bitpull 0.1.13 → 0.1.14

package/bitpull.cjs

@@ -17869,14 +17869,36 @@ var LemonadeServer = class {
       signal
     });
   }
+  /**
+   * The backend to use for a recipe on this computer (graphics card or processor), from Lemonade's
+   * own report (/v1/system-info): its preferred one, otherwise the best one that can run here.
+   */
+  async chooseBackend(recipe) {
+    let info = {};
+    try {
+      const res = await fetch(`${this.url}/v1/system-info`, { headers: this.auth(), signal: AbortSignal.timeout(2e4) });
+      if (res.ok) info = await res.json();
+    } catch {
+    }
+    const r = info.recipes?.[recipe];
+    const backends = r?.backends ?? {};
+    const usable = (name2) => backends[name2] && backends[name2].state !== "unsupported";
+    const order = ["metal", "cuda", "rocm", "vulkan", "cpu"];
+    const name = (r?.default_backend && usable(r.default_backend) ? r.default_backend : void 0) ?? order.find(usable) ?? Object.keys(backends).find(usable) ?? "default";
+    if (Object.keys(backends).length && !usable(name)) throw new Error(`no ${recipe} engine can run on this computer`);
+    return { name, install: backends[name]?.state !== "installed" };
+  }
   /** Installs the backend a catalogue model needs, then downloads the model, with progress. */
   async pull(model, onProgress) {
     const entry = catalogModel(model);
     if (!entry || entry.runtime !== "lemonade") throw new Error(`${model} is not a Lemonade model`);
     if (!await this.start()) throw new Error(this.lastError ?? "Lemonade is not installed");
-    onProgress({ detail: `installing ${entry.recipe} engine` });
-    const inst = await this.post("/v1/install", { recipe: entry.recipe });
-    if (!inst.ok) throw new Error(`engine install failed: ${await errorText(inst)}`);
+    const backend = await this.chooseBackend(entry.recipe);
+    if (backend.install) {
+      onProgress({ detail: `installing ${entry.recipe} engine (${backend.name})` });
+      const inst = await this.post("/v1/install", { recipe: entry.recipe, backend: backend.name });
+      if (!inst.ok) throw new Error(`engine install failed (${entry.recipe}, ${backend.name}): ${await errorText(inst)}`);
+    }
     onProgress({ progress: 0, detail: "downloading model" });
     const res = await this.post("/v1/pull", { model, stream: true });
     if (!res.ok || !res.body) throw new Error(`model download failed: ${await errorText(res)}`);
@@ -17954,19 +17976,38 @@ async function runAiTask(spec, timeoutSec, ollamaUrl, signal) {
         const d2 = await ollama(ollamaUrl, "/api/embed", { model: spec.model, input: (spec.input ?? []).slice(0, 256) }, sig);
         return { ok: true, result: { embeddings: d2.embeddings }, usage: { seconds: seconds(), promptTokens: d2.prompt_eval_count ?? 0 } };
       }
-      const d = await ollama(
-        ollamaUrl,
-        "/api/chat",
-        {
-          model: spec.model,
-          messages: spec.messages ?? [],
-          stream: false,
-          options: { num_predict: spec.maxTokens ?? 1024, ...spec.temperature !== void 0 ? { temperature: spec.temperature } : {} }
-        },
-        sig
-      );
+      const body = {
+        model: spec.model,
+        messages: spec.messages ?? [],
+        stream: false,
+        // Reasoning models (Qwen 3.5, gpt-oss…) can spend every token "thinking" and return an empty
+        // answer: answer directly unless the requester asks for reasoning.
+        think: spec.think ?? false,
+        options: { num_predict: spec.maxTokens ?? 1024, ...spec.temperature !== void 0 ? { temperature: spec.temperature } : {} }
+      };
+      let d;
+      try {
+        d = await ollama(ollamaUrl, "/api/chat", body, sig);
+      } catch (err) {
+        if (!/think/i.test(err.message)) throw err;
+        const { think: _t, ...plain } = body;
+        d = await ollama(ollamaUrl, "/api/chat", plain, sig);
+      }
       const usage = { seconds: seconds(), promptTokens: d.prompt_eval_count ?? 0, completionTokens: d.eval_count ?? 0 };
-      return { ok: true, result: { content: d.message?.content ?? "", usage: { promptTokens: usage.promptTokens, completionTokens: usage.completionTokens } }, usage };
+      const content = d.message?.content ?? "";
+      const reasoning = d.message?.thinking ?? "";
+      return {
+        ok: true,
+        result: {
+          // Out of tokens while reasoning: give the reasoning rather than nothing, and say so.
+          content: content || (reasoning ? `[The model used all ${usage.completionTokens} tokens reasoning before answering; raise maxTokens. Its reasoning so far:]
+${reasoning}` : ""),
+          ...reasoning && spec.think ? { reasoning } : {},
+          truncated: d.done_reason === "length",
+          usage: { promptTokens: usage.promptTokens, completionTokens: usage.completionTokens }
+        },
+        usage
+      };
     }
     const declared = spec.units ?? 0;
     if (task === "image") {
@@ -18361,7 +18402,7 @@ ${texts.join("\n")}` : input.prompt ?? "Describe this image.";
   const { images: _i, audio: _a, ...rest } = result;
   return { model, task, cost: done.cost, jobId: done.id, ...files.length ? { files } : {}, ...rest };
 }
-async function askOpenModel(ctx, input, waitMs = 18e4) {
+async function askOpenModel(ctx, input, waitMs = 6e5) {
   const model = await chooseModel(ctx, input.model, input.tier);
   const promptChars = input.messages.reduce((n, m) => n + m.content.length, 0);
   const job = await submitEncrypted(
@@ -18371,6 +18412,9 @@ async function askOpenModel(ctx, input, waitMs = 18e4) {
     { messages: input.messages, ...input.temperature !== void 0 ? { temperature: input.temperature } : {} }
   );
   const done = await waitForJob(ctx, job.id, waitMs, 500);
+  if (done.status === "running" || done.status === "queued") {
+    throw new ToolError(`The model is still working on a lender's computer (job ${done.id}). It is not lost: check later with job_status, or "bitpull ask" again.`);
+  }
   if (done.status !== "done") throw new ToolError(`Model job ${done.status}${done.error ? `: ${done.error}` : ""} (id ${done.id})`);
   const result = done.result;
   return { model, content: result.content, cost: done.cost, jobId: done.id, usage: result.usage };
@@ -19255,7 +19299,7 @@ async function doctor(cfg) {
 }
 
 // apps/agent/src/control.ts
-var APP_VERSION = true ? "0.1.13" : "dev";
+var APP_VERSION = true ? "0.1.14" : "dev";
 var UPDATE_MANIFEST = "https://github.com/SamuealPK/bitpull-releases/releases/latest/download/latest.json";
 var inDesktopApp = () => /bitpull-agent(\.exe)?$/i.test(process.execPath);
 function newerVersion(a, b) {
@@ -30675,7 +30719,7 @@ function openBrowser(url) {
   const args = (0, import_node_os10.platform)() === "win32" ? ["/c", "start", "", url] : [url];
   (0, import_node_child_process7.spawn)(cmd, args, { stdio: "ignore", detached: true }).unref();
 }
-var agentVersion = () => true ? "0.1.13" : "dev";
+var agentVersion = () => true ? "0.1.14" : "dev";
 var dashboardUrl = (cfg) => `http://127.0.0.1:${cfg.controlPort}/#token=${cfg.controlToken}`;
 var AppNotRunning = class extends Error {
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bitpull",
-  "version": "0.1.13",
+  "version": "0.1.14",
   "description": "BitPull: share signed, AI-readable code and computing power with AI tools through MCP. CLI and MCP server.",
   "mcpName": "io.github.SamuealPK/bitpull",
   "bin": {