > 0);
+    if (abstracts.length === 0) continue;
+    let title = "";
+    if (entitle) {
+      const first = sorted[0];
+      const project = first.tags.find((t) => t.root_key === "projects")?.name;
+      const facts = [
+        project ? `repo ${project}` : null,
+        `${sorted.length} part${sorted.length === 1 ? "" : "s"} on ${first.tool}`
+      ].filter(Boolean).join("; ");
+      try {
+        title = sanitiseTitle(await entitle({ abstracts: sampleAbstracts(abstracts), facts }));
+      } catch {
+        title = "";
+      }
+    }
+    if (!title) title = fallbackTitle(abstracts);
+    if (title) out[sessionId] = title;
+  }
+  return out;
+}
+var TITLER_SYSTEM_PROMPT, TITLE_MAX_CHARS, TITLER_MAX_TOKENS, TITLER_TEMPERATURE, TITLER_MAX_ABSTRACTS, TITLER_ABSTRACT_SLICE_CHARS;
+var init_title = __esm({
+  "../../packages/companion-core/src/pipeline/title.ts"() {
+    "use strict";
+    init_redact();
+    TITLER_SYSTEM_PROMPT = "You write a short TITLE for an AI coding session, given one-sentence summaries of its parts in chronological order. The title names what the session was about at a high level, in 3-8 words. If the session clearly spans more than one theme, name the top themes (at most 3) separated by ' \xB7 '. Use concrete domain keywords (features, components, subsystems). No narration, no filler words like 'session' or 'work on', no quotes, no trailing period, no PII, no code literals, no file paths. Reply with only the title.";
+    TITLE_MAX_CHARS = 80;
+    TITLER_MAX_TOKENS = 60;
+    TITLER_TEMPERATURE = 0.2;
+    TITLER_MAX_ABSTRACTS = 10;
+    TITLER_ABSTRACT_SLICE_CHARS = 240;
+  }
+});
+
 // ../../packages/companion-core/src/pipeline/index.ts
 async function buildSegmentsForSession(events, adapters2, onProgress) {
   if (events.length === 0) return [];
@@ -44682,6 +44775,7 @@ var init_pipeline = __esm({
     init_redact();
     init_prompts();
     init_cognition();
+    init_title();
     SEGMENT_TIME_GAP_MS = 15 * 6e4;
     SEGMENT_TOPIC_THRESHOLD = 0.35;
     SEGMENT_MAX_TURNS = 100;
@@ -45050,6 +45144,9 @@ async function loadOnce(cfg) {
     const cognizerContext = await model.createContext({
       contextSize: Math.min(cfg.contextSize, 2048)
     });
+    const entitlerContext = await model.createContext({
+      contextSize: cfg.contextSize
+    });
     const summarizer = new llamaMod.LlamaChatSession({
       contextSequence: summariserContext.getSequence(),
       systemPrompt: SUMMARISER_SYSTEM_PROMPT
@@ -45058,7 +45155,11 @@ async function loadOnce(cfg) {
       contextSequence: cognizerContext.getSequence(),
       systemPrompt: COGNITION_SYSTEM_PROMPT
     });
-    loaded = { summarizer, cognizer };
+    const entitler = new llamaMod.LlamaChatSession({
+      contextSequence: entitlerContext.getSequence(),
+      systemPrompt: TITLER_SYSTEM_PROMPT
+    });
+    loaded = { summarizer, cognizer, entitler };
     return loaded;
   })();
   try {
@@ -45120,12 +45221,41 @@ function llamaCognize(cfg = defaultLlamaConfig()) {
     }
   };
 }
+function llamaEntitle(cfg = defaultLlamaConfig()) {
+  return async (input) => {
+    if (input.abstracts.length === 0) return null;
+    let loadedSessions;
+    try {
+      loadedSessions = await loadOnce(cfg);
+    } catch {
+      return null;
+    }
+    const { entitler } = loadedSessions;
+    const run = inflight.then(async () => {
+      entitler.resetChatHistory();
+      const raw = await entitler.prompt(buildTitleUserPrompt(input), {
+        temperature: TITLER_TEMPERATURE,
+        // Same thinking-budget rationale as the cognition pass: the
+        // answer is tiny but Qwen3.5 reasons first.
+        maxTokens: TITLER_MAX_TOKENS + 400
+      });
+      return stripThinking(raw ?? "") || null;
+    });
+    inflight = run.catch(() => void 0);
+    try {
+      return await run;
+    } catch {
+      return null;
+    }
+  };
+}
 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();
     init_cognition();
+    init_title();
     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;
@@ -45205,6 +45335,7 @@ __export(node_exports, {
   defaultOllamaConfig: () => defaultOllamaConfig,
   ensureLlamaModel: () => ensureLlamaModel,
   llamaCognize: () => llamaCognize,
+  llamaEntitle: () => llamaEntitle,
   llamaSummarize: () => llamaSummarize,
   ollamaCognize: () => ollamaCognize,
   ollamaEmbed: () => ollamaEmbed,
@@ -45322,6 +45453,7 @@ var init_privacy_filter = __esm({
 var pipeline_exports = {};
 __export(pipeline_exports, {
   buildSegments: () => buildSegments,
+  buildSessionTitles: () => buildSessionTitles2,
   preflightSummariser: () => preflightSummariser
 });
 async function bundledAdapters() {
@@ -45337,6 +45469,11 @@ async function bundledAdapters() {
     summarize: llamaSummarize(llamaCfg),
     tokenize: (text) => Math.max(1, Math.ceil(text.length / 4)),
     cognize: llamaCognize(llamaCfg),
+    // Session-title pass — same bundled model, third chat session with
+    // TITLER_SYSTEM_PROMPT. One short call per session per upload (the
+    // sessions-list title in the dashboard). Best-effort like cognize:
+    // failures fall back to a deterministic title in buildSessionTitles.
+    entitle: llamaEntitle(llamaCfg),
     // Model-based PII redactor (OpenAI Privacy Filter via
     // transformers.js / ONNX). Runs locally on CPU after the regex
     // pass in packages/core/redact.ts. ~1 GB model downloaded on
@@ -45366,6 +45503,10 @@ async function getAdapters() {
 async function buildSegments(events, onProgress) {
   return buildSegmentsForSession(events, await getAdapters(), onProgress);
 }
+async function buildSessionTitles2(segments) {
+  const a = await getAdapters();
+  return buildSessionTitles(segments, a.entitle);
+}
 async function preflightSummariser() {
   const a = await getAdapters();
   const out = await a.summarize({
@@ -45457,16 +45598,33 @@ async function scanAll(cb = {}) {
   let segmentsUploaded = 0;
   let buffer = [];
   let pendingCursors = [];
+  const runSegmentsBySession = /* @__PURE__ */ new Map();
   async function flushBatch() {
     if (!buffer.length) return;
     const events = buffer.map(withNonNullTokens);
     const segments = await buildSegments(events, cb.onProgress);
+    for (const seg of segments) {
+      const arr = runSegmentsBySession.get(seg.session_id) ?? [];
+      arr.push(seg);
+      runSegmentsBySession.set(seg.session_id, arr);
+    }
+    const titleInput = [];
+    for (const sessionId of new Set(segments.map((s) => s.session_id))) {
+      titleInput.push(...runSegmentsBySession.get(sessionId) ?? []);
+    }
+    let sessionTitles = {};
+    try {
+      sessionTitles = await buildSessionTitles2(titleInput);
+    } catch (e) {
+      console.warn("session titling failed \u2014 shipping batch untitled:", e.message);
+    }
     const batch = {
       batch_id: batchId(),
       device_id: deviceId,
       agent_version: AGENT_VERSION,
       events,
-      segments
+      segments,
+      ...Object.keys(sessionTitles).length ? { session_titles: sessionTitles } : {}
     };
     cb.onUpload?.({ events: events.length, segments: segments.length });
     const res = await uploadBatch(batch);
@@ -45513,7 +45671,7 @@ var init_scan = __esm({
     init_pipeline2();
     init_config2();
     init_api();
-    AGENT_VERSION = true ? "agent-0.0.42" : "agent-dev";
+    AGENT_VERSION = true ? "agent-0.0.43" : "agent-dev";
     BATCH_MAX_EVENTS = 2e3;
     ZERO_TOKENS = {
       input: 0,
@@ -45704,7 +45862,7 @@ var PROCESSING_VERSION;
 var init_processing_version = __esm({
   "src/processing-version.ts"() {
     "use strict";
-    PROCESSING_VERSION = 3;
+    PROCESSING_VERSION = 4;
   }
 });
 
@@ -47713,7 +47871,7 @@ var init_daemon = __esm({
     init_lock();
     init_scan();
     init_single_flight();
-    AGENT_VERSION2 = true ? "agent-0.0.42" : "agent-dev";
+    AGENT_VERSION2 = true ? "agent-0.0.43" : "agent-dev";
     HEARTBEAT_INTERVAL_MS = 1e4;
     SCAN_INTERVAL_MS = 5 * 60 * 1e3;
     DISCOVERY_INTERVAL_MS = 6e4;
@@ -48223,7 +48381,7 @@ function tryOpenBrowser(url) {
     return false;
   }
 }
-var AGENT_VERSION3 = true ? "agent-0.0.42" : "agent-dev";
+var AGENT_VERSION3 = true ? "agent-0.0.43" : "agent-dev";
 function osFamily() {
   const p = platform4();
   if (p === "darwin") return "macos";