turbollm 1.11.6 → 1.11.7

package/dist/cli.js

@@ -127,6 +127,7 @@ function defaultConfig() {
     benchResults: {},
     vramHeadroomMb: VRAM_HEADROOM_DEFAULT_MB,
     lastLoaded: { modelKey: "", engineId: "" },
+    selectedRemoteModel: "",
     autoLoadOnStart: false,
     hf: { token: "" },
     modelDefaults: { ctx: 8192, ngl: 99, imageMaxTokens: 0, maxTokens: 0 },
@@ -262,6 +263,7 @@ function normalize(c) {
   c.hf = { ...d.hf, ...c.hf ?? {} };
   c.modelDefaults = { ...d.modelDefaults, ...c.modelDefaults ?? {} };
   c.lastLoaded = { ...d.lastLoaded, ...c.lastLoaded ?? {} };
+  if (typeof c.selectedRemoteModel !== "string") c.selectedRemoteModel = "";
   c.apiKeys ??= [];
   c.links ??= [];
   c.engines ??= [];
@@ -7042,6 +7044,8 @@ ${list}
       const refreshed = await fetchStatus(base2, _fetch);
       if (refreshed) status = refreshed;
     }
+  } else if (status?.selectedRemoteModel && (await fetchGatewayModelIds(base2, _fetch)).includes(status.selectedRemoteModel)) {
+    remoteModel = status.selectedRemoteModel;
   } else if (!alreadyRunning) {
     const models = await fetchModels(base2, _fetch);
     if (models.length === 0) {
@@ -7085,7 +7089,7 @@ Open ${base2} \u2192 Models \u2192 Load a model, then run this again.
     pinnedModel = status.model.key ?? model;
   }
   const parallelSlots = status.engine?.parallelSlots;
-  const modelNote = modelKey ? `model: ${model}` : `using loaded model: ${model}`;
+  const modelNote = modelKey ? `model: ${model}` : remoteModel ? `using selected remote model: ${model}` : `using loaded model: ${model}`;
   process.stdout.write(`\u25B8 Launching ${spec.label} \u2192 TurboLLM  (${modelNote}, ${base2})
 `);
   const libraryModels = spec.prepareConfig ? await fetchModels(base2, _fetch) : [];
@@ -9315,6 +9319,9 @@ function amdApuOnly(info = getSysInfo()) {
   const amd = info.gpus.filter((g) => g.vendor === "amd");
   return amd.length > 0 && amd.every((g) => g.unified);
 }
+function unifiedMemoryOnly(info = getSysInfo()) {
+  return info.gpus.length > 0 && info.gpus.every((g) => g.unified === true);
+}
 function classifyVendor(name) {
   const n = name.toLowerCase();
   if (/nvidia|geforce|rtx|gtx|quadro|tesla/.test(n)) return "nvidia";
@@ -9450,7 +9457,8 @@ function parseRocmSmi(memJson, nameJson = "") {
     const nameFields = names[card] ?? {};
     const nameKey = Object.keys(nameFields).find((k) => /(Card Series|Card Model|Product Name|Device Name)/i.test(k));
     const name = (nameKey ? String(nameFields[nameKey]).trim() : "") || "AMD Radeon GPU";
-    gpus.push({ name, vramMb: Math.round(bytes / 1e6), vendor: "amd" });
+    const integrated = isIntegratedGpuName(name);
+    gpus.push({ name, vramMb: Math.round(bytes / 1e6), vendor: "amd", ...integrated ? { unified: true } : {} });
   }
   return gpus;
 }
@@ -10112,9 +10120,22 @@ function reportChatBenchResult(d, ms, stats) {
   } catch {
   }
 }
-async function runGeneration(d, emit2, ctx) {
+function resilientSink(sink) {
+  let gone = false;
+  const wrapped = async (ev) => {
+    if (gone) return;
+    try {
+      await sink(ev);
+    } catch {
+      gone = true;
+    }
+  };
+  return Object.assign(wrapped, { clientGone: () => gone });
+}
+async function runGeneration(d, rawEmit, ctx) {
   const { db: db2 } = d;
   const { convId, conv, assistantMsg, upstream, ac, thinkingBudget, reasoningEffort } = ctx;
+  const emit2 = resilientSink(rawEmit);
   const localAccounting = !upstream.remote;
   const convS = conv.sampling ?? {};
   const engineKind = d.registry.active()?.kind ?? "";
@@ -10190,6 +10211,8 @@ After reviewing the search results, include a confidence assessment on a line by
   let aborted = false;
   let engineFailed = false;
   let liveOut = 0;
+  let pendingParse = null;
+  let lastPromptTotal = 0;
   noteLocalActivity();
   if (localAccounting) d.manager.generationStart();
   try {
@@ -10242,7 +10265,10 @@ After reviewing the search results, include a confidence assessment on a line by
       const reader = res.body.getReader();
       const decoder = new TextDecoder();
       let buf = "";
-      const cancelReader = () => void reader.cancel();
+      const cancelReader = () => {
+        void reader.cancel().catch(() => {
+        });
+      };
       if (ac.signal.aborted) {
         cancelReader();
       } else {
@@ -10250,6 +10276,7 @@ After reviewing the search results, include a confidence assessment on a line by
       }
       let roundContent = "";
       let parseState = initParseState();
+      pendingParse = parseState;
       let finishReason = "";
       const pendingToolCalls = /* @__PURE__ */ new Map();
       const announcedToolCalls = /* @__PURE__ */ new Set();
@@ -10271,6 +10298,7 @@ After reviewing the search results, include a confidence assessment on a line by
           }
           const pp = chunk.prompt_progress;
           if (pp && pp.total) {
+            lastPromptTotal = pp.total;
             const pct = Math.round((pp.processed ?? 0) / pp.total * 100);
             if (localAccounting) d.manager.setLiveGen({ phase: "prompt", pct, outputTokens: 0 });
             await emit2({ event: "progress", data: { phase: "prompt", processed: pp.processed, total: pp.total, pct, tps: pp.tps ?? 0 } });
@@ -10310,6 +10338,7 @@ After reviewing the search results, include a confidence assessment on a line by
           if (!raw_content) continue;
           const { state: nextState, events: parseEvents } = feedChunk(parseState, raw_content);
           parseState = nextState;
+          pendingParse = parseState;
           for (const ev of parseEvents) {
             if (ev.type === "reasoning") {
               if (!thinkStart) thinkStart = Date.now();
@@ -10328,6 +10357,7 @@ After reviewing the search results, include a confidence assessment on a line by
         }
       }
       ac.signal.removeEventListener("abort", cancelReader);
+      pendingParse = null;
       for (const ev of flushState(parseState)) {
         if (ev.type === "reasoning") {
           fullReasoning += ev.text;
@@ -10480,13 +10510,17 @@ After reviewing the search results, include a confidence assessment on a line by
         const reader = res.body.getReader();
         const decoder = new TextDecoder();
         let buf = "";
-        const cancelReader = () => void reader.cancel();
+        const cancelReader = () => {
+          void reader.cancel().catch(() => {
+          });
+        };
         if (ac.signal.aborted) {
           cancelReader();
         } else {
           ac.signal.addEventListener("abort", cancelReader, { once: true });
         }
         let parseState = initParseState();
+        pendingParse = parseState;
         roundLoop: while (true) {
           const { done, value } = await reader.read();
           if (done) break;
@@ -10505,6 +10539,8 @@ After reviewing the search results, include a confidence assessment on a line by
             }
             if (chunk.usage) finalUsage = chunk.usage;
             if (chunk.timings) finalTimings = chunk.timings;
+            const pp2 = chunk.prompt_progress;
+            if (pp2?.total) lastPromptTotal = pp2.total;
             const choices = chunk.choices;
             if (!choices?.length) continue;
             const delta = choices[0].delta ?? {};
@@ -10518,6 +10554,7 @@ After reviewing the search results, include a confidence assessment on a line by
             if (!raw_content) continue;
             const { state: nextState, events: parseEvents } = feedChunk(parseState, raw_content);
             parseState = nextState;
+            pendingParse = parseState;
             for (const ev of parseEvents) {
               if (ev.type === "reasoning") {
                 fullReasoning += ev.text;
@@ -10533,6 +10570,7 @@ After reviewing the search results, include a confidence assessment on a line by
           }
         }
         ac.signal.removeEventListener("abort", cancelReader);
+        pendingParse = null;
         for (const ev of flushState(parseState)) {
           if (ev.type === "reasoning") {
             fullReasoning += ev.text;
@@ -10549,23 +10587,38 @@ After reviewing the search results, include a confidence assessment on a line by
     aborted = isAbort;
     if (!isAbort) {
       engineFailed = true;
-      await emit2({ event: "error", data: { code: "engine_stopped", message: e.message } });
+      try {
+        await emit2({ event: "error", data: { code: "engine_stopped", message: e.message } });
+      } catch {
+      }
     }
   } finally {
     if (localAccounting) d.manager.generationEnd();
     inflight.delete(convId);
   }
+  if (pendingParse) {
+    for (const ev of flushState(pendingParse)) {
+      if (ev.type === "reasoning") {
+        fullReasoning += ev.text;
+      } else {
+        fullContent += ev.text;
+        liveOut++;
+      }
+    }
+    pendingParse = null;
+  }
   const totalMs = Date.now() - requestStart;
   const thinkMs = thinkStart && thinkEnd ? thinkEnd - thinkStart : 0;
   const ctxMax = upstream.ctxMax;
   const genTokensFallback = finalUsage.completion_tokens || liveOut;
+  const promptTokensFallback = finalUsage.prompt_tokens || lastPromptTotal;
   const stats = {
     ttftMs,
     totalMs,
     thinkMs,
     // Full context occupancy: cache-reused prompt tokens still sit in the KV cache / context
     // window (they're skipped only for recomputation), so they must stay counted here.
-    ctxUsed: (finalUsage.prompt_tokens ?? 0) + genTokensFallback,
+    ctxUsed: promptTokensFallback + genTokensFallback,
     ctxMax,
     model: upstream.modelName,
     aborted
@@ -10582,7 +10635,7 @@ After reviewing the search results, include a confidence assessment on a line by
     stats.tps = finalTimings.predicted_per_second;
     stats.cachedTokens = cachedExplicit ?? Math.max(0, (fullPrompt || processed) - processed);
   } else {
-    stats.promptTokens = fullPrompt;
+    stats.promptTokens = promptTokensFallback;
     stats.genTokens = genTokensFallback;
     stats.genMs = totalMs - ttftMs;
     stats.tps = stats.genMs > 0 ? Math.round(stats.genTokens / stats.genMs * 1e3 * 10) / 10 : 0;
@@ -10919,6 +10972,41 @@ Cannot execute bash commands.`);
   };
 }
 
+// src/code/code-upstream.ts
+var FALLBACK_CTX = 8192;
+function resolveCodeUpstream(d, requestedModel) {
+  const res = resolveChatUpstream(d, requestedModel);
+  if (!res.ok) throw new Error(res.code === "model_not_loaded" ? "model_not_loaded" : res.message);
+  const u = res.upstream;
+  const contextWindow = u.ctxMax > 0 ? u.ctxMax : FALLBACK_CTX;
+  if (u.remote) {
+    return {
+      modelId: u.modelField,
+      modelName: u.modelName || "Remote Model",
+      contextWindow,
+      remote: u.remote,
+      target: "",
+      provider: {
+        baseUrl: `${u.remote.baseUrl}/v1`,
+        apiKey: u.remote.token,
+        // The link token travels as X-TurboLLM-Auth (linkHeaders, link-proxy.ts) and NOT as
+        // a bearer: `Authorization` on the host is its own API-key credential, a different
+        // secret with different scope, and presenting a link token there would either be
+        // rejected or — worse — accepted as something it is not.
+        authHeader: false,
+        headers: { "X-TurboLLM-Auth": u.remote.token }
+      }
+    };
+  }
+  return {
+    modelId: u.modelField,
+    modelName: u.modelName || "Local Model",
+    contextWindow,
+    target: u.target,
+    provider: { baseUrl: `${u.target}/v1`, apiKey: "agent-key", authHeader: true }
+  };
+}
+
 // src/routines/routine-tools.ts
 var SCHEDULE_RULE_SCHEMA = {
   type: "object",
@@ -11553,14 +11641,12 @@ function pickPrefillProgress(slots) {
 }
 async function runCodeSession(params) {
   const { d, convId, sessionId, repoRoot, mode, thinkingBudget, reasoningEffort, task: rawTask, signal, sink } = params;
-  const ms = d.manager.status();
-  if (ms.state !== "running" || !ms.model) throw new Error("model_not_loaded");
-  const target = d.manager.target();
-  if (!target) throw new Error("model_not_loaded");
-  const engineKind = d.registry.active()?.kind ?? "";
-  const modelId = engineModelAlias(engineKind, d.manager.currentOpts()?.modelPath) ?? ms.model.key;
-  const contextWindow = ms.model.ctx > 0 ? ms.model.ctx : 8192;
-  const modelDisplayName = ms.model.name || "Local Model";
+  const requestedModel = params.model;
+  const upstream = resolveCodeUpstream(d, requestedModel);
+  const target = upstream.target;
+  const modelId = upstream.modelId;
+  const contextWindow = upstream.contextWindow;
+  const modelDisplayName = upstream.modelName;
   const skills = new SkillStore(d.store.dir()).list();
   const authStorage = AuthStorage.inMemory();
   const modelRegistry = ModelRegistry.inMemory(authStorage);
@@ -11584,9 +11670,7 @@ async function runCodeSession(params) {
   }
   const agentDir = join21(d.store.dir(), "pi-agent");
   modelRegistry.registerProvider("local", {
-    baseUrl: `${target}/v1`,
-    apiKey: "agent-key",
-    authHeader: true,
+    ...upstream.provider,
     api: "openai-completions",
     models: [
       {
@@ -11630,9 +11714,7 @@ async function runCodeSession(params) {
     const skillAuth = AuthStorage.inMemory();
     const skillRegistry = ModelRegistry.inMemory(skillAuth);
     skillRegistry.registerProvider("local", {
-      baseUrl: `${target}/v1`,
-      apiKey: "agent-key",
-      authHeader: true,
+      ...upstream.provider,
       api: "openai-completions",
       models: [{
         id: modelId,
@@ -11656,7 +11738,7 @@ async function runCodeSession(params) {
     const skillExtension = (pi) => {
       pi.on("before_provider_request", async (event) => {
         releaseSkillGate();
-        if (d.gate) skillHeldGate = await d.gate.acquire("bg", { signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
+        if (d.gate && !upstream.remote) skillHeldGate = await d.gate.acquire("bg", { signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
         const payload = { ...event.payload };
         delete payload.max_tokens;
         delete payload.max_completion_tokens;
@@ -11764,9 +11846,7 @@ async function runCodeSession(params) {
     const subAuth = AuthStorage.inMemory();
     const subRegistry = ModelRegistry.inMemory(subAuth);
     subRegistry.registerProvider("local", {
-      baseUrl: `${target}/v1`,
-      apiKey: "agent-key",
-      authHeader: true,
+      ...upstream.provider,
       api: "openai-completions",
       models: [{
         id: modelId,
@@ -11799,7 +11879,7 @@ async function runCodeSession(params) {
     const subExtension = (pi) => {
       pi.on("before_provider_request", async (event) => {
         releaseSubGate();
-        if (d.gate) subHeldGate = await d.gate.acquire("bg", { signal: subAc.signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
+        if (d.gate && !upstream.remote) subHeldGate = await d.gate.acquire("bg", { signal: subAc.signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
         const payload = { ...event.payload };
         delete payload.max_tokens;
         delete payload.max_completion_tokens;
@@ -11958,8 +12038,8 @@ async function runCodeSession(params) {
   const extension = (pi) => {
     pi.on("before_provider_request", async (event) => {
       releaseGate();
-      if (d.gate) heldGate = await d.gate.acquire("fg", { signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
-      startPrefillPoll();
+      if (d.gate && !upstream.remote) heldGate = await d.gate.acquire("fg", { signal, timeoutMs: GATE_QUEUE_TIMEOUT_MS });
+      if (!upstream.remote) startPrefillPoll();
       roundStartedAt = Date.now();
       roundFirstTokenAt = void 0;
       const payload = { ...event.payload };
@@ -12173,7 +12253,7 @@ ${lines.join("\n")}${more}]`
         const q = (params2.question ?? "").trim();
         if (!q) return { content: [{ type: "text", text: "lookback_history: `question` must be non-empty." }], details: {} };
         try {
-          const { answer } = await lookbackPreCompactionHistory({ d, convId, sessionId, repoRoot, question: q, signal });
+          const { answer } = await lookbackPreCompactionHistory({ d, convId, sessionId, repoRoot, question: q, signal, model: requestedModel });
           return { content: [{ type: "text", text: answer }], details: {} };
         } catch (e) {
           const message = e instanceof Error ? e.message : String(e);
@@ -12445,13 +12525,9 @@ ${lines.join("\n")}${more}]`
 async function compactCodeSession(params) {
   const { d, convId, sessionId, repoRoot, customInstructions } = params;
   if (d.db.getAgentRun(sessionId)?.clearedUpToMessageId) throw new Error("session_cleared");
-  const ms = d.manager.status();
-  if (ms.state !== "running" || !ms.model) throw new Error("model_not_loaded");
-  const target = d.manager.target();
-  if (!target) throw new Error("model_not_loaded");
-  const engineKind = d.registry.active()?.kind ?? "";
-  const modelId = engineModelAlias(engineKind, d.manager.currentOpts()?.modelPath) ?? ms.model.key;
-  const contextWindow = ms.model.ctx > 0 ? ms.model.ctx : 8192;
+  const upstream = resolveCodeUpstream(d, params.model);
+  const modelId = upstream.modelId;
+  const contextWindow = upstream.contextWindow;
   const { summaryText, messages: effectiveMessages } = resolveEffectiveHistory(d, convId, sessionId);
   if (effectiveMessages.length === 0) throw new Error("nothing_to_compact");
   const compactionSettings = compactionSettingsFor(contextWindow);
@@ -12459,13 +12535,11 @@ async function compactCodeSession(params) {
   const authStorage = AuthStorage.inMemory();
   const modelRegistry = ModelRegistry.inMemory(authStorage);
   modelRegistry.registerProvider("local", {
-    baseUrl: `${target}/v1`,
-    apiKey: "agent-key",
-    authHeader: true,
+    ...upstream.provider,
     api: "openai-completions",
     models: [{
       id: modelId,
-      name: ms.model.name || "Local Model",
+      name: upstream.modelName,
       reasoning: false,
       input: ["text"],
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
@@ -12548,13 +12622,9 @@ async function lookbackPreCompactionHistory(params) {
   const { d, convId, sessionId, repoRoot, question, signal } = params;
   const run = d.db.getAgentRun(sessionId);
   if (!run?.compactionUpToMessageId) throw new Error("nothing_compacted");
-  const ms = d.manager.status();
-  if (ms.state !== "running" || !ms.model) throw new Error("model_not_loaded");
-  const target = d.manager.target();
-  if (!target) throw new Error("model_not_loaded");
-  const engineKind = d.registry.active()?.kind ?? "";
-  const modelId = engineModelAlias(engineKind, d.manager.currentOpts()?.modelPath) ?? ms.model.key;
-  const contextWindow = ms.model.ctx > 0 ? ms.model.ctx : 8192;
+  const upstream = resolveCodeUpstream(d, params.model);
+  const modelId = upstream.modelId;
+  const contextWindow = upstream.contextWindow;
   const all = d.db.getConversation(convId, true)?.messages ?? [];
   const cutIdx = all.findIndex((m) => m.id === run.compactionUpToMessageId);
   const preCompaction = cutIdx === -1 ? all : all.slice(0, cutIdx + 1);
@@ -12562,13 +12632,11 @@ async function lookbackPreCompactionHistory(params) {
   const authStorage = AuthStorage.inMemory();
   const modelRegistry = ModelRegistry.inMemory(authStorage);
   modelRegistry.registerProvider("local", {
-    baseUrl: `${target}/v1`,
-    apiKey: "agent-key",
-    authHeader: true,
+    ...upstream.provider,
     api: "openai-completions",
     models: [{
       id: modelId,
-      name: ms.model.name || "Local Model",
+      name: upstream.modelName,
       reasoning: false,
       input: ["text"],
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
@@ -12730,7 +12798,7 @@ var CodeRunManager = class {
   enqueue(sessionId, params) {
     const s = this.ensure(sessionId, params.convId, params.repoRoot);
     const willQueue = s.active !== null;
-    s.queue.push({ task: params.task, userMsgId: params.userMsgId, thinkingBudget: params.thinkingBudget ?? -1, reasoningEffort: params.reasoningEffort, kind: params.kind ?? "followUp" });
+    s.queue.push({ task: params.task, userMsgId: params.userMsgId, thinkingBudget: params.thinkingBudget ?? -1, reasoningEffort: params.reasoningEffort, kind: params.kind ?? "followUp", model: params.model });
     this.emitQueue(sessionId);
     void this.pump(sessionId);
     return { queued: willQueue };
@@ -12850,8 +12918,7 @@ var CodeRunManager = class {
       push(ev.event, ev.data);
     };
     try {
-      const ms = this.d.manager.status();
-      const model = ms.model;
+      const statsModelKey = turn.model ?? this.d.manager.status().model?.key;
       const mode = this.d.db.getConversation(s.convId)?.agentMode ?? "auto";
       const result = await this.runner({
         d: this.d,
@@ -12862,6 +12929,7 @@ var CodeRunManager = class {
         thinkingBudget: turn.thinkingBudget,
         reasoningEffort: turn.reasoningEffort,
         task: turn.task,
+        model: turn.model,
         signal: ac.signal,
         sink,
         // Publish/withdraw the live turn's steer handle so steer() can inject into THIS turn.
@@ -12883,7 +12951,7 @@ var CodeRunManager = class {
         stats: {
           ctxUsed: ctxStats.ctxUsed,
           ctxMax: ctxStats.ctxMax,
-          model: model?.key,
+          model: statsModelKey,
           aborted: result.aborted,
           // Real per-turn token/timing stats (foldTurnUsage, code-session.ts) — omitted by
           // runCodeSession (not zeroed) when the engine returned no usable usage at all.
@@ -17479,7 +17547,8 @@ function spillMb(baselineSharedMb, loadedSharedMb) {
 }
 var SPILL_PROXIMITY_MB = 1024;
 var SPILL_FLOOR_CAP_MB = 2048;
-function isSpilling(spill, vramAbsMb, budgetMb, floorMb = null, pooled = false) {
+function isSpilling(spill, vramAbsMb, budgetMb, floorMb = null, pooled = false, unified = false) {
+  if (unified) return false;
   if (spill === null) return false;
   const floor = pooled ? 0 : Math.min(floorMb ?? 0, SPILL_FLOOR_CAP_MB);
   if (spill - floor <= SPILL_TOLERANCE_MB) return false;
@@ -18492,7 +18561,6 @@ var PER_TEST_TIMEOUT_MS = 10 * 6e4;
 var PREFILL_GRACE_MS = 8e3;
 var TOTAL_BUDGET_MS = 45 * 6e4;
 var OOM_RE = /out of memory|cudaMalloc|failed to allocate|unable to allocate|device not ready|CUDA error/i;
-var QUALITY_KV = ["f16", "q8_0", "turbo4"];
 var KV_BYTES = {
   f32: 4,
   f16: 2,
@@ -18757,6 +18825,7 @@ var BenchRunner = class {
     const budgetMb = gpuBudgetMb(sys, base2);
     let spillFloorMb = null;
     const pooled = sys.gpus.length > 1 && base2.gpu.splitMode !== "none";
+    const unified = unifiedMemoryOnly(sys);
     if (sys.gpus.length > 0) {
       const hi0 = entry.blockCount > 0 ? entry.blockCount : 99;
       let lo = 0, hi = hi0;
@@ -18768,7 +18837,7 @@ var BenchRunner = class {
         this.pushProbe(results, base2, "ngl", mid, probe2);
         await this.settleGpu(sys);
         spillFloorMb = spillFloor(spillFloorMb, probe2.spillMb, probe2.vramAbsMb, budgetMb, pooled);
-        const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled);
+        const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled, unified);
         if (verdict.decision === "fits") {
           bestNgl = mid;
           lo = mid + 1;
@@ -18782,7 +18851,7 @@ var BenchRunner = class {
     }
     if (bestNgl === null) return null;
     let found = await this.benchAt(entry, sys, { ...base2, ngl: bestNgl }, caps, results, `ngl=${bestNgl}`);
-    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled) && bestNgl > 0) {
+    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled, unified) && bestNgl > 0) {
       this.emit(`ngl=${bestNgl} spilled ${Math.round(found.cand.spillMb)} MB once actually generating (the load-only probe could not see this) \u2014 retrying at ngl=${bestNgl - 1}`);
       const safer = await this.benchAt(entry, sys, { ...base2, ngl: bestNgl - 1 }, caps, results, `ngl=${bestNgl - 1} (spill backoff)`);
       if (safer) {
@@ -18819,6 +18888,7 @@ var BenchRunner = class {
     let bestN = null;
     let spillFloorMb = null;
     const pooled = sys.gpus.length > 1 && base2.gpu.splitMode !== "none";
+    const unified = unifiedMemoryOnly(sys);
     while (lo <= hi && !this.cancelled && Date.now() <= this.deadline) {
       const mid = Math.floor((lo + hi) / 2);
       this.state = { ...this.state, step: `KV ${base2.kvTypeK}: probing nCpuMoe=${mid} (range ${lo}\u2013${hi})\u2026`, candidates: results };
@@ -18826,7 +18896,7 @@ var BenchRunner = class {
       this.pushProbe(results, base2, "nCpuMoe", mid, probe2);
       await this.settleGpu(sys);
       spillFloorMb = spillFloor(spillFloorMb, probe2.spillMb, probe2.vramAbsMb, budgetMb, pooled);
-      const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled);
+      const verdict = probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb, pooled, unified);
       if (verdict.decision === "fits") {
         bestN = mid;
         hi = mid - 1;
@@ -18854,7 +18924,7 @@ var BenchRunner = class {
       }
       if (n < maxN) this.emit(`nCpuMoe=${n} failed to bench (${results[results.length - 1]?.outcome ?? "crash"}) \u2014 retrying at nCpuMoe=${n + 1}`);
     }
-    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled) && foundN < maxN) {
+    if (found && isSpilling(found.cand.spillMb, found.cand.vramAbsMb, budgetMb, spillFloorMb, pooled, unified) && foundN < maxN) {
       this.emit(`nCpuMoe=${foundN} spilled ${Math.round(found.cand.spillMb)} MB once actually generating (the load-only probe could not see this) \u2014 retrying at nCpuMoe=${foundN + 1}`);
       const safer = await this.benchAt(entry, sys, { ...base2, nCpuMoe: foundN + 1 }, caps, results, `nCpuMoe=${foundN + 1} (spill backoff)`);
       if (safer) {
@@ -19487,17 +19557,20 @@ function overHeadroom(vramAbsMb, budgetMb, headroomMb) {
   if (!vramAbsMb || budgetMb <= 0) return false;
   return vramAbsMb > budgetMb - headroomMb;
 }
+var MIN_SINGLE_GPU_FRACTION_MOE = 0.5;
+var MIN_SINGLE_GPU_FRACTION_DENSE = 0.6;
 function pickSplitStrategies(entry, sys, base2, caps) {
   const has = (flag) => caps.flags.length === 0 || caps.flags.includes(flag);
   if (sys.gpus.length <= 1 || !has("--split-mode") || !has("--main-gpu")) return [base2.gpu];
   const strategies = [];
-  const kvOpts = pickKvQuants(base2.kvTypeK, caps.kvTypes);
-  const bestFitKv = kvOpts.reduce((a, b) => kvBytes(b) < kvBytes(a) ? b : a, kvOpts[0] ?? base2.kvTypeK);
   const mainGpu = base2.gpu.mainGpu >= 0 ? base2.gpu.mainGpu : 0;
   const single = { ...base2.gpu, splitMode: "none", mainGpu, tensorSplit: [] };
-  const singleFrac = maxGpuFraction(entry, sys, { ...base2, gpu: single, kvTypeK: bestFitKv, kvTypeV: bestFitKv });
-  if (singleFrac >= 0.5) strategies.push(single);
+  const singleFrac = maxGpuFraction(entry, sys, { ...base2, gpu: single });
+  if (singleFrac >= (entry.moe ? MIN_SINGLE_GPU_FRACTION_MOE : MIN_SINGLE_GPU_FRACTION_DENSE)) {
+    strategies.push(single);
+  }
   strategies.push(base2.gpu);
+  if (base2.gpu.splitMode !== "none") strategies.push({ ...base2.gpu, splitMode: "row", tensorSplit: [] });
   const seen = /* @__PURE__ */ new Set();
   return strategies.filter((g) => {
     const key = `${g.splitMode}|${g.mainGpu >= 0 ? g.mainGpu : 0}|${g.tensorSplit.join(",")}`;
@@ -19550,15 +19623,9 @@ function splitLabel(g, gpuCount) {
   if (g.splitMode === "row") return `row-split across ${gpuCount} GPUs`;
   return `layer-split across ${gpuCount} GPUs`;
 }
-function pickKvQuants(baseKv, kvTypes) {
-  const supported = (t) => kvTypes.length === 0 ? t === baseKv : kvTypes.includes(t);
-  const out = [];
-  for (const t of [baseKv, ...QUALITY_KV]) if (supported(t) && !out.includes(t)) out.push(t);
-  return out;
-}
-function probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb = null, pooled = false) {
+function probeVerdict(probe2, budgetMb, headroomMb, spillFloorMb = null, pooled = false, unified = false) {
   if (probe2.outcome === "oom") return { decision: "offload-more", reason: "oom" };
-  if (isSpilling(probe2.spillMb, probe2.vramAbsMb, budgetMb, spillFloorMb, pooled)) {
+  if (isSpilling(probe2.spillMb, probe2.vramAbsMb, budgetMb, spillFloorMb, pooled, unified)) {
     return { decision: "offload-more", reason: "spill", shortfallMb: probe2.spillMb };
   }
   if (overHeadroom(probe2.vramAbsMb, budgetMb, headroomMb)) {
@@ -22907,13 +22974,15 @@ function detectHardware(info = getSysInfo()) {
     void 0
   );
   const vramMb = ofVendor.reduce((sum, g) => sum + g.vramMb, 0);
+  const unifiedMemory = ofVendor.length > 0 && ofVendor.every((g) => g.unified === true);
   return {
     platform: process.platform,
     arch: arch2,
     gpuVendor,
     hasGpu,
     vramMb,
-    gpuName: headline?.name
+    gpuName: headline?.name,
+    unifiedMemory
   };
 }
 
@@ -23316,10 +23385,15 @@ function roleFor(profile) {
 }
 
 // src/onboarding/recommend.ts
+function jointMemoryOk(e, hw) {
+  if (!hw.unifiedMemory) return true;
+  return (e.minVramMb ?? 0) + (e.minSystemRamMb ?? 0) <= hw.systemRamMb;
+}
 function fits(e, hw) {
   if (e.minVramMb !== null && hw.usableVramMb < e.minVramMb) return false;
   if (e.maxVramMb !== null && hw.usableVramMb >= e.maxVramMb) return false;
   if (e.minSystemRamMb !== null && hw.systemRamMb < e.minSystemRamMb) return false;
+  if (!jointMemoryOk(e, hw)) return false;
   return true;
 }
 function recommend(profile, hw) {
@@ -23328,7 +23402,7 @@ function recommend(profile, hw) {
   let candidates = BLESSED.filter((e) => e.role === role && fits(e, hw));
   if (candidates.length === 0) {
     candidates = BLESSED.filter(
-      (e) => e.role === role && (e.minVramMb === null || hw.usableVramMb >= e.minVramMb) && (e.minSystemRamMb === null || hw.systemRamMb >= e.minSystemRamMb)
+      (e) => e.role === role && (e.minVramMb === null || hw.usableVramMb >= e.minVramMb) && (e.minSystemRamMb === null || hw.systemRamMb >= e.minSystemRamMb) && jointMemoryOk(e, hw)
     );
   }
   if (candidates.length === 0) return { kind: "hf-search", reason: "no-fit" };
@@ -23345,11 +23419,9 @@ function recommend(profile, hw) {
 function isT0(hw) {
   return hw.usableVramMb < 4096;
 }
-function hardwareFactsFromSysInfo() {
-  const info = getSysInfo();
-  const primaryVramMb = info.gpus[0]?.vramMb ?? 0;
-  const unifiedMemory = info.gpus.some((g) => g.unified === true);
-  return { usableVramMb: primaryVramMb, systemRamMb: info.ramMB, unifiedMemory };
+function hardwareFactsFromSysInfo(info = getSysInfo()) {
+  const { vramMb, unifiedMemory } = detectHardware(info);
+  return { usableVramMb: vramMb, systemRamMb: info.ramMB, unifiedMemory };
 }
 function applyOnboardingPatch(current, patch, now) {
   const next = { ...current };
@@ -23587,7 +23659,7 @@ function registerApi(app2, d) {
     const ms = d.manager.status();
     const launchCommand = d.manager.launchCommand();
     const engine = { ...core.engine };
-    if (ms.err) engine.error = ms.err;
+    if (ms.err) engine.error = { ...ms.err, failReason: classifyLoadFailure(ms.err) };
     if (launchCommand) engine.launchCommand = launchCommand;
     return c.json({
       version: d.version,
@@ -23596,6 +23668,11 @@ function registerApi(app2, d) {
       // Last model the user loaded (config-tracked) — lets `turbollm launch` auto-load the
       // true last-used model instead of guessing the first library entry (F-034).
       lastLoaded: d.store.snapshot().lastLoaded,
+      // Turbo Link (ADR-382): the linked-machine model this install is pointed at, or '' for
+      // local. On /status because every surface reads it here — the web app hydrates its picker
+      // from it, and `turbollm launch <cli>` (which has no access to browser state) uses it
+      // instead of auto-loading a local model the user did not ask for.
+      selectedRemoteModel: d.store.snapshot().selectedRemoteModel ?? "",
       engineStats: core.engineStats,
       liveGeneration: core.liveGeneration,
       // Auto-tune runner state (spec 09 §1): real progress while a sweep runs, then
@@ -26258,6 +26335,29 @@ function registerLinkAdminRoutes(app2, d, opts) {
     if (denied) return denied;
     return c.json({ models: d.remoteCatalog?.models() ?? [] });
   });
+  app2.put("/api/v1/links/selected-model", async (c) => {
+    const denied = gate(c, MANAGE);
+    if (denied) return denied;
+    const body10 = await c.req.json().catch(() => null);
+    const wanted = (body10?.model ?? "").trim();
+    if (!wanted) {
+      d.store.update((cfg2) => {
+        cfg2.selectedRemoteModel = "";
+      });
+      return c.json({ ok: true, selectedRemoteModel: "" });
+    }
+    const route = d.modelRouter?.resolveRemoteTarget?.(wanted);
+    if (!route) {
+      return c.json({ error: { code: "not_remote", message: `'${wanted}' does not name a model on a linked machine.` } }, 400);
+    }
+    if ("status" in route) {
+      return c.json({ error: { code: "remote_unavailable", message: route.message } }, 503);
+    }
+    d.store.update((cfg2) => {
+      cfg2.selectedRemoteModel = wanted;
+    });
+    return c.json({ ok: true, selectedRemoteModel: wanted });
+  });
   app2.post("/api/v1/links", async (c) => {
     const denied = gate(c, MANAGE);
     if (denied) return denied;
@@ -27572,16 +27672,22 @@ function registerCodeRoutes(app2, d, codeRuns) {
     if (!run.repoRoot) return err9(c, 400, "invalid_input", "Session has no repo root.");
     if (runs.isActive(id)) return err9(c, 409, "run_active", "Stop or wait for the current run before compacting.");
     if (run.clearedUpToMessageId) return err9(c, 409, "session_cleared", "Resume this session before compacting \u2014 compacting a cleared session would lose the hidden history.");
-    const ms = d.manager.status();
-    if (ms.state !== "running" || !ms.model) return err9(c, 409, "model_not_loaded", "Load a model first.");
     const b = await body8(c);
+    const requestedModel = (b.model ?? "").trim() || void 0;
+    const remoteRoute = requestedModel ? d.modelRouter.resolveRemoteTarget(requestedModel) : void 0;
+    if (remoteRoute && "status" in remoteRoute) return err9(c, 503, "remote_unavailable", remoteRoute.message);
+    if (!remoteRoute) {
+      const ms = d.manager.status();
+      if (ms.state !== "running" || !ms.model) return err9(c, 409, "model_not_loaded", "Load a model first.");
+    }
     try {
       const result = await compactCodeSession({
         d,
         convId: run.convId,
         sessionId: id,
         repoRoot: agentCwd(run),
-        customInstructions: b.instructions?.trim() || void 0
+        customInstructions: b.instructions?.trim() || void 0,
+        model: requestedModel
       });
       return c.json({ ok: true, ...result });
     } catch (e) {
@@ -27758,9 +27864,14 @@ function registerCodeRoutes(app2, d, codeRuns) {
     const conv = db2.getConversation(run.convId, true);
     if (!conv) return err9(c, 404, "not_found", "Session conversation not found.");
     if (!run.repoRoot) return err9(c, 400, "invalid_input", "Session has no repo root.");
-    const ms = d.manager.status();
-    if (ms.state !== "running" || !ms.model) return err9(c, 409, "model_not_loaded", "Load a model first.");
-    if (!d.manager.target()) return err9(c, 409, "model_not_loaded", "Engine not running.");
+    const requestedModel = (b.model ?? "").trim() || void 0;
+    const remoteRoute = requestedModel ? d.modelRouter.resolveRemoteTarget(requestedModel) : void 0;
+    if (remoteRoute && "status" in remoteRoute) return err9(c, 503, "remote_unavailable", remoteRoute.message);
+    if (!remoteRoute) {
+      const ms = d.manager.status();
+      if (ms.state !== "running" || !ms.model) return err9(c, 409, "model_not_loaded", "Load a model first.");
+      if (!d.manager.target()) return err9(c, 409, "model_not_loaded", "Engine not running.");
+    }
     const newContent = (b.content ?? "").trim();
     let userMsgId;
     let task;
@@ -27773,7 +27884,7 @@ function registerCodeRoutes(app2, d, codeRuns) {
       userMsgId = lastUser.id;
       task = contextFilesBlock(lastUser.textAttachments) + lastUser.content;
     }
-    const enqueueParams = { convId: run.convId, repoRoot: agentCwd(run), task, userMsgId, thinkingBudget: b.thinkingBudget, reasoningEffort: parseReasoningEffort(b.reasoningEffort) };
+    const enqueueParams = { convId: run.convId, repoRoot: agentCwd(run), task, userMsgId, thinkingBudget: b.thinkingBudget, reasoningEffort: parseReasoningEffort(b.reasoningEffort), model: requestedModel };
     if (kind === "steer") {
       const { steered, queued: queued2 } = await runs.steer(id, enqueueParams);
       return c.json({ ok: true, queued: queued2, steered, userMessageId: userMsgId }, 202);