modelstat 0.8.3 → 0.10.0

package/dist/cli.mjs

@@ -34556,7 +34556,9 @@ function load() {
       cursor: obj.cursor ?? {},
       segmentsSent: obj.segmentsSent ?? 0,
       processingVersion: obj.processingVersion ?? null,
-      reconcileCache: obj.reconcileCache ?? {}
+      reconcileCache: obj.reconcileCache ?? {},
+      summariserDegraded: obj.summariserDegraded ?? false,
+      summariserRecoveryAt: obj.summariserRecoveryAt ?? 0
     };
   } catch {
     cache2 = { ...DEFAULTS, cursor: {} };
@@ -34579,7 +34581,9 @@ var init_runtime_state = __esm({
       cursor: {},
       segmentsSent: 0,
       processingVersion: null,
-      reconcileCache: {}
+      reconcileCache: {},
+      summariserDegraded: false,
+      summariserRecoveryAt: 0
     };
     cache2 = null;
     runtimeState = {
@@ -34646,6 +34650,25 @@ var init_runtime_state = __esm({
         s.processingVersion = v;
         persist(s);
       },
+      /** Whether the last run shipped extractive (LLM-unavailable) abstracts. */
+      getSummariserDegraded() {
+        return load().summariserDegraded;
+      },
+      setSummariserDegraded(v) {
+        const s = load();
+        if (s.summariserDegraded === v) return;
+        s.summariserDegraded = v;
+        persist(s);
+      },
+      /** ms-epoch of the last degradation-recovery re-scan (0 = never). */
+      getSummariserRecoveryAt() {
+        return load().summariserRecoveryAt;
+      },
+      setSummariserRecoveryAt(ms) {
+        const s = load();
+        s.summariserRecoveryAt = ms;
+        persist(s);
+      },
       /** Self-healing reconcile cache (see {@link RuntimeState.reconcileCache}). */
       getReconcileCache() {
         return load().reconcileCache;
@@ -35289,6 +35312,61 @@ var init_prompts = __esm({
   }
 });
 
+// ../../packages/daemon-core/src/pipeline/heuristic-summary.ts
+function parsePromptExcerpts(prompt) {
+  const out = [];
+  const re = /\[turn \d+\]\s*"([^"]*)"/g;
+  let m = re.exec(prompt);
+  while (m) {
+    if (m[1]) out.push(m[1]);
+    m = re.exec(prompt);
+  }
+  return out;
+}
+function parsePromptFacts(prompt) {
+  const m = /Session context:\s*(.+?)\.\s*(?:\n|$)/.exec(prompt);
+  return m?.[1]?.trim() ?? "";
+}
+function pickIntent(lines) {
+  const early = lines.slice(0, 5);
+  const substantive = early.filter((l) => l.length >= 16 && !GREETING.test(l));
+  if (substantive.length > 0) return substantive[0];
+  const byLen = [...early].sort((a, b) => b.length - a.length);
+  return byLen[0] ?? lines.find((l) => l.length > 0) ?? null;
+}
+function cleanLead(s) {
+  let t = s.replace(GREETING, "").trim();
+  t = t.replace(LEAD_FILLER, "").trim();
+  if (!t) t = s.trim();
+  return t.charAt(0).toUpperCase() + t.slice(1);
+}
+function clamp(s, max) {
+  const t = s.replace(/\s+/g, " ").trim();
+  if (t.length <= max) return t;
+  const cut = t.slice(0, max - 1);
+  const sp = cut.lastIndexOf(" ");
+  return `${(sp > max * 0.6 ? cut.slice(0, sp) : cut).trimEnd()}\u2026`;
+}
+function heuristicSummarize() {
+  return async ({ prompt, excerpts, facts }) => {
+    const lines = (excerpts && excerpts.length > 0 ? excerpts : parsePromptExcerpts(prompt)).map((s) => s.replace(/\s+/g, " ").trim()).filter((s) => s.length > 0);
+    const factText = (facts?.trim() || parsePromptFacts(prompt)).replace(/\s+/g, " ").trim();
+    const intent = pickIntent(lines);
+    const lead = intent ? cleanLead(intent) : "AI coding session";
+    const body = factText ? `${lead} [${factText}]` : lead;
+    return clamp(body, ABSTRACT_OUTPUT_MAX_CHARS);
+  };
+}
+var GREETING, LEAD_FILLER;
+var init_heuristic_summary = __esm({
+  "../../packages/daemon-core/src/pipeline/heuristic-summary.ts"() {
+    "use strict";
+    init_prompts();
+    GREETING = /^(hi|hey|hello|thanks?|thank you|ok(ay)?|yes|no|sure|please|cool|great|nice)\b[\s!.,]*/i;
+    LEAD_FILLER = /^(can you|could you|could we|can we|please|i(?:'d| would)? (?:like|want)(?: you)? to|i need(?: you)? to|let'?s|help me|i'?m trying to|i am trying to|now|so|ok(?:ay)?)\s+/i;
+  }
+});
+
 // ../../packages/daemon-core/src/pipeline/script-summary.ts
 function buildScriptSummaryUserPrompt(input) {
   const body = input.content.slice(0, SCRIPT_SUMMARY_INPUT_MAX_CHARS);
@@ -35755,7 +35833,11 @@ ${excerptBlock}
 Write the SHORTEST keyword-dense paragraph (1-3 sentences, \u2264${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.`;
   const rawAbstract = await adapters2.summarize({
     prompt,
-    maxTokens: SUMMARISER_MAX_TOKENS
+    maxTokens: SUMMARISER_MAX_TOKENS,
+    // Structured inputs for the dependency-free fallback summariser (used when
+    // the bundled LLM can't load); the LLM path ignores these and uses `prompt`.
+    excerpts,
+    facts: promptFacts
   });
   if (!rawAbstract || rawAbstract.trim().length === 0) {
     throw new Error(
@@ -35916,6 +35998,7 @@ var init_pipeline = __esm({
     init_redact();
     init_cognition();
     init_prompts();
+    init_heuristic_summary();
     init_redact();
     init_cognition();
     init_prompts();
@@ -36873,10 +36956,41 @@ __export(pipeline_exports, {
   buildSessionMetadata: () => buildSessionMetadata2,
   buildSessionTitles: () => buildSessionTitles2,
   enrichScripts: () => enrichScripts,
-  preflightSummariser: () => preflightSummariser
+  preflightSummariser: () => preflightSummariser,
+  summariserDegradedThisProcess: () => summariserDegradedThisProcess
 });
 import { existsSync as existsSync8 } from "fs";
 import { readFile as fsReadFile } from "fs/promises";
+function summariserDegradedThisProcess() {
+  return degradedThisProcess;
+}
+function markDegraded(reason) {
+  if (!degradedThisProcess) {
+    degradedThisProcess = true;
+    console.warn(
+      `[modelstat] \u26A0 summariser DEGRADED \u2014 bundled LLM unavailable (${reason}); shipping extractive fallback abstracts so ingest continues. They re-summarise at model quality automatically once the LLM is healthy again.`
+    );
+  }
+  runtimeState.setSummariserDegraded(true);
+}
+function resilientSummarize(llamaCfg) {
+  const llm = llamaSummarize(llamaCfg);
+  const heuristic = heuristicSummarize();
+  return async (input) => {
+    if (Date.now() >= llmRetryAfter) {
+      try {
+        const out = await llm(input);
+        if (out && out.trim().length > 0) return out;
+      } catch (err) {
+        llmRetryAfter = Date.now() + LLM_RETRY_COOLDOWN_MS;
+        markDegraded(err.message);
+      }
+    } else {
+      markDegraded("LLM in post-failure cooldown");
+    }
+    return heuristic(input);
+  };
+}
 async function bundledAdapters() {
   const llamaCfg = defaultLlamaConfig();
   return {
@@ -36886,7 +37000,7 @@ async function bundledAdapters() {
     // vector-less with empty arrays; hooking embeddings here attaches a
     // real abstract embedding to each segment.)
     embed: createTransformersJsEmbedder(),
-    summarize: llamaSummarize(llamaCfg),
+    summarize: resilientSummarize(llamaCfg),
     tokenize: (text) => Math.max(1, Math.ceil(text.length / 4)),
     cognize: llamaCognize(llamaCfg),
     // Session-title pass — same bundled model, third chat session with
@@ -36919,14 +37033,15 @@ async function bundledAdapters() {
 }
 async function getAdapters() {
   if (adapters) return adapters;
+  let llmReady = true;
   try {
     await import("node-llama-cpp");
-  } catch (err) {
-    throw new Error(
-      `modelstat daemon can't start: the bundled summariser (node-llama-cpp) failed to load. Re-run \`modelstat connect\` (or \`npm i -g modelstat\`) so the native runtime is re-staged beside the bundle. Underlying error: ${err.message}`
-    );
+  } catch {
+    llmReady = false;
   }
-  console.log("[modelstat] using bundled local summariser (Qwen3.5-4B, runs on this machine)");
+  console.log(
+    llmReady ? "[modelstat] using bundled local summariser (Qwen3.5-4B, runs on this machine)" : "[modelstat] bundled summariser runtime not loadable \u2014 using extractive fallback (degraded) until it is"
+  );
   adapters = await bundledAdapters();
   return adapters;
 }
@@ -36965,16 +37080,19 @@ async function preflightSummariser() {
   const a = await getAdapters();
   const out = await a.summarize({
     prompt: 'Session context: smoke test. Sampled excerpts:\n  [turn 1] "hello world"\nWrite ONE sentence (\u2264240 chars) describing what the human was doing.',
-    maxTokens: 32
+    maxTokens: 32,
+    excerpts: ["smoke test \u2014 verifying the summariser is alive"],
+    facts: "preflight smoke test"
   });
+  const degraded = summariserDegradedThisProcess();
   if (!out || out.trim().length === 0) {
-    throw new Error(
-      "summariser preflight returned empty output \u2014 the configured summariser is reachable but produced no text. Check the model is loaded."
-    );
+    return { label: "summariser produced no output", degraded: true };
   }
-  return out.length > 60 ? `${out.slice(0, 57)}\u2026` : out;
+  const sample = out.length > 60 ? `${out.slice(0, 57)}\u2026` : out;
+  const engine = degraded ? "extractive fallback (LLM unavailable)" : "Qwen3.5-4B";
+  return { label: `${engine} \u2014 "${sample}"`, degraded };
 }
-var adapters, MAX_SCRIPT_READ_BYTES, scriptSummarizer;
+var adapters, degradedThisProcess, LLM_RETRY_COOLDOWN_MS, llmRetryAfter, MAX_SCRIPT_READ_BYTES, scriptSummarizer;
 var init_pipeline2 = __esm({
   "src/pipeline.ts"() {
     "use strict";
@@ -36983,7 +37101,11 @@ var init_pipeline2 = __esm({
     init_privacy_filter();
     init_src2();
     init_enrich_scripts();
+    init_runtime_state();
     adapters = null;
+    degradedThisProcess = false;
+    LLM_RETRY_COOLDOWN_MS = 10 * 6e4;
+    llmRetryAfter = 0;
     MAX_SCRIPT_READ_BYTES = 64 * 1024;
     scriptSummarizer = null;
   }
@@ -37193,7 +37315,7 @@ var init_scan = __esm({
     init_api();
     init_config2();
     init_pipeline2();
-    DAEMON_VERSION = true ? "daemon-0.8.3" : "daemon-dev";
+    DAEMON_VERSION = true ? "daemon-0.10.0" : "daemon-dev";
     BATCH_MAX_EVENTS = INGEST_BATCH_MAX_EVENTS;
     BATCH_MAX_TOOL_CALLS = 2e4;
     BATCH_BUFFER_HARD_CAP = BATCH_MAX_EVENTS * 2;
@@ -37760,7 +37882,7 @@ var PROCESSING_VERSION;
 var init_processing_version = __esm({
   "src/processing-version.ts"() {
     "use strict";
-    PROCESSING_VERSION = 7;
+    PROCESSING_VERSION = 8;
   }
 });
 
@@ -39996,14 +40118,33 @@ async function runDaemon(opts = {}) {
   const hb = setInterval(() => void sendHeartbeat(), HEARTBEAT_INTERVAL_MS);
   hb.unref();
   void sendHeartbeat();
+  const wasDegraded = runtimeState.getSummariserDegraded();
   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}"`);
+    const { label, degraded } = await preflightSummariser2();
+    if (degraded) {
+      console.warn(`[modelstat] \u26A0 summariser preflight DEGRADED \u2014 ${label}`);
+      setMessage(
+        "summariser degraded: extractive fallback (LLM unavailable) \u2014 ingest continues, self-heals when the model loads"
+      );
+    } else {
+      console.log(`[modelstat] summariser preflight ok: ${label}`);
+      if (wasDegraded) {
+        const since = Date.now() - runtimeState.getSummariserRecoveryAt();
+        if (since > SUMMARISER_RECOVERY_MIN_INTERVAL_MS) {
+          runtimeState.wipeCursors();
+          runtimeState.setSummariserRecoveryAt(Date.now());
+          console.log(
+            "[modelstat] summariser recovered \u2014 re-scanning so extractive fallback abstracts upgrade to model quality"
+          );
+        }
+      }
+      runtimeState.setSummariserDegraded(false);
+    }
   } catch (err) {
-    setPhase("error", `summariser preflight failed: ${err.message}`);
-    throw err;
+    console.warn(`[modelstat] summariser preflight error (continuing): ${err.message}`);
+    setMessage(`summariser preflight error (continuing): ${err.message}`);
   }
   const localIngest = await startLocalIngestReceiver();
   const LOCAL_DRAIN_INTERVAL_MS = 5e3;
@@ -40104,7 +40245,7 @@ async function runDaemon(opts = {}) {
   await new Promise(() => {
   });
 }
-var import_undici2, DAEMON_VERSION2, HEARTBEAT_INTERVAL_MS, SCAN_INTERVAL_MS, DISCOVERY_INTERVAL_MS, status, LOCAL_FLUSH_THROTTLE_MS, localFlushTimer, localFlushPending, lastVerdict, LOG_MAX_BYTES, LOG_TAIL_KEEP_BYTES, lastStatusPath, scanRunner;
+var import_undici2, DAEMON_VERSION2, HEARTBEAT_INTERVAL_MS, SCAN_INTERVAL_MS, DISCOVERY_INTERVAL_MS, SUMMARISER_RECOVERY_MIN_INTERVAL_MS, status, LOCAL_FLUSH_THROTTLE_MS, localFlushTimer, localFlushPending, lastVerdict, LOG_MAX_BYTES, LOG_TAIL_KEEP_BYTES, lastStatusPath, scanRunner;
 var init_daemon = __esm({
   "src/daemon.ts"() {
     "use strict";
@@ -40117,13 +40258,15 @@ var init_daemon = __esm({
     init_machine_key();
     init_receiver();
     init_reconcile();
+    init_runtime_state();
     init_scan();
     init_single_flight();
     init_update();
-    DAEMON_VERSION2 = true ? "daemon-0.8.3" : "daemon-dev";
+    DAEMON_VERSION2 = true ? "daemon-0.10.0" : "daemon-dev";
     HEARTBEAT_INTERVAL_MS = 1e4;
     SCAN_INTERVAL_MS = 5 * 60 * 1e3;
     DISCOVERY_INTERVAL_MS = 6e4;
+    SUMMARISER_RECOVERY_MIN_INTERVAL_MS = 6 * 60 * 6e4;
     status = {
       phase: "starting",
       message: null,
@@ -40305,15 +40448,16 @@ function setupRuntime() {
   return installBundle();
 }
 var NODE_LLAMA_CPP_FALLBACK_VERSION = "3.18.1";
-function sourceLlamaVersion(sourceCli) {
+var HF_TRANSFORMERS_FALLBACK_VERSION = "3.8.1";
+function sourcePkgVersion(sourceCli, pkgName) {
   try {
     const req = createRequire(sourceCli);
-    let d = dirname6(realpathSync(req.resolve("node-llama-cpp")));
+    let d = dirname6(realpathSync(req.resolve(pkgName)));
     for (let i = 0; i < 10; i++) {
       const pj = join8(d, "package.json");
       if (existsSync9(pj)) {
         const p = JSON.parse(readFileSync5(pj, "utf8"));
-        if (p.name === "node-llama-cpp" && p.version) return p.version;
+        if (p.name === pkgName && p.version) return p.version;
       }
       const up = dirname6(d);
       if (up === d) break;
@@ -40323,21 +40467,21 @@ function sourceLlamaVersion(sourceCli) {
   }
   return null;
 }
-function installNativeRuntime(sourceCli) {
-  const version = sourceLlamaVersion(sourceCli) ?? NODE_LLAMA_CPP_FALLBACK_VERSION;
+function stageNativePkg(sourceCli, pkgName, fallbackVersion, critical) {
+  const version = sourcePkgVersion(sourceCli, pkgName) ?? fallbackVersion;
   const dest = binDir();
+  const pkgJson = join8(dest, "node_modules", ...pkgName.split("/"), "package.json");
   try {
-    const have = JSON.parse(
-      readFileSync5(join8(dest, "node_modules", "node-llama-cpp", "package.json"), "utf8")
-    );
-    if (have.version === version) return [`node-llama-cpp@${version} (cached)`];
+    const have = JSON.parse(readFileSync5(pkgJson, "utf8"));
+    if (have.version === version) return [`${pkgName}@${version} (cached)`];
   } catch {
   }
   mkdirSync4(dest, { recursive: true });
   const childEnv = { ...process.env };
   delete childEnv.npm_config_global;
   delete childEnv.npm_config_prefix;
-  process.stderr.write(`  \xB7 staging summariser runtime (node-llama-cpp@${version})\u2026
+  const label = critical ? "summariser runtime" : "on-device embedder + PII/NER redactor";
+  process.stderr.write(`  \xB7 staging ${label} (${pkgName}@${version})\u2026
 `);
   const r = spawnSync2(
     "npm",
@@ -40350,27 +40494,38 @@ function installNativeRuntime(sourceCli) {
       "--omit=dev",
       "--no-audit",
       "--no-fund",
-      // Prefer the npm cache the current install already populated (node-llama-cpp
-      // is a direct dep, so the platform prebuilt is cached) — an offline ~3s copy
-      // instead of a redundant network re-fetch. Only a genuine cache miss touches
-      // the network, and a capped per-request timeout makes that fail fast (the
-      // daemon's summariser preflight re-stages) rather than hang indefinitely.
+      // Prefer the npm cache the current install already populated — an offline
+      // ~3s copy instead of a redundant network re-fetch. Only a genuine cache
+      // miss touches the network, and a capped per-request timeout makes that
+      // fail fast (the daemon re-stages on its next preflight) rather than hang.
       "--prefer-offline",
       "--fetch-timeout=60000",
       "--loglevel=error",
-      `node-llama-cpp@${version}`
+      `${pkgName}@${version}`
     ],
     { encoding: "utf8", stdio: "pipe", env: childEnv }
   );
   if (r.status !== 0) {
     process.stderr.write(
-      `[modelstat] couldn't stage the bundled summariser runtime via npm (node-llama-cpp@${version}); the daemon's summariser preflight will fail until this is resolved.
+      critical ? `[modelstat] couldn't stage the bundled summariser runtime via npm (${pkgName}@${version}); the daemon's summariser preflight will fail until this is resolved.
 ${(r.stderr || r.stdout || "").trim()}
+` : `[modelstat] couldn't stage the on-device embedder/redactor (${pkgName}@${version}); the daemon falls back to server-side embedding + regex/LLM redaction (non-fatal).
 `
     );
     return [];
   }
-  return [`node-llama-cpp@${version}`];
+  return [`${pkgName}@${version}`];
+}
+function installNativeRuntime(sourceCli) {
+  return [
+    ...stageNativePkg(sourceCli, "node-llama-cpp", NODE_LLAMA_CPP_FALLBACK_VERSION, true),
+    ...stageNativePkg(
+      sourceCli,
+      "@huggingface/transformers",
+      HF_TRANSFORMERS_FALLBACK_VERSION,
+      false
+    )
+  ];
 }
 function nodeBinary() {
   return process.execPath;
@@ -40724,7 +40879,7 @@ function tryOpenBrowser(url) {
     return false;
   }
 }
-var DAEMON_VERSION3 = true ? "daemon-0.8.3" : "daemon-dev";
+var DAEMON_VERSION3 = true ? "daemon-0.10.0" : "daemon-dev";
 function osFamily() {
   const p = platform6();
   if (p === "darwin") return "macos";
@@ -41097,8 +41252,10 @@ async function cmdDiscover() {
 }
 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 { label, degraded } = await preflightSummariser2();
+  console.log(
+    degraded ? `[modelstat] \u26A0 summariser DEGRADED \u2014 ${label}; extractive fallback, ingest continues` : `[modelstat] summariser preflight ok: ${label}`
+  );
   const { reconcileProcessingVersion: reconcileProcessingVersion2 } = await Promise.resolve().then(() => (init_processing_version(), processing_version_exports));
   const pv = reconcileProcessingVersion2(state);
   if (pv.changed) {