open-agents-ai 0.185.34 → 0.185.36

package/dist/index.js

@@ -26404,94 +26404,103 @@ If you're stuck, try a completely different approach. Do NOT repeat what failed
               this.emit({ type: "error", content: `Backend error: ${reqErr.message}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
               break;
             }
-            const recovered = await this.retryOnTransient(reqErr, chatRequest, turn);
-            if (!recovered) {
-              const errMsg = reqErr instanceof Error ? reqErr.message : String(reqErr);
-              const cause = reqErr instanceof Error && reqErr.cause ? ` (${reqErr.cause.message ?? ""} ${reqErr.cause?.code ?? ""})` : "";
-              this.emit({ type: "error", content: `Backend error: ${errMsg}${cause}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
-              if (/HTTP 404|not found|model.*not found/i.test(errMsg)) {
-                this.emit({ type: "error", content: `Model not available. Use /model to select a different model.`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+            if (this.handleMaxTokensError(reqErr, chatRequest)) {
+              try {
+                response = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
+              } catch (retryErr) {
+                this.emit({ type: "error", content: `Retry with reduced max_tokens also failed: ${retryErr instanceof Error ? retryErr.message : String(retryErr)}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
                 break;
               }
-              let imageRecovered = false;
-              if (/invalid image|image.*invalid|image_url.*unsupported|does not support.*image|image.*not supported/i.test(errMsg)) {
-                imageRecovered = await this._recoverFromImageError(messages, chatRequest, turn);
-                if (imageRecovered) {
-                  try {
-                    const imgRetry = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
-                    response = imgRetry;
-                  } catch (imgRetryErr) {
-                    const msg2 = imgRetryErr instanceof Error ? imgRetryErr.message : String(imgRetryErr);
-                    this.emit({ type: "error", content: `Retry after image fallback also failed: ${msg2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
-                    imageRecovered = false;
-                    break;
-                  }
-                } else {
+            } else {
+              const recovered = await this.retryOnTransient(reqErr, chatRequest, turn);
+              if (!recovered) {
+                const errMsg = reqErr instanceof Error ? reqErr.message : String(reqErr);
+                const cause = reqErr instanceof Error && reqErr.cause ? ` (${reqErr.cause.message ?? ""} ${reqErr.cause?.code ?? ""})` : "";
+                this.emit({ type: "error", content: `Backend error: ${errMsg}${cause}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                if (/HTTP 404|not found|model.*not found/i.test(errMsg)) {
+                  this.emit({ type: "error", content: `Model not available. Use /model to select a different model.`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
                   break;
                 }
-              }
-              if (imageRecovered) {
-              } else if (/does not support tools|HTTP 400.*tools/i.test(errMsg)) {
-                this.emit({
-                  type: "status",
-                  content: `Model lacks native tool support \u2014 switching to prompt-injected tool mode`,
-                  timestamp: (/* @__PURE__ */ new Date()).toISOString()
-                });
-                const toolDescriptions = Array.from(this.tools.values()).map((t) => `- ${t.name}: ${t.description}`).join("\n");
-                const toolInjectMsg = [
-                  "\n\n[TOOL MODE \u2014 PROMPT INJECTION]",
-                  "This model does not have native tool-calling. To use tools, output a JSON block:",
-                  "```json",
-                  '{"tool": "tool_name", "args": {"param": "value"}}',
-                  "```",
-                  "\nAvailable tools:",
-                  toolDescriptions,
-                  "\nOutput EXACTLY ONE tool call per response in the JSON format above.",
-                  "After seeing the tool result, continue or call another tool.",
-                  'When done, output: {"tool": "task_complete", "args": {"summary": "what you did"}}'
-                ].join("\n");
-                messages.push({ role: "system", content: toolInjectMsg });
-                chatRequest.tools = [];
-                try {
-                  response = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
-                  const content = response.choices?.[0]?.message?.content ?? "";
-                  const jsonMatch = content.match(/```json\s*\n?([\s\S]*?)```/);
-                  if (jsonMatch) {
+                let imageRecovered = false;
+                if (/invalid image|image.*invalid|image_url.*unsupported|does not support.*image|image.*not supported/i.test(errMsg)) {
+                  imageRecovered = await this._recoverFromImageError(messages, chatRequest, turn);
+                  if (imageRecovered) {
                     try {
-                      const parsed = JSON.parse(jsonMatch[1]);
-                      if (parsed.tool && this.tools.has(parsed.tool)) {
-                        const tool = this.tools.get(parsed.tool);
-                        const result = await tool.execute(parsed.args ?? {});
-                        messages.push({ role: "assistant", content });
-                        messages.push({ role: "user", content: `Tool result (${parsed.tool}): ${result.output.slice(0, 2e3)}` });
-                        if (parsed.tool === "task_complete") {
-                          completed = true;
-                          summary = String(parsed.args?.summary ?? content);
+                      const imgRetry = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
+                      response = imgRetry;
+                    } catch (imgRetryErr) {
+                      const msg2 = imgRetryErr instanceof Error ? imgRetryErr.message : String(imgRetryErr);
+                      this.emit({ type: "error", content: `Retry after image fallback also failed: ${msg2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                      imageRecovered = false;
+                      break;
+                    }
+                  } else {
+                    break;
+                  }
+                }
+                if (imageRecovered) {
+                } else if (/does not support tools|HTTP 400.*tools/i.test(errMsg)) {
+                  this.emit({
+                    type: "status",
+                    content: `Model lacks native tool support \u2014 switching to prompt-injected tool mode`,
+                    timestamp: (/* @__PURE__ */ new Date()).toISOString()
+                  });
+                  const toolDescriptions = Array.from(this.tools.values()).map((t) => `- ${t.name}: ${t.description}`).join("\n");
+                  const toolInjectMsg = [
+                    "\n\n[TOOL MODE \u2014 PROMPT INJECTION]",
+                    "This model does not have native tool-calling. To use tools, output a JSON block:",
+                    "```json",
+                    '{"tool": "tool_name", "args": {"param": "value"}}',
+                    "```",
+                    "\nAvailable tools:",
+                    toolDescriptions,
+                    "\nOutput EXACTLY ONE tool call per response in the JSON format above.",
+                    "After seeing the tool result, continue or call another tool.",
+                    'When done, output: {"tool": "task_complete", "args": {"summary": "what you did"}}'
+                  ].join("\n");
+                  messages.push({ role: "system", content: toolInjectMsg });
+                  chatRequest.tools = [];
+                  try {
+                    response = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
+                    const content = response.choices?.[0]?.message?.content ?? "";
+                    const jsonMatch = content.match(/```json\s*\n?([\s\S]*?)```/);
+                    if (jsonMatch) {
+                      try {
+                        const parsed = JSON.parse(jsonMatch[1]);
+                        if (parsed.tool && this.tools.has(parsed.tool)) {
+                          const tool = this.tools.get(parsed.tool);
+                          const result = await tool.execute(parsed.args ?? {});
+                          messages.push({ role: "assistant", content });
+                          messages.push({ role: "user", content: `Tool result (${parsed.tool}): ${result.output.slice(0, 2e3)}` });
+                          if (parsed.tool === "task_complete") {
+                            completed = true;
+                            summary = String(parsed.args?.summary ?? content);
+                          }
+                          toolCallCount++;
+                          continue;
                         }
-                        toolCallCount++;
-                        continue;
+                      } catch {
                       }
-                    } catch {
                     }
+                    messages.push({ role: "assistant", content });
+                    continue;
+                  } catch (retryErr2) {
+                    const msg2 = retryErr2 instanceof Error ? retryErr2.message : String(retryErr2);
+                    this.emit({ type: "error", content: `Prompt-injected tool mode also failed: ${msg2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                    break;
                   }
-                  messages.push({ role: "assistant", content });
-                  continue;
-                } catch (retryErr2) {
-                  const msg2 = retryErr2 instanceof Error ? retryErr2.message : String(retryErr2);
-                  this.emit({ type: "error", content: `Prompt-injected tool mode also failed: ${msg2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                }
+                if (!imageRecovered) {
+                  this.emit({
+                    type: "error",
+                    content: `Backend unavailable \u2014 stopping task. Fix the issue and retry.`,
+                    timestamp: (/* @__PURE__ */ new Date()).toISOString()
+                  });
                   break;
                 }
               }
-              if (!imageRecovered) {
-                this.emit({
-                  type: "error",
-                  content: `Backend unavailable \u2014 stopping task. Fix the issue and retry.`,
-                  timestamp: (/* @__PURE__ */ new Date()).toISOString()
-                });
-                break;
-              }
+              response = recovered ?? response;
             }
-            response = recovered ?? response;
           }
           totalTokens += response.usage?.totalTokens ?? 0;
           promptTokens += response.usage?.promptTokens ?? 0;
@@ -27056,15 +27065,24 @@ Integrate this guidance into your current approach. Continue working on the task
                 this.emit({ type: "error", content: "Task aborted by user", timestamp: (/* @__PURE__ */ new Date()).toISOString() });
                 break;
               }
-              const recovered = await this.retryOnTransient(reqErr, chatRequest, turn);
-              if (!recovered) {
-                const errMsg2 = reqErr instanceof Error ? reqErr.message : String(reqErr);
-                const cause2 = reqErr instanceof Error && reqErr.cause ? ` (${reqErr.cause.message ?? ""} ${reqErr.cause?.code ?? ""})` : "";
-                this.emit({ type: "error", content: `Backend error: ${errMsg2}${cause2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
-                this.emit({ type: "error", content: `Backend unavailable \u2014 stopping task.`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
-                break;
+              if (this.handleMaxTokensError(reqErr, chatRequest)) {
+                try {
+                  response = this.options.streamEnabled && this.hasStreamingSupport() ? await this.streamingRequest(chatRequest, turn) : await this.backend.chatCompletion(chatRequest);
+                } catch (retryErr) {
+                  this.emit({ type: "error", content: `Retry with reduced max_tokens also failed: ${retryErr instanceof Error ? retryErr.message : String(retryErr)}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                  break;
+                }
+              } else {
+                const recovered = await this.retryOnTransient(reqErr, chatRequest, turn);
+                if (!recovered) {
+                  const errMsg2 = reqErr instanceof Error ? reqErr.message : String(reqErr);
+                  const cause2 = reqErr instanceof Error && reqErr.cause ? ` (${reqErr.cause.message ?? ""} ${reqErr.cause?.code ?? ""})` : "";
+                  this.emit({ type: "error", content: `Backend error: ${errMsg2}${cause2}`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                  this.emit({ type: "error", content: `Backend unavailable \u2014 stopping task.`, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
+                  break;
+                }
+                response = recovered;
               }
-              response = recovered;
             }
             totalTokens += response.usage?.totalTokens ?? 0;
             promptTokens += response.usage?.promptTokens ?? 0;
@@ -28443,6 +28461,28 @@ ${transcript}`
       // -------------------------------------------------------------------------
       // Transient error recovery — retry on 502, fetch failed, timeouts
       // -------------------------------------------------------------------------
+      /**
+       * Detect max_completion_tokens rejection (HTTP 400) and auto-reduce to the server's limit.
+       * Returns true if maxTokens was reduced and the caller should retry.
+       */
+      handleMaxTokensError(err, chatRequest) {
+        const msg = err instanceof Error ? err.message : String(err);
+        const match = msg.match(/max_?(?:completion_?)?tokens\s+is\s+too\s+large.*?allows?\s+up\s+to\s+(\d+)/i);
+        if (!match)
+          return false;
+        const serverLimit = parseInt(match[1], 10);
+        if (isNaN(serverLimit) || serverLimit <= 0)
+          return false;
+        const prev = this.options.maxTokens;
+        this.options.maxTokens = serverLimit;
+        chatRequest.maxTokens = serverLimit;
+        this.emit({
+          type: "status",
+          content: `Server max_tokens limit is ${serverLimit} (was ${prev}) \u2014 auto-adjusted`,
+          timestamp: (/* @__PURE__ */ new Date()).toISOString()
+        });
+        return true;
+      }
       /** Detect whether an error is transient (worth retrying) */
       isTransientError(err) {
         if (err instanceof Error && err.fatal)
@@ -41301,11 +41341,11 @@ function execAsync(cmd, opts = {}) {
     child.stderr?.on("data", (d) => {
       stderr += d.toString();
     });
-    child.on("close", (code) => {
+    child.on("close", (code, signal) => {
       if (code === 0)
         resolve36(stdout.trim());
       else
-        reject(new Error(`Exit ${code}: ${stderr.slice(0, 500)}`));
+        reject(new Error(`Exit ${code}${signal ? ` (signal: ${signal})` : ""}: ${stderr.slice(0, 500)}`));
     });
     child.on("error", reject);
   });
@@ -41314,8 +41354,8 @@ function selectWeightTier(vramGB) {
   if (vramGB >= 48)
     return "original";
   if (vramGB >= 16)
-    return "nf4";
-  return "turbo2bit";
+    return "nf4-distilled";
+  return "nf4";
 }
 function detectJetson() {
   try {
@@ -41499,9 +41539,21 @@ async function installPersonaPlex(onInfo, weightTier) {
         return false;
       }
     }
-    await execAsync(`"${pip}" install --quiet "${join54(repoDir, "moshi")}/."`, { timeout: 3e5 });
+    if (isAarch64) {
+      log("ARM64: Installing moshi (--no-deps to preserve JetPack torch)...");
+      await execAsync(`"${pip}" install --quiet --no-deps "${join54(repoDir, "moshi")}/."`, { timeout: 3e5 });
+      log("ARM64: Installing remaining moshi dependencies...");
+      await execAsync(`"${pip}" install --quiet "numpy>=1.26,<2.2" "safetensors>=0.4.0,<0.5" "huggingface-hub>=0.24,<0.25" "einops==0.7" "sentencepiece==0.2" "sounddevice==0.5" "aiohttp>=3.10.5,<3.11"`, { timeout: 3e5 });
+    } else {
+      await execAsync(`"${pip}" install --quiet "${join54(repoDir, "moshi")}/."`, { timeout: 3e5 });
+    }
   } catch (err) {
     log(`Moshi install failed: ${err instanceof Error ? err.message : String(err)}`);
+    if (isAarch64) {
+      log("ARM64: This often means the pip process was OOM-killed.");
+      log("Check: dmesg | grep -i 'oom\\|killed' | tail -5");
+      log("Ensure JetPack PyTorch is installed: pip3 show torch");
+    }
     try {
       await execAsync(`"${pip}" install --quiet torch torchaudio websockets soundfile huggingface_hub`, { timeout: 3e5, stdio: "pipe" });
     } catch {
@@ -41526,6 +41578,104 @@ async function installPersonaPlex(onInfo, weightTier) {
     }
   } catch {
   }
+  try {
+    const sitePackages = execSync27(`"${python}" -c "import moshi, os; print(os.path.dirname(moshi.__file__))"`, {
+      encoding: "utf8",
+      timeout: 5e3,
+      stdio: "pipe"
+    }).trim();
+    const loadersFile = join54(sitePackages, "models", "loaders.py");
+    if (existsSync37(loadersFile)) {
+      let src = readFileSync28(loadersFile, "utf8");
+      if (!src.includes("_dequantize_2bit_state_dict")) {
+        const dequantPatch = `
+import math
+
+# NF2 centroids (Lloyd-Max optimal for Gaussian distribution)
+_NF2_CENTROIDS = torch.tensor([-1.5104, -0.4528, 0.4528, 1.5104])
+
+
+def _is_2bit_quantized(filename):
+    return "turbo2bit" in str(filename).lower() or "2bit" in str(filename).lower()
+
+
+def _fast_wht(x):
+    n = x.shape[-1]
+    h = 1
+    while h < n:
+        x_view = x.view(*x.shape[:-1], -1, 2, h)
+        a = x_view[..., 0, :].clone()
+        b = x_view[..., 1, :].clone()
+        x_view[..., 0, :] = a + b
+        x_view[..., 1, :] = a - b
+        x = x_view.reshape(*x.shape)
+        h *= 2
+    return x / math.sqrt(n)
+
+
+def _dequantize_2bit_state_dict(state_dict):
+    result = {}
+    processed = set()
+    meta_suffixes = (".packed", ".scales", ".shape", ".numel", ".gs", ".np2")
+    base_names = set()
+    for key in state_dict:
+        if key.endswith(".packed"):
+            base_names.add(key[:-len(".packed")])
+    for name in base_names:
+        packed_key = f"{name}.packed"
+        if packed_key in state_dict:
+            gs = state_dict[f"{name}.gs"].item()
+            gs_pow2 = state_dict[f"{name}.np2"].item()
+            numel = state_dict[f"{name}.numel"].item()
+            shape = [s for s in state_dict[f"{name}.shape"].tolist() if s > 0]
+            scales = state_dict[f"{name}.scales"].float()
+            packed = state_dict[packed_key]
+            n_groups = scales.numel()
+            p = packed.reshape(n_groups, gs // 4)
+            codes = torch.zeros(n_groups, gs, dtype=torch.long)
+            for i in range(4):
+                codes[:, i::4] = (p >> (2 * i)) & 0x03
+            dequant = _NF2_CENTROIDS[codes]
+            if gs_pow2 > gs:
+                dequant = torch.cat([dequant, torch.zeros(n_groups, gs_pow2 - gs)], dim=1)
+            dequant = _fast_wht(dequant)
+            dequant = dequant[:, :gs]
+            dequant = dequant * scales.unsqueeze(1)
+            result[name] = dequant.reshape(-1)[:numel].reshape(shape).to(torch.bfloat16)
+            processed.add(name)
+    for name, tensor in state_dict.items():
+        if any(name.endswith(s) for s in meta_suffixes):
+            continue
+        if name not in processed:
+            result[name] = tensor.to(torch.bfloat16)
+    return result
+`;
+        const insertPoint = src.indexOf("\nSAMPLE_RATE");
+        if (insertPoint > 0) {
+          src = src.slice(0, insertPoint) + dequantPatch + src.slice(insertPoint);
+        }
+        src = src.replace(/( +)# Load state_dict\n( +)if filename\.endswith\("\.safetensors"\):/, `$1is_2bit = _is_2bit_quantized(filename)
+$1# Load state_dict \u2014 2-bit must load to CPU for dequant
+$2load_device = "cpu" if is_2bit else dev.type
+$2if filename.endswith(".safetensors"):`);
+        if (src.includes("device=dev.type)")) {
+          src = src.replace("device=dev.type)", "device=load_device)");
+        }
+        const patchPoint = "# Patch 1: expand depformer";
+        if (src.includes(patchPoint) && !src.includes("_dequantize_2bit_state_dict(state_dict)")) {
+          src = src.replace(patchPoint, `# Dequantize 2-bit weights if needed
+    if is_2bit:
+        logger.info("Dequantizing 2-bit TurboQuant weights...")
+        state_dict = _dequantize_2bit_state_dict(state_dict)
+
+    ${patchPoint}`);
+        }
+        writeFileSync16(loadersFile, src);
+        log("Patched loaders.py with 2-bit TurboQuant native dequant support.");
+      }
+    }
+  } catch {
+  }
   if (isAarch64) {
     log("ARM64: Installing bitsandbytes for INT4 inference...");
     try {
@@ -41533,6 +41683,10 @@ async function installPersonaPlex(onInfo, weightTier) {
     } catch {
     }
   }
+  try {
+    await execAsync(`"${pip}" install --quiet accelerate`, { timeout: 12e4, stdio: "pipe" });
+  } catch {
+  }
   try {
     await execAsync(`"${pip}" install --quiet pyloudnorm noisereduce torchaudio`, { timeout: 12e4, stdio: "pipe" });
   } catch {
@@ -41616,49 +41770,101 @@ async function startPersonaPlexDaemon(onInfo) {
   const repoInfo = WEIGHT_REPOS[tier];
   const extraArgs = [];
   if (tier !== "original") {
-    log(`Weight tier: ${tier} (${repoInfo.sizeGB}GB) \u2014 dequantizing to bf16 cache...`);
-    const dequantScript = join54(PERSONAPLEX_DIR, "dequant-loader.py");
     const cachedBf16 = join54(PERSONAPLEX_DIR, "model-bf16-cache.safetensors");
-    if (!existsSync37(dequantScript)) {
-      const shipped = getShippedVoicesDir();
-      if (shipped) {
-        const src = join54(shipped, "dequant-loader.py");
-        if (existsSync37(src))
-          copyFileSync2(src, dequantScript);
-      }
-    }
-    try {
-      const weightPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}'${repoInfo.needsToken ? "" : ", token=False"}))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
-      if (existsSync37(dequantScript) && existsSync37(weightPath)) {
-        try {
-          execSync27(`"${venvPython2}" "${dequantScript}" --input "${weightPath}" --output "${cachedBf16}"`, { timeout: 3e5, stdio: "pipe" });
+    if (tier === "nf4-distilled") {
+      log(`Weight tier: ${tier} \u2014 distilled NF4 (90% token match, ${repoInfo.sizeGB}GB)...`);
+      try {
+        const weightPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}', token=False))"`, { encoding: "utf8", timeout: 6e4, stdio: "pipe" }).trim();
+        if (existsSync37(weightPath)) {
+          if (!existsSync37(cachedBf16)) {
+            log("Converting .pt checkpoint to safetensors (one-time)...");
+            execSync27(`"${venvPython2}" -c "
+import torch; from safetensors.torch import save_file
+state = torch.load('${weightPath}', map_location='cpu', weights_only=True)
+state = {k: v.to(torch.bfloat16) if v.is_floating_point() else v for k, v in state.items()}
+save_file(state, '${cachedBf16}')
+print('Converted')
+"`, { timeout: 18e4, stdio: "pipe" });
+          }
           if (existsSync37(cachedBf16)) {
             extraArgs.push("--moshi-weight", cachedBf16);
-            log(`Using dequantized cache: ${(statSync13(cachedBf16).size / 1024 ** 3).toFixed(1)}GB`);
+            log(`Using distilled weights: ${(statSync13(cachedBf16).size / 1024 ** 3).toFixed(1)}GB`);
+          } else {
+            extraArgs.push("--moshi-weight", weightPath);
           }
-        } catch (e) {
-          log(`Dequantization failed \u2014 server will try to load original weights`);
         }
+      } catch (e) {
+        log(`Failed to load distilled weights \u2014 falling back to standard NF4`);
       }
-      try {
-        const mimiPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer-e351c8d8-checkpoint125.safetensors', token=False))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
-        if (existsSync37(mimiPath))
-          extraArgs.push("--mimi-weight", mimiPath);
-      } catch {
+    } else {
+      log(`Weight tier: ${tier} (${repoInfo.sizeGB}GB) \u2014 dequantizing to bf16 cache...`);
+      const dequantScript = join54(PERSONAPLEX_DIR, "dequant-loader.py");
+      if (!existsSync37(dequantScript)) {
+        const shipped = getShippedVoicesDir();
+        if (shipped) {
+          const src = join54(shipped, "dequant-loader.py");
+          if (existsSync37(src))
+            copyFileSync2(src, dequantScript);
+        }
       }
       try {
-        const tokPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer_spm_32k_3.model', token=False))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
-        if (existsSync37(tokPath))
-          extraArgs.push("--tokenizer", tokPath);
+        const weightPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}'${repoInfo.needsToken ? "" : ", token=False"}))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
+        if (existsSync37(dequantScript) && existsSync37(weightPath)) {
+          try {
+            execSync27(`"${venvPython2}" "${dequantScript}" --input "${weightPath}" --output "${cachedBf16}"`, { timeout: 3e5, stdio: "pipe" });
+            if (existsSync37(cachedBf16)) {
+              extraArgs.push("--moshi-weight", cachedBf16);
+              log(`Using dequantized cache: ${(statSync13(cachedBf16).size / 1024 ** 3).toFixed(1)}GB`);
+            }
+          } catch (e) {
+            log(`Dequantization failed \u2014 server will try to load original weights`);
+          }
+        }
+        try {
+          const mimiPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer-e351c8d8-checkpoint125.safetensors', token=False))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
+          if (existsSync37(mimiPath))
+            extraArgs.push("--mimi-weight", mimiPath);
+        } catch {
+        }
+        try {
+          const tokPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer_spm_32k_3.model', token=False))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
+          if (existsSync37(tokPath))
+            extraArgs.push("--tokenizer", tokPath);
+        } catch {
+        }
       } catch {
+        log(`Weight file not found \u2014 server will download on first run`);
       }
+      extraArgs.push("--hf-repo", repoInfo.repo);
+    }
+  }
+  let hybridEnabled = false;
+  let ollamaModel = process.env["HYBRID_LLM_MODEL"] || "";
+  if (!ollamaModel) {
+    try {
+      const oaConfig = JSON.parse(readFileSync28(join54(homedir13(), ".open-agents", "config.json"), "utf8"));
+      if (oaConfig.model)
+        ollamaModel = oaConfig.model;
     } catch {
-      log(`Weight file not found \u2014 server will download on first run`);
     }
-    extraArgs.push("--hf-repo", repoInfo.repo);
   }
-  log(`Starting PersonaPlex daemon (${tier} tier)...`);
-  const child = spawn19(venvPython2, [
+  if (!ollamaModel)
+    ollamaModel = "qwen3.5:4b";
+  try {
+    const ollamaCheck = execSync27("curl -s http://localhost:11434/api/tags", {
+      timeout: 3e3,
+      stdio: "pipe",
+      encoding: "utf8"
+    });
+    if (ollamaCheck.includes("models")) {
+      hybridEnabled = true;
+      log(`Hybrid mode: PersonaPlex voice + ${ollamaModel} reasoning`);
+    }
+  } catch {
+    log("Ollama not detected \u2014 running PersonaPlex standalone (no hybrid)");
+  }
+  log(`Starting PersonaPlex daemon (${tier} tier${hybridEnabled ? ", hybrid" : ""})...`);
+  const serverArgs = [
     "-m",
     "moshi.server",
     "--host",
@@ -41670,10 +41876,19 @@ async function startPersonaPlexDaemon(onInfo) {
     "--device",
     "cuda",
     ...extraArgs
-  ], {
+  ];
+  if (hybridEnabled)
+    serverArgs.push("--hybrid");
+  const serverEnv = { ...process.env };
+  if (hybridEnabled) {
+    serverEnv["HYBRID_ENABLED"] = "1";
+    serverEnv["HYBRID_LLM_MODEL"] = ollamaModel;
+    serverEnv["HYBRID_MODEL_FAST"] = "qwen3.5:4b";
+  }
+  const child = spawn19(venvPython2, serverArgs, {
     stdio: ["ignore", "pipe", "pipe"],
     detached: true,
-    env: { ...process.env },
+    env: serverEnv,
     cwd: PERSONAPLEX_DIR
   });
   if (child.pid) {
@@ -41990,7 +42205,7 @@ var init_personaplex = __esm({
     WEIGHT_REPOS = {
       original: { repo: "nvidia/personaplex-7b-v1", file: "model.safetensors", sizeGB: 15.6, needsToken: true },
       nf4: { repo: "cudabenchmarktest/personaplex-7b-nf4", file: "model-nf4.safetensors", sizeGB: 4.1, needsToken: false },
-      turbo2bit: { repo: "cudabenchmarktest/personaplex-7b-turbo2bit", file: "model-turbo2bit.safetensors", sizeGB: 2.1, needsToken: false }
+      "nf4-distilled": { repo: "cudabenchmarktest/personaplex-7b-nf4-distilled", file: "student_best.pt", sizeGB: 16.7, needsToken: false }
     };
     PERSONAPLEX_DIR = join54(homedir13(), ".open-agents", "voice", "personaplex");
     PID_FILE = join54(PERSONAPLEX_DIR, "daemon.pid");
@@ -43757,7 +43972,8 @@ function fitToWidth(text, width) {
   return text + " ".repeat(width - visible.length);
 }
 function showDropPanel(opts) {
-  const { title, instruction = "Drag and drop a file here, or type/paste a path", allowedExtensions = [], typeLabel, rl } = opts;
+  const { title, instruction = "Drag and drop a file here, or type/paste a path", allowedExtensions = [], typeLabel, rl, borderColor } = opts;
+  const bc = borderColor ?? dc.cyan;
   return new Promise((resolve_) => {
     const stdin = process.stdin;
     const hadRawMode = stdin.isRaw;
@@ -43809,16 +44025,16 @@ function showDropPanel(opts) {
       const bottomPad = Math.max(0, availableForPadding - topPad);
       const lines = [];
       const borderH = "\u2508".repeat(Math.max(2, cols - 4));
-      const emptyPipe = `  ${dc.cyan("\u250A")}${" ".repeat(innerSpace)}${dc.cyan("\u250A")}`;
-      lines.push(`  ${dc.cyan(borderH)}`);
+      const emptyPipe = `  ${bc("\u250A")}${" ".repeat(innerSpace)}${bc("\u250A")}`;
+      lines.push(`  ${bc(borderH)}`);
       for (let i = 0; i < topPad; i++)
         lines.push(emptyPipe);
       for (const line of content) {
-        lines.push(`  ${dc.cyan("\u250A")}${line}${dc.cyan("\u250A")}`);
+        lines.push(`  ${bc("\u250A")}${line}${bc("\u250A")}`);
       }
       for (let i = 0; i < bottomPad; i++)
         lines.push(emptyPipe);
-      lines.push(`  ${dc.cyan(borderH)}`);
+      lines.push(`  ${bc(borderH)}`);
       lines.push(`  ${dc.dim("Enter confirm  Esc cancel")}`);
       overlayWrite(lines.join("\n") + "\n");
     }
@@ -49644,18 +49860,22 @@ async function handleSlashCommand(input, ctx) {
           if (currentVoiceModel === "personaplex") {
             if (!cloneArg) {
               const dropResult = await showDropPanel({
-                title: "PersonaPlex Voice Clone \u2014 Drop Audio File",
-                instruction: "Drop a WAV file (4-10s clean speech) to clone into PersonaPlex",
+                title: "PersonaPlex Voice Clone",
+                instruction: "Drop an audio file (4-10s clean speech) to clone a voice",
                 allowedExtensions: [".wav", ".mp3", ".ogg", ".flac", ".m4a", ".opus", ".aac"],
                 typeLabel: "Audio files",
-                rl: ctx.rl
+                rl: ctx.rl,
+                borderColor: c2.green
               });
               if (dropResult.confirmed && dropResult.path) {
-                const voiceName2 = dropResult.path.replace(/.*[\\/]/, "").replace(/\.[^.]+$/, "").replace(/[^a-zA-Z0-9_-]/g, "_");
+                const defaultName = dropResult.path.replace(/.*[\\/]/, "").replace(/\.[^.]+$/, "").replace(/[^a-zA-Z0-9_-]/g, "_");
+                renderInfo(`File: ${dropResult.path}`);
+                renderInfo(`Voice name: ${defaultName} (derived from filename)`);
+                renderInfo("Cloning voice with preprocessing (denoise + normalize + multi-segment)...");
                 const { clonePersonaPlexVoice: clonePersonaPlexVoice3 } = await Promise.resolve().then(() => (init_personaplex(), personaplex_exports));
-                const result2 = await clonePersonaPlexVoice3(dropResult.path, voiceName2, (m) => renderInfo(m));
+                const result2 = await clonePersonaPlexVoice3(dropResult.path, defaultName, (m) => renderInfo(m));
                 if (result2)
-                  renderInfo(`Voice "${voiceName2}" ready \u2014 use /voice list to see all voices`);
+                  renderInfo(`Voice "${defaultName}" ready \u2014 use /voice list to see all voices`);
               } else {
                 renderInfo("Voice clone cancelled.");
               }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.185.34",
+  "version": "0.185.36",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",