omnius 1.0.133 → 1.0.134

package/dist/index.js

@@ -1328,6 +1328,577 @@ var init_tool_executor = __esm({
   }
 });
 
+// packages/execution/dist/model-broker.js
+import { EventEmitter } from "node:events";
+import { totalmem, freemem } from "node:os";
+import { exec } from "node:child_process";
+function ramSnapshotMB() {
+  const total = Math.round(totalmem() / (1024 * 1024));
+  const free = Math.round(freemem() / (1024 * 1024));
+  return { total, free, used: total - free };
+}
+async function vramSnapshotMB() {
+  if (_nvSmiAvailable === false)
+    return null;
+  try {
+    const out = await new Promise((resolve55, reject) => {
+      exec("nvidia-smi --query-gpu=memory.total,memory.used,memory.free --format=csv,noheader,nounits 2>/dev/null", { encoding: "utf8", timeout: 3e3 }, (err, stdout) => err ? reject(err) : resolve55(stdout));
+    });
+    _nvSmiAvailable = true;
+    let total = 0, used = 0, free = 0;
+    for (const line of out.trim().split("\n")) {
+      const parts = line.split(",").map((s2) => s2.trim());
+      if (parts.length < 3)
+        continue;
+      total += parseInt(parts[0] ?? "0", 10) || 0;
+      used += parseInt(parts[1] ?? "0", 10) || 0;
+      free += parseInt(parts[2] ?? "0", 10) || 0;
+    }
+    if (total <= 0)
+      return null;
+    return { total, used, free };
+  } catch {
+    _nvSmiAvailable = false;
+    return null;
+  }
+}
+function getModelBroker() {
+  return ModelBroker.getInstance();
+}
+var DEFAULT_RAM_HEADROOM_MB, DEFAULT_VRAM_HEADROOM_MB, DEFAULT_IDLE_EVICT_MS, DEFAULT_POLL_MS, DEFAULT_INFLIGHT_WAIT_MS, ModelBroker, _nvSmiAvailable;
+var init_model_broker = __esm({
+  "packages/execution/dist/model-broker.js"() {
+    "use strict";
+    DEFAULT_RAM_HEADROOM_MB = 2048;
+    DEFAULT_VRAM_HEADROOM_MB = 1024;
+    DEFAULT_IDLE_EVICT_MS = 5 * 60 * 1e3;
+    DEFAULT_POLL_MS = 4e3;
+    DEFAULT_INFLIGHT_WAIT_MS = 6e4;
+    ModelBroker = class _ModelBroker {
+      static _instance = null;
+      /** Loaded model registry keyed by `${host}:${name}`. */
+      _loaded = /* @__PURE__ */ new Map();
+      /** In-flight load promises keyed by `${host}:${name}`. */
+      _inflight = /* @__PURE__ */ new Map();
+      /** Fallback chains keyed by domain. */
+      _fallbacks = /* @__PURE__ */ new Map();
+      /** Cached n_ctx_train per Ollama model. */
+      _ctxTrainCache = /* @__PURE__ */ new Map();
+      /** Cached "model exists in Ollama" probes (true / false). */
+      _ollamaModelExists = /* @__PURE__ */ new Map();
+      /** Event emitter — typed via `on<K>(event: K, listener: BrokerEvents[K])`. */
+      _events = new EventEmitter();
+      /** Poll timer. */
+      _pollTimer = null;
+      /** Last full snapshot. */
+      _lastSnapshot = null;
+      /** Configured Ollama base URL. */
+      _ollamaBaseUrl = process.env["OLLAMA_HOST"] || "http://127.0.0.1:11434";
+      /** Configured thresholds (mutable for /broker tuning). */
+      ramHeadroomMB = DEFAULT_RAM_HEADROOM_MB;
+      vramHeadroomMB = DEFAULT_VRAM_HEADROOM_MB;
+      idleEvictMs = DEFAULT_IDLE_EVICT_MS;
+      static getInstance() {
+        if (!_ModelBroker._instance)
+          _ModelBroker._instance = new _ModelBroker();
+        return _ModelBroker._instance;
+      }
+      /** Reset (test-only). */
+      static resetInstance() {
+        if (_ModelBroker._instance?._pollTimer)
+          clearInterval(_ModelBroker._instance._pollTimer);
+        _ModelBroker._instance = null;
+      }
+      constructor() {
+        this.registerDefaultFallbacks();
+      }
+      // ------------------------------------------------------------------
+      // Public API — events
+      // ------------------------------------------------------------------
+      on(event, listener) {
+        this._events.on(event, listener);
+        return this;
+      }
+      off(event, listener) {
+        this._events.off(event, listener);
+        return this;
+      }
+      emit(event, ...args) {
+        this._events.emit(event, ...args);
+      }
+      // ------------------------------------------------------------------
+      // Public API — polling
+      // ------------------------------------------------------------------
+      /** Start background polling of Ollama /api/ps and nvidia-smi. */
+      startPolling(intervalMs = DEFAULT_POLL_MS) {
+        if (this._pollTimer)
+          return;
+        this._pollTimer = setInterval(() => {
+          this.pollOnce().catch(() => {
+          });
+        }, intervalMs);
+        this.pollOnce().catch(() => {
+        });
+      }
+      stopPolling() {
+        if (this._pollTimer) {
+          clearInterval(this._pollTimer);
+          this._pollTimer = null;
+        }
+      }
+      /** Configure Ollama base URL (called from cli config wiring). */
+      setOllamaBaseUrl(url) {
+        this._ollamaBaseUrl = url;
+      }
+      /** One poll cycle — refreshes /api/ps and emits snapshot. */
+      async pollOnce() {
+        await Promise.all([
+          this.refreshOllamaPs().catch(() => {
+          })
+          // VRAM total/free comes from system-metrics; broker computes its own snapshot
+        ]);
+        const snapshot = this.buildSnapshot();
+        this._lastSnapshot = snapshot;
+        this.emit("snapshot", snapshot);
+        this.checkPressure(snapshot);
+        return snapshot;
+      }
+      /** Best-known current snapshot. */
+      snapshot() {
+        return this._lastSnapshot ?? this.buildSnapshot();
+      }
+      // ------------------------------------------------------------------
+      // Public API — fallback registry
+      // ------------------------------------------------------------------
+      /** Register a fallback chain for a domain. Later entries are tried later. */
+      setFallbackChain(domain, chain) {
+        this._fallbacks.set(domain, [...chain]);
+      }
+      getFallbackChain(domain) {
+        return this._fallbacks.get(domain) ?? [];
+      }
+      // ------------------------------------------------------------------
+      // Public API — load decisioning
+      // ------------------------------------------------------------------
+      /**
+       * Pre-flight a model-load request. Always call this before allocating a model.
+       *
+       * Decisions:
+       *   - ok:                 proceed; use `effectiveNumCtx` if returned
+       *   - wait-for-inflight:  another caller is loading the same model; await `promise`
+       *   - evict:              caller should free `evictTargets` (broker calls evict
+       *                         hooks itself when possible) then retry
+       *   - degrade:            caller should reissue with `fallback`
+       *   - reject:             nothing viable — caller should error out
+       */
+      async ensureModelLoadable(spec) {
+        const key = this.keyOf(spec);
+        const inflight = this._inflight.get(key);
+        if (inflight) {
+          return { kind: "wait-for-inflight", promise: inflight.promise };
+        }
+        const existing = this._loaded.get(key);
+        if (existing) {
+          existing.lastUsedAt = Date.now();
+          return { kind: "ok", effectiveNumCtx: existing.numCtx, note: "already-loaded" };
+        }
+        let effectiveNumCtx;
+        if (spec.host === "ollama" && spec.requestedNumCtx) {
+          const trainCtx = await this.getNctxTrain(spec.name);
+          if (trainCtx && spec.requestedNumCtx > trainCtx) {
+            effectiveNumCtx = trainCtx;
+          } else {
+            effectiveNumCtx = spec.requestedNumCtx;
+          }
+        } else if (spec.host === "ollama") {
+          const trainCtx = await this.getNctxTrain(spec.name);
+          if (trainCtx)
+            effectiveNumCtx = trainCtx;
+        }
+        const estVram = spec.estimatedVramMB ?? this.estimateFootprintVramMB(spec);
+        const estRam = spec.estimatedRamMB ?? this.estimateFootprintRamMB(spec);
+        const ram = ramSnapshotMB();
+        const vram = await vramSnapshotMB();
+        const ramFitsAfter = ram.free - estRam >= this.ramHeadroomMB;
+        const vramFitsAfter = vram ? vram.free - estVram >= this.vramHeadroomMB : true;
+        if (ramFitsAfter && vramFitsAfter) {
+          const promise = Promise.resolve({ kind: "ok", effectiveNumCtx });
+          this._inflight.set(key, { startedMs: Date.now(), owner: spec.owner, promise });
+          setTimeout(() => this._inflight.delete(key), spec.loadTimeoutMs ?? DEFAULT_INFLIGHT_WAIT_MS).unref?.();
+          return { kind: "ok", effectiveNumCtx };
+        }
+        const evictTargets = this.pickEvictionCandidates({
+          needVramMB: vramFitsAfter ? 0 : estVram + this.vramHeadroomMB - (vram?.free ?? 0),
+          needRamMB: ramFitsAfter ? 0 : estRam + this.ramHeadroomMB - ram.free,
+          requestingPriority: spec.priority ?? 0,
+          requestingDomain: spec.domain
+        });
+        if (evictTargets.length > 0) {
+          return { kind: "evict", evictTargets, effectiveNumCtx };
+        }
+        const fallback = await this.findRunnableFallback(spec);
+        if (fallback) {
+          this.emit("degraded", spec, fallback, "insufficient-memory-no-evictable");
+          return { kind: "degrade", fallback, reason: "insufficient-memory-no-evictable" };
+        }
+        const reason = `insufficient resources (need ~${estRam}MB RAM, ~${estVram}MB VRAM; free ${ram.free}MB RAM, ${vram ? vram.free : "?"}MB VRAM) and no evictable / fallback models`;
+        this.emit("rejected", spec, reason);
+        return { kind: "reject", reason };
+      }
+      /**
+       * Register a model that has been successfully loaded.
+       * Callers MUST call this after a successful load so the broker can track LRU.
+       */
+      registerLoaded(model) {
+        const now = Date.now();
+        const m2 = {
+          ...model,
+          loadedAt: model.loadedAt ?? now,
+          lastUsedAt: model.lastUsedAt ?? now
+        };
+        this._loaded.set(m2.key, m2);
+        this._inflight.delete(m2.key);
+        this.emit("loaded", m2);
+        return m2;
+      }
+      /** Update last-used timestamp on every successful inference. */
+      touch(host, name10) {
+        const m2 = this._loaded.get(`${host}:${name10}`);
+        if (m2)
+          m2.lastUsedAt = Date.now();
+      }
+      /** Unregister a model (called when caller knows it has unloaded). */
+      unregisterLoaded(host, name10, reason = "caller-unloaded") {
+        const key = `${host}:${name10}`;
+        const m2 = this._loaded.get(key);
+        if (m2) {
+          this._loaded.delete(key);
+          this.emit("evicted", m2, reason);
+        }
+      }
+      /** Clear an inflight marker without registering a load (failed/aborted). */
+      clearInflight(host, name10) {
+        this._inflight.delete(`${host}:${name10}`);
+      }
+      /**
+       * Best-effort eviction of a tracked model. Returns true if the broker was
+       * able to actively unload (e.g. Ollama keep_alive=0); false if it just
+       * unregistered (caller must clean up its own subprocess).
+       */
+      async evict(host, name10, reason = "broker-evict") {
+        const key = `${host}:${name10}`;
+        const m2 = this._loaded.get(key);
+        if (!m2)
+          return false;
+        let actively = false;
+        if (host === "ollama") {
+          actively = await this.ollamaUnload(name10).catch(() => false);
+        }
+        this._loaded.delete(key);
+        this.emit("evicted", m2, reason);
+        return actively;
+      }
+      // ------------------------------------------------------------------
+      // Internal — Ollama
+      // ------------------------------------------------------------------
+      /** Fetch Ollama's runtime model list and reconcile against our registry. */
+      async refreshOllamaPs() {
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/ps`, {
+            signal: AbortSignal.timeout(3e3)
+          });
+          if (!res.ok)
+            return;
+          const data = await res.json();
+          const seen = /* @__PURE__ */ new Set();
+          const now = Date.now();
+          for (const m2 of data.models ?? []) {
+            const key = `ollama:${m2.name}`;
+            seen.add(key);
+            const vramMB = Math.round((m2.size_vram ?? 0) / (1024 * 1024));
+            const ramMB = Math.round(((m2.size ?? 0) - (m2.size_vram ?? 0)) / (1024 * 1024));
+            const existing = this._loaded.get(key);
+            if (existing) {
+              existing.vramMB = vramMB || existing.vramMB;
+              existing.ramMB = ramMB || existing.ramMB;
+            } else {
+              const tracked = this.registerLoaded({
+                key,
+                name: m2.name,
+                domain: this.guessOllamaDomain(m2.name),
+                host: "ollama",
+                owner: "external-ollama",
+                vramMB,
+                ramMB,
+                priority: 0,
+                loadedAt: now,
+                lastUsedAt: now
+              });
+              void tracked;
+            }
+          }
+          for (const [key, m2] of this._loaded) {
+            if (m2.host === "ollama" && !seen.has(key)) {
+              this._loaded.delete(key);
+              this.emit("evicted", m2, "ollama-unloaded");
+            }
+          }
+        } catch {
+        }
+      }
+      /** Force Ollama to unload a model by calling /api/generate keep_alive=0. */
+      async ollamaUnload(modelName) {
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/generate`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({ model: modelName, keep_alive: 0 }),
+            signal: AbortSignal.timeout(5e3)
+          });
+          return res.ok;
+        } catch {
+          return false;
+        }
+      }
+      /** Pull n_ctx_train for an Ollama model via /api/show; cached. */
+      async getNctxTrain(modelName) {
+        if (this._ctxTrainCache.has(modelName))
+          return this._ctxTrainCache.get(modelName);
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/show`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({ name: modelName }),
+            signal: AbortSignal.timeout(5e3)
+          });
+          if (!res.ok)
+            return null;
+          const data = await res.json();
+          const info = data.model_info ?? {};
+          const arch3 = info["general.architecture"];
+          let trainCtx;
+          if (arch3 && typeof info[`${arch3}.context_length`] === "number") {
+            trainCtx = info[`${arch3}.context_length`];
+          } else {
+            for (const [k, v] of Object.entries(info)) {
+              if (k.endsWith(".context_length") && typeof v === "number") {
+                trainCtx = v;
+                break;
+              }
+            }
+          }
+          if (trainCtx && Number.isFinite(trainCtx) && trainCtx > 0) {
+            this._ctxTrainCache.set(modelName, trainCtx);
+            return trainCtx;
+          }
+          return null;
+        } catch {
+          return null;
+        }
+      }
+      /** Probe whether a model exists in Ollama (cached). */
+      async ollamaModelExists(modelName) {
+        if (this._ollamaModelExists.has(modelName))
+          return this._ollamaModelExists.get(modelName);
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/show`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({ name: modelName }),
+            signal: AbortSignal.timeout(3e3)
+          });
+          const exists2 = res.ok;
+          this._ollamaModelExists.set(modelName, exists2);
+          return exists2;
+        } catch {
+          this._ollamaModelExists.set(modelName, false);
+          return false;
+        }
+      }
+      // ------------------------------------------------------------------
+      // Internal — LRU eviction selection
+      // ------------------------------------------------------------------
+      pickEvictionCandidates(req2) {
+        const now = Date.now();
+        const sameDomainOk = (m2) => (
+          // never evict the requesting domain's only loaded model unless multiple exist
+          m2.domain !== req2.requestingDomain || this.countByDomain(req2.requestingDomain) > 1
+        );
+        const idle = (m2) => now - m2.lastUsedAt > this.idleEvictMs;
+        const evictable = [...this._loaded.values()].filter((m2) => m2.priority <= req2.requestingPriority).filter(sameDomainOk).sort((a2, b) => {
+          const aIdle = idle(a2) ? 0 : 1;
+          const bIdle = idle(b) ? 0 : 1;
+          if (aIdle !== bIdle)
+            return aIdle - bIdle;
+          return a2.lastUsedAt - b.lastUsedAt;
+        });
+        const targets = [];
+        let vramFreed = 0;
+        let ramFreed = 0;
+        for (const m2 of evictable) {
+          if (vramFreed >= req2.needVramMB && ramFreed >= req2.needRamMB)
+            break;
+          targets.push(m2);
+          vramFreed += m2.vramMB;
+          ramFreed += m2.ramMB;
+        }
+        if (vramFreed >= req2.needVramMB && ramFreed >= req2.needRamMB)
+          return targets;
+        return [];
+      }
+      countByDomain(domain) {
+        let n2 = 0;
+        for (const m2 of this._loaded.values())
+          if (m2.domain === domain)
+            n2++;
+        return n2;
+      }
+      // ------------------------------------------------------------------
+      // Internal — fallback resolution
+      // ------------------------------------------------------------------
+      async findRunnableFallback(original) {
+        const chain = this._fallbacks.get(original.domain) ?? [];
+        for (const entry of chain) {
+          if (entry.spec.host === original.host && entry.spec.name === original.name)
+            continue;
+          const ok3 = entry.available ? await Promise.resolve(entry.available()).catch(() => false) : true;
+          if (!ok3)
+            continue;
+          return { ...entry.spec, owner: original.owner };
+        }
+        return null;
+      }
+      registerDefaultFallbacks() {
+        this.setFallbackChain("vision", [
+          { spec: { name: "moondream2", domain: "vision", host: "moondream-station" }, note: "local Moondream Station REST" },
+          { spec: { name: "moondream", domain: "vision", host: "ollama", estimatedVramMB: 1800 }, note: "ollama moondream (small VRAM)" },
+          { spec: { name: "tesseract-ocr-fallback", domain: "ocr", host: "subprocess", estimatedVramMB: 0, estimatedRamMB: 100 }, note: "OCR-only — no visual reasoning" }
+        ]);
+        this.setFallbackChain("image-gen", [
+          { spec: { name: "flux1-schnell", domain: "image-gen", host: "diffusers-py", estimatedVramMB: 12e3 } },
+          { spec: { name: "sdxl-turbo", domain: "image-gen", host: "diffusers-py", estimatedVramMB: 8e3 } },
+          { spec: { name: "sd-turbo", domain: "image-gen", host: "diffusers-py", estimatedVramMB: 4e3 } }
+        ]);
+        this.setFallbackChain("music", [
+          { spec: { name: "facebook/musicgen-medium", domain: "music", host: "audiocraft", estimatedVramMB: 6e3 } },
+          { spec: { name: "facebook/musicgen-small", domain: "music", host: "audiocraft", estimatedVramMB: 3e3 } }
+        ]);
+        this.setFallbackChain("sound", [
+          { spec: { name: "cvssp/audioldm-s-full-v2", domain: "sound", host: "diffusers-py", estimatedVramMB: 4e3 } },
+          { spec: { name: "facebook/audiogen-medium", domain: "sound", host: "audiocraft", estimatedVramMB: 3e3 } }
+        ]);
+        this.setFallbackChain("asr", [
+          { spec: { name: "base", domain: "asr", host: "whisper-cli", estimatedRamMB: 800 } },
+          { spec: { name: "tiny", domain: "asr", host: "whisper-cli", estimatedRamMB: 300 } }
+        ]);
+        this.setFallbackChain("tts", [
+          { spec: { name: "piper-default", domain: "tts", host: "piper", estimatedRamMB: 200 } }
+        ]);
+      }
+      // ------------------------------------------------------------------
+      // Internal — footprint estimation
+      // ------------------------------------------------------------------
+      estimateFootprintVramMB(spec) {
+        if (spec.estimatedVramMB !== void 0)
+          return spec.estimatedVramMB;
+        switch (spec.domain) {
+          case "embedding":
+            return 300;
+          case "vision":
+            return spec.host === "moondream-station" ? 2e3 : 2500;
+          case "image-gen":
+            return 8e3;
+          case "video-gen":
+            return 12e3;
+          case "music":
+            return 4e3;
+          case "sound":
+            return 3e3;
+          case "asr":
+            return 1e3;
+          case "tts":
+            return 300;
+          case "subagent":
+            return 4e3;
+          case "ocr":
+            return 0;
+          case "chat":
+          default:
+            return 5e3;
+        }
+      }
+      estimateFootprintRamMB(spec) {
+        if (spec.estimatedRamMB !== void 0)
+          return spec.estimatedRamMB;
+        switch (spec.domain) {
+          case "ocr":
+            return 100;
+          case "tts":
+            return 200;
+          case "embedding":
+            return 500;
+          case "asr":
+            return 800;
+          case "music":
+          case "sound":
+            return 2e3;
+          case "vision":
+            return 1500;
+          case "image-gen":
+            return 4e3;
+          case "video-gen":
+            return 6e3;
+          case "subagent":
+            return 1500;
+          case "chat":
+          default:
+            return 2e3;
+        }
+      }
+      guessOllamaDomain(name10) {
+        const n2 = name10.toLowerCase();
+        if (/embed|nomic|bge|e5/.test(n2))
+          return "embedding";
+        if (/moondream|llava|vlm|vision|qwen.*vl|minicpm-v|gemma3|pixtral|cogvlm|internvl/.test(n2))
+          return "vision";
+        if (/flux|stable.diffusion|sdxl|z-image/.test(n2))
+          return "image-gen";
+        return "chat";
+      }
+      // ------------------------------------------------------------------
+      // Internal — snapshot + pressure
+      // ------------------------------------------------------------------
+      buildSnapshot() {
+        const ram = ramSnapshotMB();
+        const vram = this._lastSnapshot?.vramMB ?? null;
+        return {
+          loaded: [...this._loaded.values()],
+          inflight: [...this._inflight.entries()].map(([key, v]) => ({ key, owner: v.owner, startedMs: v.startedMs })),
+          ramMB: ram,
+          vramMB: vram,
+          lastPollAt: Date.now()
+        };
+      }
+      async checkPressure(snap) {
+        if (snap.ramMB.free < this.ramHeadroomMB) {
+          this.emit("pressure", "ram", snap.ramMB.free, this.ramHeadroomMB);
+        }
+        const v = await vramSnapshotMB();
+        if (v) {
+          snap.vramMB = v;
+          if (v.free < this.vramHeadroomMB) {
+            this.emit("pressure", "vram", v.free, this.vramHeadroomMB);
+          }
+        }
+      }
+      keyOf(spec) {
+        return `${spec.host}:${spec.name}`;
+      }
+    };
+    _nvSmiAvailable = null;
+  }
+});
+
 // packages/execution/dist/tools/security-classifier.js
 function classifyTool(name10) {
   for (const rule of RULES) {
@@ -19513,6 +20084,20 @@ import { existsSync as existsSync25, mkdirSync as mkdirSync10, writeFileSync as
 import { join as join28, basename as basename5, extname as extname3, resolve as resolve16 } from "node:path";
 import { homedir as homedir9 } from "node:os";
 import { execFileSync as execFileSync3, execSync as execSync15 } from "node:child_process";
+function whisperRamEstimate(model) {
+  const m2 = model.toLowerCase();
+  if (m2.includes("large"))
+    return 3200;
+  if (m2.includes("medium"))
+    return 1700;
+  if (m2.includes("small"))
+    return 800;
+  if (m2.includes("base"))
+    return 400;
+  if (m2.includes("tiny"))
+    return 200;
+  return 800;
+}
 function isTranscribable(path12) {
   const ext = extname3(path12).toLowerCase();
   return AUDIO_EXTS.has(ext) || VIDEO_EXTS.has(ext);
@@ -19582,6 +20167,7 @@ var AUDIO_EXTS, VIDEO_EXTS, MAX_TRANSCRIBE_URL_BYTES, _tcModule, _tcChecked, Tra
 var init_transcribe_tool = __esm({
   "packages/execution/dist/tools/transcribe-tool.js"() {
     "use strict";
+    init_model_broker();
     init_network_egress_policy();
     AUDIO_EXTS = /* @__PURE__ */ new Set([
       ".mp3",
@@ -19636,7 +20222,7 @@ var init_transcribe_tool = __esm({
       async execute(args) {
         const start2 = performance.now();
         const filePath = resolve16(this.workingDir, String(args["path"] ?? ""));
-        const model = String(args["model"] ?? "base");
+        let model = String(args["model"] ?? "base");
         const diarize = Boolean(args["diarize"] ?? false);
         if (!existsSync25(filePath)) {
           return {
@@ -19654,6 +20240,32 @@ var init_transcribe_tool = __esm({
             durationMs: performance.now() - start2
           };
         }
+        const broker = getModelBroker();
+        const askedModel = model;
+        let effectiveModel = model;
+        const whisperDecision = await broker.ensureModelLoadable({
+          name: askedModel,
+          domain: "asr",
+          host: "whisper-cli",
+          owner: "transcribe-file-tool",
+          estimatedRamMB: whisperRamEstimate(askedModel)
+        });
+        if (whisperDecision.kind === "degrade") {
+          effectiveModel = whisperDecision.fallback.name;
+        } else if (whisperDecision.kind === "evict") {
+          for (const target of whisperDecision.evictTargets) {
+            await broker.evict(target.host, target.name, "asr-needs-room");
+          }
+        } else if (whisperDecision.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Transcription blocked by resource broker: ${whisperDecision.reason}`,
+            durationMs: performance.now() - start2
+          };
+        }
+        if (effectiveModel !== askedModel)
+          model = effectiveModel;
         const tc = await loadTranscribeCli();
         if (!tc) {
           return this.execViaCli(filePath, model, diarize, start2);
@@ -46186,11 +46798,11 @@ var require_eventemitter3 = __commonJS({
       if (--emitter._eventsCount === 0) emitter._events = new Events();
       else delete emitter._events[evt];
     }
-    function EventEmitter14() {
+    function EventEmitter15() {
       this._events = new Events();
       this._eventsCount = 0;
     }
-    EventEmitter14.prototype.eventNames = function eventNames() {
+    EventEmitter15.prototype.eventNames = function eventNames() {
       var names = [], events, name10;
       if (this._eventsCount === 0) return names;
       for (name10 in events = this._events) {
@@ -46201,7 +46813,7 @@ var require_eventemitter3 = __commonJS({
       }
       return names;
     };
-    EventEmitter14.prototype.listeners = function listeners(event) {
+    EventEmitter15.prototype.listeners = function listeners(event) {
       var evt = prefix ? prefix + event : event, handlers = this._events[evt];
       if (!handlers) return [];
       if (handlers.fn) return [handlers.fn];
@@ -46210,13 +46822,13 @@ var require_eventemitter3 = __commonJS({
       }
       return ee;
     };
-    EventEmitter14.prototype.listenerCount = function listenerCount(event) {
+    EventEmitter15.prototype.listenerCount = function listenerCount(event) {
       var evt = prefix ? prefix + event : event, listeners = this._events[evt];
       if (!listeners) return 0;
       if (listeners.fn) return 1;
       return listeners.length;
     };
-    EventEmitter14.prototype.emit = function emit2(event, a1, a2, a3, a4, a5) {
+    EventEmitter15.prototype.emit = function emit2(event, a1, a2, a3, a4, a5) {
       var evt = prefix ? prefix + event : event;
       if (!this._events[evt]) return false;
       var listeners = this._events[evt], len = arguments.length, args, i2;
@@ -46267,13 +46879,13 @@ var require_eventemitter3 = __commonJS({
       }
       return true;
     };
-    EventEmitter14.prototype.on = function on2(event, fn, context2) {
+    EventEmitter15.prototype.on = function on2(event, fn, context2) {
       return addListener2(this, event, fn, context2, false);
     };
-    EventEmitter14.prototype.once = function once(event, fn, context2) {
+    EventEmitter15.prototype.once = function once(event, fn, context2) {
       return addListener2(this, event, fn, context2, true);
     };
-    EventEmitter14.prototype.removeListener = function removeListener2(event, fn, context2, once) {
+    EventEmitter15.prototype.removeListener = function removeListener2(event, fn, context2, once) {
       var evt = prefix ? prefix + event : event;
       if (!this._events[evt]) return this;
       if (!fn) {
@@ -46296,7 +46908,7 @@ var require_eventemitter3 = __commonJS({
       }
       return this;
     };
-    EventEmitter14.prototype.removeAllListeners = function removeAllListeners(event) {
+    EventEmitter15.prototype.removeAllListeners = function removeAllListeners(event) {
       var evt;
       if (event) {
         evt = prefix ? prefix + event : event;
@@ -46307,12 +46919,12 @@ var require_eventemitter3 = __commonJS({
       }
       return this;
     };
-    EventEmitter14.prototype.off = EventEmitter14.prototype.removeListener;
-    EventEmitter14.prototype.addListener = EventEmitter14.prototype.on;
-    EventEmitter14.prefixed = prefix;
-    EventEmitter14.EventEmitter = EventEmitter14;
+    EventEmitter15.prototype.off = EventEmitter15.prototype.removeListener;
+    EventEmitter15.prototype.addListener = EventEmitter15.prototype.on;
+    EventEmitter15.prefixed = prefix;
+    EventEmitter15.EventEmitter = EventEmitter15;
     if ("undefined" !== typeof module) {
-      module.exports = EventEmitter14;
+      module.exports = EventEmitter15;
     }
   }
 });
@@ -119084,10 +119696,10 @@ var require_wrap_handler = __commonJS({
 var require_dispatcher = __commonJS({
   "../node_modules/undici/lib/dispatcher/dispatcher.js"(exports, module) {
     "use strict";
-    var EventEmitter14 = __require("node:events");
+    var EventEmitter15 = __require("node:events");
     var WrapHandler = require_wrap_handler();
     var wrapInterceptor = (dispatch) => (opts, handler) => dispatch(opts, WrapHandler.wrap(handler));
-    var Dispatcher2 = class extends EventEmitter14 {
+    var Dispatcher2 = class extends EventEmitter15 {
       dispatch() {
         throw new Error("not implemented");
       }
@@ -126370,7 +126982,7 @@ var require_socks5_utils = __commonJS({
 var require_socks5_client = __commonJS({
   "../node_modules/undici/lib/core/socks5-client.js"(exports, module) {
     "use strict";
-    var { EventEmitter: EventEmitter14 } = __require("node:events");
+    var { EventEmitter: EventEmitter15 } = __require("node:events");
     var { Buffer: Buffer7 } = __require("node:buffer");
     var { InvalidArgumentError, Socks5ProxyError } = require_errors2();
     var { debuglog } = __require("node:util");
@@ -126413,7 +127025,7 @@ var require_socks5_client = __commonJS({
       ERROR: "error",
       CLOSED: "closed"
     };
-    var Socks5Client = class extends EventEmitter14 {
+    var Socks5Client = class extends EventEmitter15 {
       constructor(socket, options2 = {}) {
         super();
         if (!socket) {
@@ -132798,9 +133410,9 @@ var require_memory_cache_store = __commonJS({
   "../node_modules/undici/lib/cache/memory-cache-store.js"(exports, module) {
     "use strict";
     var { Writable } = __require("node:stream");
-    var { EventEmitter: EventEmitter14 } = __require("node:events");
+    var { EventEmitter: EventEmitter15 } = __require("node:events");
     var { assertCacheKey, assertCacheValue } = require_cache();
-    var MemoryCacheStore = class extends EventEmitter14 {
+    var MemoryCacheStore = class extends EventEmitter15 {
       #maxCount = 1024;
       #maxSize = 104857600;
       // 100MB
@@ -229661,7 +230273,7 @@ var init_send_ssdp_message = __esm({
 });
 
 // ../node_modules/@achingbrain/ssdp/dist/src/ssdp.js
-import { EventEmitter as EventEmitter2, on } from "node:events";
+import { EventEmitter as EventEmitter3, on } from "node:events";
 import { createRequire } from "node:module";
 var req, name8, version2, DEFAULT_SSDP_SIGNATURE, SSDP;
 var init_ssdp = __esm({
@@ -229679,7 +230291,7 @@ var init_ssdp = __esm({
     req = createRequire(import.meta.url);
     ({ name: name8, version: version2 } = req("../../package.json"));
     DEFAULT_SSDP_SIGNATURE = `node.js/${process.version.substring(1)} UPnP/1.1 ${name8}/${version2}`;
-    SSDP = class extends EventEmitter2 {
+    SSDP = class extends EventEmitter3 {
       udn;
       signature;
       sockets;
@@ -236837,7 +237449,7 @@ var require_extension2 = __commonJS({
 var require_websocket2 = __commonJS({
   "../node_modules/ws/lib/websocket.js"(exports, module) {
     "use strict";
-    var EventEmitter14 = __require("events");
+    var EventEmitter15 = __require("events");
     var https4 = __require("https");
     var http6 = __require("http");
     var net5 = __require("net");
@@ -236869,7 +237481,7 @@ var require_websocket2 = __commonJS({
     var protocolVersions = [8, 13];
     var readyStates = ["CONNECTING", "OPEN", "CLOSING", "CLOSED"];
     var subprotocolRegex = /^[!#$%&'*+\-.0-9A-Z^_`|a-z~]+$/;
-    var WebSocket6 = class _WebSocket extends EventEmitter14 {
+    var WebSocket6 = class _WebSocket extends EventEmitter15 {
       /**
        * Create a new `WebSocket`.
        *
@@ -237866,7 +238478,7 @@ var require_subprotocol = __commonJS({
 var require_websocket_server = __commonJS({
   "../node_modules/ws/lib/websocket-server.js"(exports, module) {
     "use strict";
-    var EventEmitter14 = __require("events");
+    var EventEmitter15 = __require("events");
     var http6 = __require("http");
     var { Duplex: Duplex3 } = __require("stream");
     var { createHash: createHash31 } = __require("crypto");
@@ -237879,7 +238491,7 @@ var require_websocket_server = __commonJS({
     var RUNNING = 0;
     var CLOSING = 1;
     var CLOSED = 2;
-    var WebSocketServer4 = class extends EventEmitter14 {
+    var WebSocketServer4 = class extends EventEmitter15 {
       /**
        * Create a `WebSocketServer` instance.
        *
@@ -254016,6 +254628,7 @@ var DEFAULT_DIFFUSERS_IMAGE_MODEL, DEFAULT_OLLAMA_IMAGE_MODEL, LEGACY_SDXL_TURBO
 var init_image_generate = __esm({
   "packages/execution/dist/tools/image-generate.js"() {
     "use strict";
+    init_model_broker();
     init_venv_paths();
     init_model_store();
     DEFAULT_DIFFUSERS_IMAGE_MODEL = "Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers";
@@ -254830,6 +255443,28 @@ if __name__ == "__main__":
         const requestedBackend = args["backend"] ? String(args["backend"]) : this.defaultBackend;
         const seed = optionalNumberArg(args["seed"]);
         const candidates = imageGenerationFallbackCandidates(requestedModel, requestedBackend, generationFallbackEnabled(args));
+        const broker = getModelBroker();
+        const firstCandidate = candidates[0];
+        if (firstCandidate) {
+          const decision2 = await broker.ensureModelLoadable({
+            name: firstCandidate.model,
+            domain: "image-gen",
+            host: firstCandidate.backend === "ollama" ? "ollama" : "diffusers-py",
+            owner: "image-generate-tool"
+          });
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              await broker.evict(target.host, target.name, "image-gen-needs-room");
+            }
+          } else if (decision2.kind === "reject") {
+            return {
+              success: false,
+              output: "",
+              error: `Image generation blocked by resource broker: ${decision2.reason}`,
+              durationMs: performance.now() - start2
+            };
+          }
+        }
         try {
           return await this.generateCandidateLadder({ candidates, prompt, args, seed, start: start2 });
         } catch (err) {
@@ -256270,6 +256905,7 @@ var DEFAULT_SOUND_MODEL, DEFAULT_MUSIC_MODEL, DIFFUSERS_AUDIO_PACKAGES, TRANSFOR
 var init_audio_generate = __esm({
   "packages/execution/dist/tools/audio-generate.js"() {
     "use strict";
+    init_model_broker();
     init_venv_paths();
     init_model_store();
     DEFAULT_SOUND_MODEL = "cvssp/audioldm-s-full-v2";
@@ -257217,6 +257853,28 @@ if __name__ == "__main__":
         const candidates = audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, generationFallbackEnabled2(args));
         const seed = optionalNumberArg2(args["seed"]);
         const playback = playbackRequested(args);
+        const broker = getModelBroker();
+        const firstCandidate = candidates[0];
+        if (firstCandidate) {
+          const decision2 = await broker.ensureModelLoadable({
+            name: firstCandidate.model,
+            domain: kind === "music" ? "music" : "sound",
+            host: firstCandidate.backend === "audiocraft" ? "audiocraft" : firstCandidate.backend === "tangoflux" ? "tangoflux" : firstCandidate.backend === "transformers" ? "diffusers-py" : "diffusers-py",
+            owner: `audio-generate-tool/${kind}`
+          });
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              await broker.evict(target.host, target.name, `${kind}-gen-needs-room`);
+            }
+          } else if (decision2.kind === "reject") {
+            return {
+              success: false,
+              output: "",
+              error: `${kind === "music" ? "Music" : "Sound"} generation blocked by resource broker: ${decision2.reason}`,
+              durationMs: performance.now() - start2
+            };
+          }
+        }
         try {
           return await this.generateCandidateLadder({ kind, candidates, prompt, args, seed, playback, start: start2 });
         } catch (err) {
@@ -258160,6 +258818,7 @@ var DEFAULT_DIFFUSERS_VIDEO_MODEL, SANA_VIDEO_480P_MODEL, SANA_VIDEO_720P_MODEL,
 var init_video_generate = __esm({
   "packages/execution/dist/tools/video-generate.js"() {
     "use strict";
+    init_model_broker();
     init_venv_paths();
     init_model_store();
     DEFAULT_DIFFUSERS_VIDEO_MODEL = "Efficient-Large-Model/SANA-Video_2B_480p";
@@ -259497,6 +260156,28 @@ if __name__ == "__main__":
         const withAudio = booleanArg3(args["with_audio"], false);
         const audioInput = typeof args["audio_input"] === "string" && args["audio_input"].trim() ? String(args["audio_input"]).trim() : void 0;
         const candidates = videoGenerationFallbackCandidates(requestedModel, requestedBackend, inferredKind, generationFallbackEnabled3(args), { preferNativeAudioVideo: withAudio || Boolean(audioInput) });
+        const broker = getModelBroker();
+        const firstCandidate = candidates[0];
+        if (firstCandidate) {
+          const decision2 = await broker.ensureModelLoadable({
+            name: firstCandidate.model,
+            domain: "video-gen",
+            host: firstCandidate.backend === "comfyui" ? "comfyui" : "diffusers-py",
+            owner: "video-generate-tool"
+          });
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              await broker.evict(target.host, target.name, "video-gen-needs-room");
+            }
+          } else if (decision2.kind === "reject") {
+            return {
+              success: false,
+              output: "",
+              error: `Video generation blocked by resource broker: ${decision2.reason}`,
+              durationMs: performance.now() - start2
+            };
+          }
+        }
         if (candidates.length === 0) {
           return {
             success: false,
@@ -261097,6 +261778,7 @@ var moondreamClient, moondreamError, stationProcess, hfPointUnavailable, IMAGE_E
 var init_vision = __esm({
   "packages/execution/dist/tools/vision.js"() {
     "use strict";
+    init_model_broker();
     moondreamClient = null;
     moondreamError = null;
     stationProcess = null;
@@ -261195,14 +261877,43 @@ var init_vision = __esm({
               };
             }
           }
+          const broker = getModelBroker();
+          const moondreamDecision = await broker.ensureModelLoadable({
+            name: "moondream2",
+            domain: "vision",
+            host: "moondream-station",
+            owner: "vision-tool"
+          });
+          const forceDegradeToOllama = moondreamDecision.kind === "degrade" && moondreamDecision.fallback.host === "ollama";
+          const forceReject = moondreamDecision.kind === "reject" && this._activeModelHasVision !== true;
           let client = null;
-          try {
-            client = await getMoondreamClient();
-          } catch {
+          if (!forceDegradeToOllama) {
+            try {
+              client = await getMoondreamClient();
+            } catch {
+            }
           }
           if (client) {
+            broker.registerLoaded({
+              key: "moondream-station:moondream2",
+              name: "moondream2",
+              domain: "vision",
+              host: "moondream-station",
+              owner: "vision-tool",
+              vramMB: 2e3,
+              ramMB: 1500,
+              priority: 1
+            });
             return await this.runMoondream(client, buffer2, filename, action, prompt, length4, start2);
           }
+          if (forceReject) {
+            return {
+              success: false,
+              output: "",
+              error: `Vision blocked by resource broker: ${moondreamDecision.kind === "reject" ? moondreamDecision.reason : "insufficient memory"}`,
+              durationMs: performance.now() - start2
+            };
+          }
           const ollamaResult = await this.tryOllamaVision(buffer2, filename, action, prompt, length4, start2);
           if (ollamaResult)
             return ollamaResult;
@@ -261300,7 +262011,41 @@ Coordinates are normalized (0-1). Multiply by image width/height for pixel value
       async tryOllamaVision(buffer2, filename, action, prompt, length4, start2) {
         const ollamaHost = process.env["OLLAMA_HOST"] || "http://127.0.0.1:11434";
         const envModel = process.env["OLLAMA_VISION_MODEL"];
-        const model = envModel || (this._activeModelHasVision && this._activeModel ? this._activeModel : "moondream");
+        let model = envModel || (this._activeModelHasVision && this._activeModel ? this._activeModel : "moondream");
+        const broker = getModelBroker();
+        const decision2 = await broker.ensureModelLoadable({
+          name: model,
+          domain: "vision",
+          host: "ollama",
+          owner: "vision-tool/ollama",
+          requestedNumCtx: 2048,
+          estimatedVramMB: 2e3
+        });
+        let numCtx;
+        if (decision2.kind === "reject")
+          return null;
+        if (decision2.kind === "degrade") {
+          model = decision2.fallback.name;
+        } else if (decision2.kind === "evict") {
+          for (const target of decision2.evictTargets) {
+            await broker.evict(target.host, target.name, "vision-needs-room");
+          }
+          numCtx = decision2.effectiveNumCtx;
+        } else if (decision2.kind === "ok") {
+          numCtx = decision2.effectiveNumCtx;
+        } else if (decision2.kind === "wait-for-inflight") {
+          const inner = await decision2.promise;
+          if (inner.kind === "ok")
+            numCtx = inner.effectiveNumCtx;
+          else if (inner.kind === "degrade")
+            model = inner.fallback.name;
+          else if (inner.kind === "reject")
+            return null;
+        }
+        if (numCtx === void 0) {
+          const trainCtx = await broker.getNctxTrain(model);
+          numCtx = trainCtx && trainCtx > 0 ? Math.min(trainCtx, 4096) : 2048;
+        }
         const imageBase64 = buffer2.toString("base64");
         let ollamaPrompt;
         switch (action) {
@@ -261323,7 +262068,7 @@ Coordinates are normalized (0-1). Multiply by image width/height for pixel value
           let res = await fetch(`${ollamaHost}/api/generate`, {
             method: "POST",
             headers: { "Content-Type": "application/json" },
-            body: JSON.stringify({ model, prompt: ollamaPrompt, images: [imageBase64], stream: false }),
+            body: JSON.stringify({ model, prompt: ollamaPrompt, images: [imageBase64], stream: false, options: { num_ctx: numCtx } }),
             signal: AbortSignal.timeout(6e4)
           });
           if (!res.ok && model === "moondream") {
@@ -261335,15 +262080,18 @@ Coordinates are normalized (0-1). Multiply by image width/height for pixel value
                 res = await fetch(`${ollamaHost}/api/generate`, {
                   method: "POST",
                   headers: { "Content-Type": "application/json" },
-                  body: JSON.stringify({ model, prompt: ollamaPrompt, images: [imageBase64], stream: false }),
+                  body: JSON.stringify({ model, prompt: ollamaPrompt, images: [imageBase64], stream: false, options: { num_ctx: numCtx } }),
                   signal: AbortSignal.timeout(6e4)
                 });
               } catch {
               }
             }
           }
-          if (!res.ok)
+          if (!res.ok) {
+            broker.clearInflight("ollama", model);
             return null;
+          }
+          broker.touch("ollama", model);
           const data = await res.json();
           const response = data.response || "";
           if (!response)
@@ -513030,7 +513778,7 @@ var init_dist4 = __esm({
 });
 
 // packages/execution/dist/tools/code-graph-events.js
-import { EventEmitter as EventEmitter3 } from "node:events";
+import { EventEmitter as EventEmitter4 } from "node:events";
 function emitIndexed(payload) {
   codeGraphEvents.publish({ type: "indexed", timestamp: Date.now(), ...payload });
 }
@@ -513048,7 +513796,7 @@ var init_code_graph_events = __esm({
   "packages/execution/dist/tools/code-graph-events.js"() {
     "use strict";
     MAX_RECENT = 256;
-    CodeGraphEventBus = class extends EventEmitter3 {
+    CodeGraphEventBus = class extends EventEmitter4 {
       ring = [];
       constructor() {
         super();
@@ -519570,6 +520318,7 @@ var init_full_sub_agent = __esm({
   "packages/execution/dist/tools/full-sub-agent.js"() {
     "use strict";
     init_process_kill();
+    init_model_broker();
     _activeSubProcesses = /* @__PURE__ */ new Map();
     FullSubAgentTool = class {
       name = "full_sub_agent";
@@ -519631,10 +520380,45 @@ var init_full_sub_agent = __esm({
             if (!task)
               return { success: false, output: "", error: "task is required", durationMs: performance.now() - start2 };
             const model = String(args["model"] ?? this.model);
+            const broker = getModelBroker();
+            const decision2 = await broker.ensureModelLoadable({
+              name: model || "default",
+              domain: "subagent",
+              host: model ? "ollama" : "subprocess",
+              owner: "full-sub-agent-tool",
+              estimatedVramMB: 4e3,
+              estimatedRamMB: 1500,
+              priority: 1
+              // sub-agents are higher priority than idle background models
+            });
+            if (decision2.kind === "evict") {
+              for (const target of decision2.evictTargets) {
+                await broker.evict(target.host, target.name, "sub-agent-spawn-needs-room");
+              }
+            } else if (decision2.kind === "reject") {
+              return {
+                success: false,
+                output: "",
+                error: `Sub-agent spawn blocked by resource broker: ${decision2.reason}`,
+                durationMs: performance.now() - start2
+              };
+            }
             const entry = spawnFullSubAgent(task, { model, backendUrl: this.backendUrl, workingDir: this.workingDir }, (text) => this.onViewWrite?.(entry.id, text), (id, exitCode, output) => {
               this.onViewStatus?.(id, exitCode === 0 ? "completed" : "failed");
+              broker.unregisterLoaded("subprocess", id, "sub-agent-exited");
               this.onComplete?.(id, task, exitCode, output);
             });
+            broker.registerLoaded({
+              key: `subprocess:${entry.id}`,
+              name: entry.id,
+              domain: "subagent",
+              host: "subprocess",
+              owner: "full-sub-agent-tool",
+              vramMB: 4e3,
+              ramMB: 1500,
+              pid: entry.pid,
+              priority: 1
+            });
             this.onViewRegister?.(entry.id, entry.id, "full");
             return {
               success: true,
@@ -519708,6 +520492,7 @@ var _idCounter, AgentTool;
 var init_agent_tool = __esm({
   "packages/execution/dist/tools/agent-tool.js"() {
     "use strict";
+    init_model_broker();
     _idCounter = 0;
     AgentTool = class {
       name = "agent";
@@ -519798,6 +520583,31 @@ var init_agent_tool = __esm({
         }
         const model = modelOverride ?? this.config.model;
         const agentId = generateAgentId(subagentType);
+        {
+          const broker = getModelBroker();
+          const decision2 = await broker.ensureModelLoadable({
+            name: model || "default",
+            domain: "subagent",
+            host: "ollama",
+            owner: `agent-tool/${subagentType}`,
+            estimatedVramMB: 2500,
+            estimatedRamMB: 1e3,
+            priority: 1
+          });
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              await broker.evict(target.host, target.name, "agent-tool-needs-room");
+            }
+          } else if (decision2.kind === "reject") {
+            return {
+              success: false,
+              output: "",
+              error: `Agent spawn blocked by resource broker: ${decision2.reason}`,
+              durationMs: performance.now() - start2
+            };
+          }
+          broker.touch("ollama", model);
+        }
         const label = description ?? `${subagentType}: ${prompt.slice(0, 40)}`;
         const preloadedFiles = [];
         if (relevantFilePaths.length > 0) {
@@ -522473,13 +523283,13 @@ var init_notebook_edit = __esm({
 
 // packages/execution/dist/tools/environment-snapshot.js
 import { execSync as execSync42 } from "node:child_process";
-import { cpus, totalmem, freemem, hostname as hostname2, platform as platform2, arch, uptime } from "node:os";
+import { cpus, totalmem as totalmem2, freemem as freemem2, hostname as hostname2, platform as platform2, arch, uptime } from "node:os";
 import { statfsSync as statfsSync2 } from "node:fs";
 function collectSnapshot(workingDir) {
   const now = /* @__PURE__ */ new Date();
   const cpuInfo = cpus();
-  const totalRAM = totalmem();
-  const freeRAM = freemem();
+  const totalRAM = totalmem2();
+  const freeRAM = freemem2();
   let load1 = 0, load5 = 0, load15 = 0;
   try {
     const loadavg4 = __require("node:os").loadavg();
@@ -522674,6 +523484,7 @@ var VideoUnderstandTool;
 var init_video_understand = __esm({
   "packages/execution/dist/tools/video-understand.js"() {
     "use strict";
+    init_model_broker();
     VideoUnderstandTool = class {
       name = "video_understand";
       description = "Analyze a video from URL or local file. Produces timestamped transcript aligned with keyframe descriptions. Supports YouTube URLs and direct video files. Pipeline: download → transcribe (Whisper) → extract keyframes (scene detection) → describe frames → align timestamps → save structured output.";
@@ -522704,6 +523515,36 @@ var init_video_understand = __esm({
         if (!url && !localPath) {
           return { success: false, output: "", error: "url or path required", durationMs: performance.now() - start2 };
         }
+        const broker = getModelBroker();
+        const asrDecision = await broker.ensureModelLoadable({
+          name: whisperModel,
+          domain: "asr",
+          host: "whisper-cli",
+          owner: "video-understand-tool"
+        });
+        if (asrDecision.kind === "evict") {
+          for (const target of asrDecision.evictTargets) {
+            await broker.evict(target.host, target.name, "video-understand-needs-asr-room");
+          }
+        } else if (asrDecision.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Video understanding blocked by resource broker (ASR): ${asrDecision.reason}`,
+            durationMs: performance.now() - start2
+          };
+        }
+        const visionDecision = await broker.ensureModelLoadable({
+          name: "moondream2",
+          domain: "vision",
+          host: "moondream-station",
+          owner: "video-understand-tool"
+        });
+        if (visionDecision.kind === "evict") {
+          for (const target of visionDecision.evictTargets) {
+            await broker.evict(target.host, target.name, "video-understand-needs-vision-room");
+          }
+        }
         const outDir = join71(this.workingDir, ".omnius", "video-analysis");
         mkdirSync28(outDir, { recursive: true });
         const tmpDir = join71(outDir, `tmp-${Date.now()}`);
@@ -524314,6 +525155,7 @@ __export(dist_exports, {
   MemorySearchTool: () => MemorySearchTool,
   MemoryWriteTool: () => MemoryWriteTool,
   MeshtasticTool: () => MeshtasticTool,
+  ModelBroker: () => ModelBroker,
   MultimodalMemoryTool: () => MultimodalMemoryTool,
   NetworkEgressPolicyError: () => NetworkEgressPolicyError,
   NexusTool: () => NexusTool,
@@ -524451,6 +525293,7 @@ __export(dist_exports, {
   getFileNotes: () => getFileNotes,
   getFullSubAgent: () => getFullSubAgent,
   getImageGenerationPreset: () => getImageGenerationPreset,
+  getModelBroker: () => getModelBroker,
   getModelStoreDiskInfo: () => getModelStoreDiskInfo,
   getRecentChangesSummary: () => getRecentChangesSummary,
   getSecretsFilePath: () => getSecretsFilePath,
@@ -524580,6 +525423,7 @@ var init_dist5 = __esm({
   "packages/execution/dist/index.js"() {
     "use strict";
     init_tool_executor();
+    init_model_broker();
     init_security_classifier();
     init_tool_manifest();
     init_provenance();
@@ -527339,12 +528183,12 @@ var init_ollama_pool_cleanup = __esm({
 });
 
 // packages/orchestrator/dist/ollama-pool.js
-import { spawn as spawn23, exec } from "node:child_process";
+import { spawn as spawn23, exec as exec2 } from "node:child_process";
 import { existsSync as existsSync63, readFileSync as readFileSync50, readdirSync as readdirSync20, statfsSync as statfsSync3, statSync as statSync25 } from "node:fs";
 import { homedir as homedir28 } from "node:os";
 import { join as join77 } from "node:path";
 import { createServer as createServer3 } from "node:net";
-import { EventEmitter as EventEmitter4 } from "node:events";
+import { EventEmitter as EventEmitter5 } from "node:events";
 function discoverSystemOllamaModelStore() {
   const fromEnv = process.env["OLLAMA_MODELS"]?.trim();
   if (fromEnv && isDirectory(fromEnv))
@@ -527436,7 +528280,7 @@ async function detectGpus() {
   if (_nvidiaSmiAvailable === false)
     return [];
   return new Promise((resolve55) => {
-    exec("nvidia-smi --query-gpu=index,uuid,name,memory.total,memory.free,utilization.gpu --format=csv,noheader,nounits 2>/dev/null", { encoding: "utf8", timeout: 3e3 }, (err, stdout) => {
+    exec2("nvidia-smi --query-gpu=index,uuid,name,memory.total,memory.free,utilization.gpu --format=csv,noheader,nounits 2>/dev/null", { encoding: "utf8", timeout: 3e3 }, (err, stdout) => {
       if (err) {
         _nvidiaSmiAvailable = false;
         resolve55([]);
@@ -527465,7 +528309,7 @@ async function detectGpus() {
   });
 }
 async function getHardwareSnapshot() {
-  const { totalmem: totalmem7, freemem: freemem6, cpus: cpus5 } = await import("node:os");
+  const { totalmem: totalmem8, freemem: freemem7, cpus: cpus5 } = await import("node:os");
   const gpus = await detectGpus();
   const diskPath = discoverSystemOllamaModelStore() ?? homedir28();
   const disk = snapshotDisk(diskPath);
@@ -527473,8 +528317,8 @@ async function getHardwareSnapshot() {
   return {
     gpus,
     cpuCores: cpus5().length,
-    ramTotalMB: Math.round(totalmem7() / (1024 * 1024)),
-    ramFreeMB: Math.round(freemem6() / (1024 * 1024)),
+    ramTotalMB: Math.round(totalmem8() / (1024 * 1024)),
+    ramFreeMB: Math.round(freemem7() / (1024 * 1024)),
     disk,
     network,
     takenAtMs: Date.now()
@@ -527741,7 +528585,7 @@ var init_ollama_pool = __esm({
       return { proc, ready };
     };
     _gpuCursor = 0;
-    OllamaPool = class extends EventEmitter4 {
+    OllamaPool = class extends EventEmitter5 {
       config;
       instances = [];
       reaperHandle = null;
@@ -540331,8 +541175,8 @@ var init_streaming_executor = __esm({
       startExecution(entry) {
         entry.state = "executing";
         entry.startedAt = Date.now();
-        const exec5 = this.executeFn;
-        entry.promise = exec5(entry.name, entry.args).then((result) => {
+        const exec6 = this.executeFn;
+        entry.promise = exec6(entry.name, entry.args).then((result) => {
           entry.state = "completed";
           entry.result = result;
           entry.completedAt = Date.now();
@@ -540621,7 +541465,7 @@ __export(preflightSnapshot_exports, {
 });
 import { existsSync as existsSync79, readFileSync as readFileSync62, statSync as statSync31 } from "node:fs";
 import { execSync as execSync46 } from "node:child_process";
-import { homedir as homedir29, platform as platform3, arch as arch2, totalmem as totalmem2, freemem as freemem2, hostname as hostname3 } from "node:os";
+import { homedir as homedir29, platform as platform3, arch as arch2, totalmem as totalmem3, freemem as freemem3, hostname as hostname3 } from "node:os";
 import { join as join92 } from "node:path";
 import { createHash as createHash17 } from "node:crypto";
 function capturePreflightSnapshot(workingDir) {
@@ -540668,8 +541512,8 @@ function capturePreflightSnapshot(workingDir) {
       platform: platform3(),
       arch: arch2(),
       nodeVersion: process.version,
-      totalMemBytes: totalmem2(),
-      freeMemBytes: freemem2(),
+      totalMemBytes: totalmem3(),
+      freeMemBytes: freemem3(),
       hostname: hostname3()
     },
     toolchain: captureToolchainVersions(),
@@ -559212,7 +560056,7 @@ import { existsSync as existsSync85, mkdirSync as mkdirSync46, writeFileSync as
 import { join as join100, dirname as dirname27 } from "node:path";
 import { homedir as homedir32 } from "node:os";
 import { fileURLToPath as fileURLToPath11 } from "node:url";
-import { EventEmitter as EventEmitter5 } from "node:events";
+import { EventEmitter as EventEmitter6 } from "node:events";
 import { createInterface as createInterface2 } from "node:readline";
 function isAudioPath(path12) {
   const ext = path12.toLowerCase().split(".").pop();
@@ -559451,9 +560295,9 @@ function ensureTranscribeCliBackground() {
     } catch {
     }
     try {
-      const { exec: exec5 } = await import("node:child_process");
+      const { exec: exec6 } = await import("node:child_process");
       return new Promise((resolve55) => {
-        exec5("npm i -g transcribe-cli", { timeout: 18e4 }, (err) => {
+        exec6("npm i -g transcribe-cli", { timeout: 18e4 }, (err) => {
           resolve55(!err);
         });
       });
@@ -559501,7 +560345,7 @@ var init_listen = __esm({
       ".m4v",
       ".ts"
     ]);
-    WhisperFallbackTranscriber = class extends EventEmitter5 {
+    WhisperFallbackTranscriber = class extends EventEmitter6 {
       constructor(model, scriptPath2) {
         super();
         this.model = model;
@@ -559610,7 +560454,7 @@ var init_listen = __esm({
         this._ready = false;
       }
     };
-    ListenEngine = class extends EventEmitter5 {
+    ListenEngine = class extends EventEmitter6 {
       config;
       micProcess = null;
       liveTranscriber = null;
@@ -562355,7 +563199,7 @@ var require_extension3 = __commonJS({
 var require_websocket3 = __commonJS({
   "node_modules/.pnpm/ws@8.20.1/node_modules/ws/lib/websocket.js"(exports, module) {
     "use strict";
-    var EventEmitter14 = __require("events");
+    var EventEmitter15 = __require("events");
     var https4 = __require("https");
     var http6 = __require("http");
     var net5 = __require("net");
@@ -562387,7 +563231,7 @@ var require_websocket3 = __commonJS({
     var protocolVersions = [8, 13];
     var readyStates = ["CONNECTING", "OPEN", "CLOSING", "CLOSED"];
     var subprotocolRegex = /^[!#$%&'*+\-.0-9A-Z^_`|a-z~]+$/;
-    var WebSocket6 = class _WebSocket extends EventEmitter14 {
+    var WebSocket6 = class _WebSocket extends EventEmitter15 {
       /**
        * Create a new `WebSocket`.
        *
@@ -563384,7 +564228,7 @@ var require_subprotocol2 = __commonJS({
 var require_websocket_server2 = __commonJS({
   "node_modules/.pnpm/ws@8.20.1/node_modules/ws/lib/websocket-server.js"(exports, module) {
     "use strict";
-    var EventEmitter14 = __require("events");
+    var EventEmitter15 = __require("events");
     var http6 = __require("http");
     var { Duplex: Duplex3 } = __require("stream");
     var { createHash: createHash31 } = __require("crypto");
@@ -563397,7 +564241,7 @@ var require_websocket_server2 = __commonJS({
     var RUNNING = 0;
     var CLOSING = 1;
     var CLOSED = 2;
-    var WebSocketServer4 = class extends EventEmitter14 {
+    var WebSocketServer4 = class extends EventEmitter15 {
       /**
        * Create a `WebSocketServer` instance.
        *
@@ -564285,6 +565129,9 @@ var init_command_registry = __esm({
       ["/wizard", "Alias for /setup"],
       ["/parallel", "Show current Ollama parallel inference slots"],
       ["/parallel <1-15>", "Set parallel slots (restarts Ollama, max 15)"],
+      ["/broker", "Show resource broker status — loaded models, in-flight loads, RAM/VRAM headroom"],
+      ["/broker evict <host>:<name>", "Force eviction of a tracked model (e.g. /broker evict ollama:moondream)"],
+      ["/broker threshold <ram|vram|idle> <v>", "Tune broker thresholds (RAM/VRAM in MB, idle in seconds)"],
       ["/ollama cleanup", "Dry-run stale Ollama pool process cleanup"],
       ["/ollama cleanup --execute", "Terminate guarded stale Ollama pool runners"],
       ["/mcp", "Show MCP server/tool status and controls"],
@@ -564554,6 +565401,7 @@ var init_command_registry = __esm({
       endpoint: "network",
       provider: "network",
       parallel: "runtime",
+      broker: "runtime",
       setup: "runtime",
       wizard: "runtime",
       mcp: "tools",
@@ -564695,6 +565543,7 @@ var init_command_registry = __esm({
       "model",
       "endpoint",
       "parallel",
+      "broker",
       "commands",
       "cmds",
       "selfmodify",
@@ -565588,7 +566437,7 @@ var init_task_complete_box = __esm({
 });
 
 // packages/cli/src/tui/model-picker.ts
-import { totalmem as totalmem3 } from "node:os";
+import { totalmem as totalmem4 } from "node:os";
 function isImageGenModel(name10, family) {
   return IMAGE_GEN_PATTERNS.some((p2) => p2.test(name10) || family && p2.test(family));
 }
@@ -565977,7 +566826,7 @@ async function queryModelContextSize(baseUrl, modelName) {
   }
 }
 function estimateRealisticContext(kvBytesPerToken, archMax, modelSizeGB2) {
-  const totalMemGB = totalmem3() / 1024 ** 3;
+  const totalMemGB = totalmem4() / 1024 ** 3;
   const usableBytes = totalMemGB * 0.7 * 1024 ** 3;
   const maxTokens = Math.floor(usableBytes / kvBytesPerToken);
   let numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
@@ -567316,7 +568165,7 @@ var init_render = __esm({
 // packages/cli/src/tui/voice-session.ts
 import { createServer as createServer4 } from "node:http";
 import { spawn as spawn25, execSync as execSync49 } from "node:child_process";
-import { EventEmitter as EventEmitter6 } from "node:events";
+import { EventEmitter as EventEmitter7 } from "node:events";
 function generateFrontendHTML() {
   return `<!DOCTYPE html>
 <html lang="en">
@@ -568051,7 +568900,7 @@ var init_voice_session = __esm({
     init_wrapper2();
     init_render();
     init_typed_node_events();
-    VoiceSession = class extends EventEmitter6 {
+    VoiceSession = class extends EventEmitter7 {
       state;
       server = null;
       wss = null;
@@ -569166,11 +570015,11 @@ var init_voice_soul = __esm({
 
 // packages/cli/src/tui/expose.ts
 import { createServer as createServer5, request as httpRequest } from "node:http";
-import { spawn as spawn26, exec as exec2 } from "node:child_process";
-import { EventEmitter as EventEmitter7 } from "node:events";
+import { spawn as spawn26, exec as exec3 } from "node:child_process";
+import { EventEmitter as EventEmitter8 } from "node:events";
 import { randomBytes as randomBytes19, timingSafeEqual } from "node:crypto";
 import { URL as URL2 } from "node:url";
-import { loadavg, cpus as cpus2, totalmem as totalmem4, freemem as freemem3 } from "node:os";
+import { loadavg, cpus as cpus2, totalmem as totalmem5, freemem as freemem4 } from "node:os";
 import { existsSync as existsSync88, readFileSync as readFileSync70, writeFileSync as writeFileSync44, unlinkSync as unlinkSync14, mkdirSync as mkdirSync48, readdirSync as readdirSync29, statSync as statSync34, statfsSync as statfsSync4 } from "node:fs";
 import { join as join103 } from "node:path";
 function cleanForwardHeaders(raw, targetHost) {
@@ -569276,8 +570125,8 @@ function parseRateLimitHeaders(headers) {
 async function collectSystemMetricsAsync() {
   const [l1, l5, l15] = loadavg();
   const cores = cpus2().length;
-  const totalMem = totalmem4();
-  const freeMem = freemem3();
+  const totalMem = totalmem5();
+  const freeMem = freemem4();
   const usedMem = totalMem - freeMem;
   let disk = {
     path: process.cwd(),
@@ -569310,7 +570159,7 @@ async function collectSystemMetricsAsync() {
   };
   try {
     const smi = await new Promise((resolve55, reject) => {
-      exec2(
+      exec3(
         "nvidia-smi --query-gpu=utilization.gpu,memory.used,memory.total,name --format=csv,noheader,nounits 2>/dev/null",
         { encoding: "utf8", timeout: 3e3 },
         (err, stdout) => err ? reject(err) : resolve55(stdout)
@@ -569398,7 +570247,7 @@ var init_expose = __esm({
       custom: "http://127.0.0.1:11434"
     };
     STATE_FILE_NAME = "expose-state.json";
-    ExposeGateway = class _ExposeGateway extends EventEmitter7 {
+    ExposeGateway = class _ExposeGateway extends EventEmitter8 {
       constructor(options2) {
         super();
         this.options = options2;
@@ -570277,7 +571126,7 @@ ${this.formatConnectionInfo()}`);
       }
     };
     P2P_STATE_FILE_NAME = "expose-p2p-state.json";
-    ExposeP2PGateway = class _ExposeP2PGateway extends EventEmitter7 {
+    ExposeP2PGateway = class _ExposeP2PGateway extends EventEmitter8 {
       _nexusTool;
       // NexusTool instance
       _kind;
@@ -571048,7 +571897,7 @@ var init_secret_vault = __esm({
 });
 
 // packages/cli/src/tui/p2p/peer-mesh.ts
-import { EventEmitter as EventEmitter8 } from "node:events";
+import { EventEmitter as EventEmitter9 } from "node:events";
 import { createServer as createServer6 } from "node:http";
 import { randomBytes as randomBytes21, createHash as createHash21, generateKeyPairSync } from "node:crypto";
 var PING_INTERVAL_MS, PEER_TIMEOUT_MS, GOSSIP_INTERVAL_MS, MAX_PEERS, PeerMesh;
@@ -571060,7 +571909,7 @@ var init_peer_mesh = __esm({
     PEER_TIMEOUT_MS = 9e4;
     GOSSIP_INTERVAL_MS = 6e4;
     MAX_PEERS = 50;
-    PeerMesh = class extends EventEmitter8 {
+    PeerMesh = class extends EventEmitter9 {
       constructor(options2) {
         super();
         this.options = options2;
@@ -571506,7 +572355,7 @@ var init_peer_mesh = __esm({
 });
 
 // packages/cli/src/tui/p2p/inference-router.ts
-import { EventEmitter as EventEmitter9 } from "node:events";
+import { EventEmitter as EventEmitter10 } from "node:events";
 var TRUST_WEIGHTS, InferenceRouter;
 var init_inference_router = __esm({
   "packages/cli/src/tui/p2p/inference-router.ts"() {
@@ -571518,7 +572367,7 @@ var init_inference_router = __esm({
       verified: 5,
       public: 2
     };
-    InferenceRouter = class extends EventEmitter9 {
+    InferenceRouter = class extends EventEmitter10 {
       mesh;
       vault;
       defaultTimeoutMs;
@@ -571716,7 +572565,7 @@ var init_p2p = __esm({
 });
 
 // packages/cli/src/tui/call-agent.ts
-import { EventEmitter as EventEmitter10 } from "node:events";
+import { EventEmitter as EventEmitter11 } from "node:events";
 import crypto13 from "node:crypto";
 function adaptTool(tool) {
   return {
@@ -571774,7 +572623,7 @@ var init_call_agent = __esm({
       }
     };
     _globalFeed = null;
-    CallSubAgent = class extends EventEmitter10 {
+    CallSubAgent = class extends EventEmitter11 {
       tier;
       clientId;
       runner = null;
@@ -573876,8 +574725,8 @@ __export(system_metrics_exports, {
   getInstantSnapshot: () => getInstantSnapshot,
   instantaneousCpuPct: () => instantaneousCpuPct
 });
-import { loadavg as loadavg2, cpus as cpus3, totalmem as totalmem5, freemem as freemem4, platform as platform4 } from "node:os";
-import { exec as exec3 } from "node:child_process";
+import { loadavg as loadavg2, cpus as cpus3, totalmem as totalmem6, freemem as freemem5, platform as platform4 } from "node:os";
+import { exec as exec4 } from "node:child_process";
 import { readFile as readFile22 } from "node:fs/promises";
 function formatRate(bytesPerSec) {
   if (bytesPerSec < 1024) return `${Math.round(bytesPerSec)}B`;
@@ -573924,7 +574773,7 @@ async function collectNetworkMetrics() {
   if (plat === "darwin") {
     try {
       const output = await new Promise((resolve55, reject) => {
-        exec3("netstat -ib 2>/dev/null | head -30", { encoding: "utf8", timeout: 3e3 }, (err, stdout) => err ? reject(err) : resolve55(stdout));
+        exec4("netstat -ib 2>/dev/null | head -30", { encoding: "utf8", timeout: 3e3 }, (err, stdout) => err ? reject(err) : resolve55(stdout));
       });
       let rxBytes = 0, txBytes = 0;
       for (const line of output.split("\n")) {
@@ -573967,7 +574816,7 @@ async function collectGpuMetrics() {
   if (_nvidiaSmiAvailable2 === false) return noGpu;
   try {
     const smi = await new Promise((resolve55, reject) => {
-      exec3(
+      exec4(
         "nvidia-smi --query-gpu=index,uuid,utilization.gpu,memory.used,memory.total,name --format=csv,noheader,nounits 2>/dev/null",
         { encoding: "utf8", timeout: 3e3 },
         (err, stdout) => err ? reject(err) : resolve55(stdout)
@@ -574066,8 +574915,8 @@ function instantaneousCpuPct() {
 function collectCpuRam() {
   const cores = cpus3().length;
   const cpuModel = cpus3()[0]?.model ?? "";
-  const totalMem = totalmem5();
-  const usedMem = totalMem - freemem4();
+  const totalMem = totalmem6();
+  const usedMem = totalMem - freemem5();
   let cpuUtil = instantaneousCpuPct();
   if (cpuUtil < 0) {
     const [l1] = loadavg2();
@@ -574862,6 +575711,7 @@ var init_status_bar = __esm({
     init_text_selection();
     init_daemon_registry();
     init_overlay_lock();
+    init_dist5();
     init_theme();
     init_layout2();
     EXPERT_TOOL_BASELINES = {
@@ -576073,6 +576923,10 @@ var init_status_bar = __esm({
           this._unifiedMetrics = m2;
           if (this.active) this.renderFooterPreserveCursor();
         }, intervalMs);
+        try {
+          getModelBroker().startPolling(Math.max(2e3, intervalMs * 2));
+        } catch {
+        }
       }
       /** Stop all metrics collection (local and remote) */
       stopAllMetrics() {
@@ -580498,7 +581352,7 @@ __export(setup_exports, {
   updateOllama: () => updateOllama
 });
 import * as readline from "node:readline";
-import { execSync as execSync51, spawn as spawn28, exec as exec4 } from "node:child_process";
+import { execSync as execSync51, spawn as spawn28, exec as exec5 } from "node:child_process";
 import { promisify as promisify6 } from "node:util";
 import { existsSync as existsSync95, writeFileSync as writeFileSync49, readFileSync as readFileSync78, appendFileSync as appendFileSync6, mkdirSync as mkdirSync53 } from "node:fs";
 import { join as join111 } from "node:path";
@@ -583250,7 +584104,7 @@ var init_setup = __esm({
     init_dist();
     init_tui_select();
     init_listen();
-    execAsync2 = promisify6(exec4);
+    execAsync2 = promisify6(exec5);
     OMNIUS_FIRST_RUN_BANNER = [
       " ░▒▓██████▓▒░░▒▓██████████████▓▒░░▒▓███████▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░░▒▓███████▓▒░ ",
       "░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░▒▓█▓▒░░▒▓█▓▒░▒▓█▓▒░        ",
@@ -595704,6 +596558,9 @@ async function handleSlashCommand(input, ctx3) {
     case "parallel":
       await handleParallel(arg, ctx3);
       return "handled";
+    case "broker":
+      await handleBroker(arg, ctx3);
+      return "handled";
     case "ollama":
       await handleOllama(arg, ctx3);
       return "handled";
@@ -603161,6 +604018,83 @@ async function handlePeerEndpoint(peerId, authKey, ctx3, local) {
     );
   }
 }
+async function handleBroker(arg, _ctx) {
+  const broker = getModelBroker();
+  const sub = (arg || "").trim().toLowerCase();
+  const snap = await broker.pollOnce();
+  if (!sub || sub === "status" || sub === "list" || sub === "ls") {
+    safeLog("");
+    safeLog(`  ${c3.bold("Resource Broker")}`);
+    safeLog("");
+    safeLog(`  ${c3.dim("RAM:")}  ${snap.ramMB.used} / ${snap.ramMB.total} MB used (${snap.ramMB.free} MB free)`);
+    if (snap.vramMB) {
+      safeLog(`  ${c3.dim("VRAM:")} ${snap.vramMB.used} / ${snap.vramMB.total} MB used (${snap.vramMB.free} MB free)`);
+    } else {
+      safeLog(`  ${c3.dim("VRAM:")} ${c3.dim("(no GPU detected)")}`);
+    }
+    safeLog(`  ${c3.dim("RAM headroom threshold:")}  ${broker.ramHeadroomMB} MB`);
+    safeLog(`  ${c3.dim("VRAM headroom threshold:")} ${broker.vramHeadroomMB} MB`);
+    safeLog(`  ${c3.dim("Idle-evict threshold:")} ${Math.round(broker.idleEvictMs / 1e3)}s`);
+    safeLog("");
+    if (snap.loaded.length === 0) {
+      safeLog(`  ${c3.dim("No loaded models tracked.")}`);
+    } else {
+      safeLog(`  ${c3.bold("Loaded models:")}`);
+      const now = Date.now();
+      for (const m2 of snap.loaded) {
+        const idle = Math.round((now - m2.lastUsedAt) / 1e3);
+        const owner = m2.owner ? c3.dim(` [owner=${m2.owner}]`) : "";
+        const ctx3 = m2.numCtx ? c3.dim(` n_ctx=${m2.numCtx}`) : "";
+        safeLog(`    ${c3.cyan(m2.name)} (${m2.host}/${m2.domain})  vram=${m2.vramMB}MB ram=${m2.ramMB}MB  idle=${idle}s${ctx3}${owner}`);
+      }
+    }
+    if (snap.inflight.length > 0) {
+      safeLog("");
+      safeLog(`  ${c3.bold("In-flight loads:")}`);
+      for (const f2 of snap.inflight) {
+        const age = Math.round((Date.now() - f2.startedMs) / 1e3);
+        safeLog(`    ${c3.yellow(f2.key)}  owner=${f2.owner}  ${age}s ago`);
+      }
+    }
+    safeLog("");
+    safeLog(`  ${c3.dim("Subcommands: /broker [status|evict <key>|threshold ram|vram|idle <value>]")}`);
+    safeLog("");
+    return;
+  }
+  const parts = (arg || "").trim().split(/\s+/);
+  const op = parts[0]?.toLowerCase();
+  if (op === "evict") {
+    const key = parts[1];
+    if (!key || !key.includes(":")) {
+      renderWarning("Usage: /broker evict <host>:<name>  e.g. /broker evict ollama:moondream");
+      return;
+    }
+    const [host, ...rest] = key.split(":");
+    const name10 = rest.join(":");
+    const ok3 = await broker.evict(host, name10, "user-requested");
+    if (ok3) renderInfo(`Evicted ${key} (actively unloaded)`);
+    else renderInfo(`Unregistered ${key} (could not actively unload; subprocess may need manual cleanup)`);
+    return;
+  }
+  if (op === "threshold") {
+    const which3 = parts[1]?.toLowerCase();
+    const value2 = parts[2] ? Number(parts[2]) : NaN;
+    if (!which3 || !Number.isFinite(value2) || value2 < 0) {
+      renderWarning("Usage: /broker threshold <ram|vram|idle> <value>  (ram/vram in MB, idle in seconds)");
+      return;
+    }
+    if (which3 === "ram") broker.ramHeadroomMB = value2;
+    else if (which3 === "vram") broker.vramHeadroomMB = value2;
+    else if (which3 === "idle") broker.idleEvictMs = value2 * 1e3;
+    else {
+      renderWarning("Unknown threshold; use ram|vram|idle");
+      return;
+    }
+    renderInfo(`Updated broker threshold ${which3} = ${value2}`);
+    return;
+  }
+  renderWarning("Unknown /broker subcommand. Try: status | evict <host>:<name> | threshold <ram|vram|idle> <value>");
+}
 async function handleParallel(arg, ctx3) {
   const { execSync: execSync61 } = await import("node:child_process");
   const baseUrl = ctx3.config.backendUrl || "http://localhost:11434";
@@ -604173,9 +605107,9 @@ async function handleUpdate(subcommand, ctx3) {
       }
     };
   }
-  const { exec: exec5, spawn: spawn34, execSync: es2 } = await import("node:child_process");
+  const { exec: exec6, spawn: spawn34, execSync: es2 } = await import("node:child_process");
   const execA = (cmd, opts) => new Promise(
-    (res, rej) => exec5(
+    (res, rej) => exec6(
       cmd,
       {
         encoding: "utf8",
@@ -604869,7 +605803,7 @@ async function handleUpdate(subcommand, ctx3) {
     installOverlay.setPhase("Native Modules");
     installOverlay.setStatus("Rebuilding native modules...");
     await new Promise((resolve55) => {
-      const child = exec5(
+      const child = exec6(
         `${sudoPrefix}npm rebuild -g omnius 2>/dev/null || true`,
         { timeout: 12e4 },
         () => resolve55(true)
@@ -604911,7 +605845,7 @@ async function handleUpdate(subcommand, ctx3) {
     if (fsExists(venvPip2)) {
       installOverlay.setStatus("Upgrading Python packages...");
       await new Promise((resolve55) => {
-        const child = exec5(
+        const child = exec6(
           `"${venvPip2}" install --upgrade moondream-station pytesseract Pillow opencv-python-headless numpy 2>/dev/null || true`,
           { timeout: 3e5 },
           (err) => resolve55(!err)
@@ -610886,6 +611820,9 @@ var init_bless_engine = __esm({
       async pingModel() {
         try {
           const url = `${this.config.backendUrl}/api/chat`;
+          const broker = getModelBroker();
+          const trainCtx = await broker.getNctxTrain(this.config.model);
+          const numCtx = trainCtx && trainCtx > 0 ? Math.min(trainCtx, 8192) : void 0;
           await fetch(url, {
             method: "POST",
             headers: { "Content-Type": "application/json" },
@@ -610893,11 +611830,12 @@ var init_bless_engine = __esm({
               model: this.config.model,
               messages: [{ role: "user", content: "." }],
               stream: false,
-              options: { num_predict: 1 },
+              options: numCtx ? { num_predict: 1, num_ctx: numCtx } : { num_predict: 1 },
               keep_alive: "30m"
             }),
             signal: AbortSignal.timeout(15e3)
           });
+          broker.touch("ollama", this.config.model);
           this.state.keepAlivePings++;
         } catch {
         }
@@ -616279,19 +617217,51 @@ async function queryVisionModel(modelName, imagePath, prompt = "Describe what yo
   if (!existsSync116(imagePath)) return "";
   const imageBuffer = readFileSync95(imagePath);
   const base64Image = imageBuffer.toString("base64");
+  const broker = getModelBroker();
+  const decision2 = await broker.ensureModelLoadable({
+    name: modelName,
+    domain: "vision",
+    host: "ollama",
+    owner: "vision-ingress",
+    requestedNumCtx: 2048
+  });
+  let effectiveModel = modelName;
+  let numCtx;
+  if (decision2.kind === "reject") {
+    return "";
+  } else if (decision2.kind === "degrade") {
+    effectiveModel = decision2.fallback.name;
+  } else if (decision2.kind === "evict") {
+    for (const target of decision2.evictTargets) {
+      await broker.evict(target.host, target.name, "make-room-for-vision");
+    }
+    numCtx = decision2.effectiveNumCtx;
+  } else if (decision2.kind === "ok") {
+    numCtx = decision2.effectiveNumCtx;
+  } else if (decision2.kind === "wait-for-inflight") {
+    const inner = await decision2.promise;
+    if (inner.kind === "ok") numCtx = inner.effectiveNumCtx;
+    else if (inner.kind === "degrade") effectiveModel = inner.fallback.name;
+    else if (inner.kind === "reject") return "";
+  }
+  if (numCtx === void 0) {
+    const trainCtx = await broker.getNctxTrain(effectiveModel);
+    numCtx = trainCtx && trainCtx > 0 ? Math.min(trainCtx, 4096) : 2048;
+  }
   try {
     const response = await fetch("http://localhost:11434/api/generate", {
       method: "POST",
       headers: { "Content-Type": "application/json" },
       body: JSON.stringify({
-        model: modelName,
+        model: effectiveModel,
         prompt,
         images: [base64Image],
         stream: false,
-        options: { temperature: 0.3, num_predict: 1024 }
+        options: { temperature: 0.3, num_predict: 1024, num_ctx: numCtx }
       })
     });
     if (!response.ok) return "";
+    broker.touch("ollama", effectiveModel);
     const data = await response.json();
     return (data.response || "").trim();
   } catch {
@@ -616344,6 +617314,7 @@ function formatImageContextPrefix(result) {
 var init_vision_ingress = __esm({
   "packages/cli/src/tui/vision-ingress.ts"() {
     "use strict";
+    init_dist5();
   }
 });
 
@@ -629030,12 +630001,12 @@ var direct_input_exports = {};
 __export(direct_input_exports, {
   DirectInput: () => DirectInput
 });
-import { EventEmitter as EventEmitter11 } from "node:events";
+import { EventEmitter as EventEmitter12 } from "node:events";
 var DirectInput;
 var init_direct_input = __esm({
   "packages/cli/src/tui/direct-input.ts"() {
     "use strict";
-    DirectInput = class extends EventEmitter11 {
+    DirectInput = class extends EventEmitter12 {
       /** Current input line text */
       line = "";
       /** Cursor position within .line (0-based) */
@@ -629754,7 +630725,7 @@ var voicechat_exports = {};
 __export(voicechat_exports, {
   VoiceChatSession: () => VoiceChatSession
 });
-import { EventEmitter as EventEmitter12 } from "node:events";
+import { EventEmitter as EventEmitter13 } from "node:events";
 function clamp0110(x) {
   return x < 0 ? 0 : x > 1 ? 1 : x;
 }
@@ -629864,7 +630835,7 @@ Rules:
 - Prefer tools for factual queries; otherwise, answer directly with a short reply.`;
     MIN_SIGNAL_SCORE = 0.4;
     NOISE_ONLY_RE = /^(?:[.·…\s,;:!?\-–—_()\[\]{}"'`]+|(?:uh|um|erm|hmm|mm+|uhh+|umm+)[\s.!?]*)+$/i;
-    VoiceChatSession = class extends EventEmitter12 {
+    VoiceChatSession = class extends EventEmitter13 {
       voice;
       listen;
       backendUrl;
@@ -630393,7 +631364,7 @@ __export(voice_runtime_exports, {
   synthesizeToWav: () => synthesizeToWav,
   unregisterClient: () => unregisterClient
 });
-import { EventEmitter as EventEmitter13 } from "node:events";
+import { EventEmitter as EventEmitter14 } from "node:events";
 function getVoiceEngine() {
   if (!_voiceEngine) {
     _voiceEngine = new VoiceEngine();
@@ -630405,7 +631376,7 @@ function getDaemonListenEngine() {
   return _listenEngine;
 }
 function getVoiceBus() {
-  if (!_bus) _bus = new EventEmitter13();
+  if (!_bus) _bus = new EventEmitter14();
   return _bus;
 }
 function getRuntimeStatus() {
@@ -661368,8 +662339,8 @@ NEW TASK: ${fullInput}`;
         const updateInfo = await checkForUpdate(version4);
         if (updateInfo) {
           _autoUpdatedThisSession = true;
-          const { exec: exec5 } = await import("node:child_process");
-          exec5(
+          const { exec: exec6 } = await import("node:child_process");
+          exec6(
             `npm install -g omnius@latest --prefer-online`,
             { timeout: 18e4 },
             (err) => {