omnius 1.0.885 → 1.0.888

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 38501421 bytes
-sha256: 93169deae92540705e32001d313d4ce473ab593892b4961028c6d743b672aa17
+size: 38545361 bytes
+sha256: eba5e49632d5ab67fbacfa057dddcb9ca05e8dc40ece7a60a65566d776e1c8e6

package/dist/library.js

@@ -345015,7 +345015,7 @@ var StreamInterruptedError = class extends Error {
   fatal = true;
   retryable = false;
   constructor(failure) {
-    super("Inference ended after streaming output began. Automatic replay is disabled.");
+    super("Inference ended after model progress began. Automatic replay is disabled.");
     this.failure = failure;
     this.name = "InferenceStreamInterruptedError";
   }
@@ -345464,7 +345464,10 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     }
     let localProgress = false;
     const localPacket = { ...this.localRequest(request), onLocalInferenceSemanticProgress: () => {
-      localProgress = true;
+      if (!localProgress) {
+        localProgress = true;
+        this.lifecycle(request, "selected", void 0, void 0, void 0, void 0, void 0, true);
+      }
       request.onLocalInferenceSemanticProgress?.();
     } };
     if (!this.eligible(request)) {
@@ -345472,9 +345475,13 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         return local(request);
       this.fence(request);
       try {
-        return await this.attempt(localPacket, this.localAttemptTimeoutMs(request), local);
+        return await this.attempt(localPacket, this.localAttemptTimeoutMs(request), local, "local_first_frame", true);
       } catch (error) {
         this.fence(request);
+        if (this.retainedLocalAdmission(error, request))
+          throw error;
+        if (localProgress && this.initialWaitTimeout(error))
+          throw new StreamInterruptedError(safeFailure(error, request));
         throw this.localWaiting(error) ?? (!localProgress && this.initialWaitTimeout(error) ? new LocalInferenceNotReadyError("initial_response_timeout", error) : error);
       }
     }
@@ -345487,7 +345494,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         if (bounded3.localAdmissionFailure !== void 0)
           throw bounded3.localAdmissionFailure;
         return local(bounded3);
-      });
+      }, "local_first_frame", true);
       if (!usableResult(result))
         throw new EmptyInferenceResponse();
       this.lifecycle(request, "succeeded");
@@ -345505,6 +345512,8 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         this.lifecycle(request, "failed", void 0, classified);
       if (classified instanceof LocalInferenceNotReadyError)
         throw classified;
+      if (localProgress && this.initialWaitTimeout(error))
+        throw new StreamInterruptedError(safeFailure(error, request));
       if (!transportFailure(error))
         throw error;
     }
@@ -345588,7 +345597,10 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     }
     let localProgress = false;
     const localPacket = { ...this.localRequest(request), onLocalInferenceSemanticProgress: () => {
-      localProgress = true;
+      if (!localProgress) {
+        localProgress = true;
+        this.lifecycle(request, "selected", void 0, void 0, void 0, void 0, void 0, true);
+      }
       request.onLocalInferenceSemanticProgress?.();
     } };
     if (!this.eligible(request)) {
@@ -345601,6 +345613,10 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
         yield* this.boundedStream(localPacket, this.localAttemptTimeoutMs(request), local, true, "local_first_frame");
       } catch (error) {
         this.fence(request);
+        if (this.retainedLocalAdmission(error, request))
+          throw error;
+        if (localProgress && this.initialWaitTimeout(error))
+          throw new StreamInterruptedError(safeFailure(error, request));
         throw this.localWaiting(error) ?? (!localProgress && this.initialWaitTimeout(error) ? new LocalInferenceNotReadyError("initial_response_timeout", error) : error);
       }
       return;
@@ -345623,7 +345639,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
       return;
     } catch (error) {
       this.fence(request);
-      if (emitted) {
+      if (emitted || localProgress && this.initialWaitTimeout(error)) {
         if (localSelected)
           this.lifecycle(request, "failed", void 0, error, frames);
         throw new StreamInterruptedError(safeFailure(error, request, void 0, frames));
@@ -345960,7 +345976,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     }
     this.lifecycle(request, "selected", route);
   }
-  lifecycle(request, phase, route, error, frames, completedBinding, decision2) {
+  lifecycle(request, phase, route, error, frames, completedBinding, decision2, semanticProgress = false) {
     this.fence(request);
     if (!request.onInferenceRouteLifecycle)
       return;
@@ -345973,6 +345989,7 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
       providerId: route ? route.keyless?.id ?? route.providerId ?? "remote" : this.primary.providerId ?? "ollama",
       backendType: route ? "vllm" : this.primary.backendType ?? "ollama",
       ...request.logicalRequestId ? { logicalRequestId: request.logicalRequestId } : {},
+      ...semanticProgress ? { semanticProgress: true } : {},
       ...route && (this.schemaRestrictions(route).length || promptOnly) ? { schemaAdapter: {
         omittedNativeKeywords: this.schemaRestrictions(route),
         hostValidation: "unchanged",
@@ -346039,11 +346056,16 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
       route.apiKey = current2.apiKey;
     }
   }
-  async attempt(request, timeoutMs, operation, deadlineKind = "request") {
+  async attempt(request, timeoutMs, operation, deadlineKind = "request", initialResponseOnly = false) {
     const controller = new AbortController();
     const signal = request.signal ? AbortSignal.any([request.signal, controller.signal]) : controller.signal;
     let timer = setTimeout(() => controller.abort(new AttemptTimeout(deadlineKind)), Math.max(1, timeoutMs));
-    const bounded3 = { ...request, signal };
+    const bounded3 = { ...request, signal, ...initialResponseOnly ? { onLocalInferenceSemanticProgress: () => {
+      if (signal.aborted)
+        return;
+      clearTimeout(timer);
+      request.onLocalInferenceSemanticProgress?.();
+    } } : {} };
     try {
       return await this.abortable(operation(bounded3), signal);
     } finally {
@@ -346055,15 +346077,24 @@ var RemoteInferenceFallback = class _RemoteInferenceFallback {
     const controller = new AbortController();
     const signal = request.signal ? AbortSignal.any([request.signal, controller.signal]) : controller.signal;
     let timer = setTimeout(() => controller.abort(new AttemptTimeout(deadlineKind)), Math.max(1, timeoutMs));
-    const iterator = operation({ ...request, signal });
+    const progress = () => {
+      if (signal.aborted)
+        return;
+      clearTimeout(timer);
+      request.onLocalInferenceSemanticProgress?.();
+    };
+    const iterator = operation({
+      ...request,
+      signal,
+      ...firstFrameOnly ? { onLocalInferenceSemanticProgress: progress } : {}
+    });
     try {
       for (; ; ) {
         const step = await this.abortable(iterator.next(), signal);
         if (step.done)
           return;
         if (firstFrameOnly && (step.value.type === "tool_call_delta" || step.value.type === "content" && Boolean(step.value.content))) {
-          clearTimeout(timer);
-          request.onLocalInferenceSemanticProgress?.();
+          progress();
         }
         yield step.value;
       }
@@ -368550,155 +368581,6 @@ async function runStuckAnalyzer(opts) {
   };
 }
 
-// packages/orchestrator/dist/thinking-mode-collapse.js
-function renderModeCollapsePrompt(inputs) {
-  const lines = [];
-  lines.push("# THINKING STREAM MODE-COLLAPSE ANALYSIS");
-  lines.push("");
-  lines.push("You are a META-ANALYSIS sub-agent. A model (the generator) has been");
-  lines.push("producing internal reasoning tokens for an extended period without");
-  lines.push("emitting any visible output or tool calls. A sample of its reasoning");
-  lines.push("stream is provided below.");
-  lines.push("");
-  lines.push("Your job: classify whether the generator is experiencing a MODE COLLAPSE");
-  lines.push("(a reasoning loop where it repeats the same idea, goes in circles, or");
-  lines.push("degrades into noise) or is still PRODUCTIVELY REASONING (making linear");
-  lines.push("progress toward a conclusion).");
-  lines.push("");
-  lines.push("## Collapse categories");
-  lines.push("");
-  lines.push("- mode_collapse_exact_repeat: The model emits the exact same sentence or");
-  lines.push("  paragraph 2+ times. The text is verbatim identical.");
-  lines.push("");
-  lines.push("- mode_collapse_semantic_loop: The model rephrases the same idea");
-  lines.push("  repeatedly with different wording. Same concept, new sentences.");
-  lines.push("  No new reasoning steps or conclusions.");
-  lines.push("");
-  lines.push("- mode_collapse_circular: The model alternates between two or more");
-  lines.push("  positions/approaches without converging. A -> B -> A -> B pattern.");
-  lines.push("");
-  lines.push("- mode_collapse_degraded: The reasoning quality is visibly degrading —");
-  lines.push("  sentences get shorter, less coherent, fragmented, or trail off.");
-  lines.push("  The model is losing the thread.");
-  lines.push("");
-  lines.push("- mode_collapse_stuck_at_start: The model keeps restating the problem or");
-  lines.push("  the first step but never moves to subsequent steps. Stuck in a");
-  lines.push("  preamble loop.");
-  lines.push("");
-  lines.push("- healthy: The reasoning is making linear progress. Each sentence builds");
-  lines.push("  on or advances from the previous one. The model is analyzing,");
-  lines.push("  comparing, planning, or synthesizing in a productive direction.");
-  lines.push("");
-  lines.push("## Context");
-  lines.push(`Goal: ${inputs.goal.slice(0, 500)}`);
-  lines.push(`Total thinking produced: ${inputs.totalThinkingChars} characters over ${(inputs.thinkingDurationMs / 1e3).toFixed(0)}s`);
-  if (inputs.lastToolNames.length > 0) {
-    lines.push(`Recent tool calls before thinking: ${inputs.lastToolNames.join(", ")}`);
-  }
-  lines.push("");
-  lines.push("## Thinking sample (most recent portion)");
-  lines.push("");
-  lines.push("```");
-  lines.push(inputs.thinkingSample.slice(0, 3e3));
-  lines.push("```");
-  lines.push("");
-  lines.push("## Output format");
-  lines.push("");
-  lines.push("Emit a SINGLE JSON code block with this exact shape:");
-  lines.push("");
-  lines.push("```json");
-  lines.push("{");
-  lines.push('  "category": "<one of the categories above>",');
-  lines.push('  "diagnosis": "<1-2 sentence diagnosis>",');
-  lines.push('  "confidence": <0.0-1.0>,');
-  lines.push('  "suggested_action": "<one of: none, abort_and_retry_no_think, abort_and_nudge, abort_and_force_tool_call>"');
-  lines.push("}");
-  lines.push("```");
-  lines.push("");
-  lines.push("Only flag as collapse if confidence > 0.7. When unsure, classify as healthy");
-  lines.push("with low confidence — false positives (aborting productive reasoning) are");
-  lines.push("more harmful than false negatives (letting a collapse continue briefly).");
-  return lines.join("\n");
-}
-function parseModeCollapseVerdict(rawResponse) {
-  const fallback = (msg) => ({
-    category: "healthy",
-    diagnosis: `(sub-agent parse failed: ${msg})`,
-    confidence: 0,
-    suggested_action: "none",
-    raw: rawResponse,
-    parseFallback: true
-  });
-  if (!rawResponse || typeof rawResponse !== "string" || rawResponse.trim().length === 0) {
-    return fallback("empty response");
-  }
-  const fenceMatch = rawResponse.match(/```(?:json)?\s*\n([\s\S]*?)\n```/);
-  let jsonText = null;
-  if (fenceMatch) {
-    jsonText = fenceMatch[1].trim();
-  } else {
-    const first = rawResponse.indexOf("{");
-    const last = rawResponse.lastIndexOf("}");
-    if (first !== -1 && last > first)
-      jsonText = rawResponse.slice(first, last + 1);
-  }
-  if (!jsonText)
-    return fallback("no JSON block found");
-  let parsed;
-  try {
-    parsed = JSON.parse(jsonText);
-  } catch (e) {
-    return fallback(`JSON parse: ${e instanceof Error ? e.message : String(e)}`);
-  }
-  if (!parsed || typeof parsed !== "object")
-    return fallback("not an object");
-  const validCategories = [
-    "healthy",
-    "mode_collapse_exact_repeat",
-    "mode_collapse_semantic_loop",
-    "mode_collapse_circular",
-    "mode_collapse_degraded",
-    "mode_collapse_stuck_at_start"
-  ];
-  const category = typeof parsed.category === "string" && validCategories.includes(parsed.category) ? parsed.category : "healthy";
-  const diagnosis = typeof parsed.diagnosis === "string" && parsed.diagnosis.trim().length > 0 ? parsed.diagnosis.slice(0, 500) : "(no diagnosis)";
-  const confidence2 = typeof parsed.confidence === "number" ? Math.max(0, Math.min(1, parsed.confidence)) : 0;
-  const validActions = [
-    "none",
-    "abort_and_retry_no_think",
-    "abort_and_nudge",
-    "abort_and_force_tool_call"
-  ];
-  const suggested_action = typeof parsed.suggested_action === "string" && validActions.includes(parsed.suggested_action) ? parsed.suggested_action : "none";
-  return {
-    category,
-    diagnosis,
-    confidence: confidence2,
-    suggested_action,
-    raw: rawResponse,
-    parseFallback: false
-  };
-}
-async function runModeCollapseAnalysis(opts) {
-  const startMs = Date.now();
-  const prompt = renderModeCollapsePrompt(opts.inputs);
-  const promptBytes = Buffer.byteLength(prompt, "utf-8");
-  let raw = "";
-  try {
-    raw = await opts.callable(prompt);
-  } catch {
-    raw = "";
-  }
-  const responseBytes = Buffer.byteLength(raw, "utf-8");
-  const verdict = parseModeCollapseVerdict(raw);
-  return {
-    verdict,
-    promptBytes,
-    responseBytes,
-    durationMs: Date.now() - startMs
-  };
-}
-
 // packages/orchestrator/dist/native-ollama-messages.js
 function nativeOllamaMessages(messages2) {
   const toolNames = /* @__PURE__ */ new Map();
@@ -428134,28 +428016,6 @@ ${description}`
     });
     this._streamingExecutor.reset();
     let lastFinalizedIdx = -1;
-    let firstThinkingTime = 0;
-    let watchdogTripped = false;
-    let thinkingSample = "";
-    let watchdogMs = typeof backend.getThinkingWatchdogMs === "function" ? backend.getThinkingWatchdogMs() : 6e4;
-    if (watchdogMs > 15e3) {
-      let estimatedHistoryTokens = 0;
-      for (const m of request.messages) {
-        const c = typeof m.content === "string" ? m.content : "";
-        estimatedHistoryTokens += c.length;
-        if (m.tool_calls) {
-          for (const tc of m.tool_calls) {
-            estimatedHistoryTokens += tc.function?.arguments?.length ?? 0;
-          }
-        }
-      }
-      estimatedHistoryTokens = Math.round(estimatedHistoryTokens / 4);
-      if (estimatedHistoryTokens > 28e3) {
-        watchdogMs = Math.min(watchdogMs, 3e4);
-      } else if (estimatedHistoryTokens > 24e3) {
-        watchdogMs = Math.min(watchdogMs, 45e3);
-      }
-    }
     for await (const chunk of backend.chatCompletionStream(request)) {
       request.signal?.throwIfAborted();
       if (this.aborted || this._pendingSteeringPreemption)
@@ -428189,25 +428049,6 @@ ${description}`
         } else {
           thinkingContent += fragment;
         }
-        if (kind === "thinking") {
-          if (firstThinkingTime === 0)
-            firstThinkingTime = Date.now();
-          if (!watchdogTripped) {
-            const elapsed = Date.now() - firstThinkingTime;
-            if (elapsed >= watchdogMs) {
-              watchdogTripped = true;
-              thinkingSample = thinkingContent.slice(-3e3);
-              this.emit({
-                type: "status",
-                content: `Thinking watchdog tripped after ${Math.round(elapsed / 1e3)}s (threshold=${(watchdogMs / 1e3).toFixed(0)}s) — model may be in a reasoning loop`,
-                timestamp: (/* @__PURE__ */ new Date()).toISOString()
-              });
-              break;
-            }
-          }
-        } else {
-          firstThinkingTime = 0;
-        }
         const buf = kind === "content" ? content : thinkingContent;
         if (buf.length > 200 && buf.length % 100 < fragment.length) {
           const tail = buf.slice(-600);
@@ -428397,80 +428238,9 @@ arguments=` : ""}` + (chunk.toolCallArgs ?? "");
     };
     if (thinkOnlyResponse)
       resp._thinkOnly = true;
-    if (watchdogTripped)
-      resp._watchdogTripped = true;
     resp._rawTypedModelContent = content;
-    if (watchdogTripped && thinkingSample.length > 0) {
-      this._fireModeCollapseAnalysis(backend, thinkingSample, thinkingContent.length, Date.now() - (firstThinkingTime || Date.now()), turn);
-    }
     return resp;
   }
-  /**
-   * Fire-and-forget mode-collapse analysis. Called when the streaming
-   * watchdog detects extended thinking without visible output. Fires a
-   * sub-agent to classify the thinking pattern and updates the backend's
-   * elevated scrutiny state for future turns.
-   *
-   * The sub-agent call runs asynchronously and does NOT block the current
-   * response — the watchdog already aborted it. This is purely diagnostic
-   * and adjusts thresholds for subsequent turns.
-   *
-   * Gated behind OMNIUS_MODE_COLLAPSE_ANALYZER env (default "on") so users
-   * who want zero extra round-trips can disable it.
-   */
-  _fireModeCollapseAnalysis(backend, thinkingSample, totalThinkingChars, thinkingDurationMs, turn) {
-    const gate = (process.env["OMNIUS_MODE_COLLAPSE_ANALYZER"] || "on").toLowerCase();
-    if (gate !== "on" && gate !== "1" && gate !== "true")
-      return;
-    const goal = this._taskState?.originalGoal || this._taskState?.goal || "";
-    const callable = async (prompt) => {
-      try {
-        const r = await this._dispatchInterruptionOwnedModelRequest(backend, {
-          messages: [
-            {
-              role: "system",
-              content: "You are a META-ANALYSIS sub-agent. Classify the generator's thinking stream and emit a structured JSON verdict."
-            },
-            { role: "user", content: prompt }
-          ],
-          tools: [],
-          temperature: 0,
-          maxTokens: parseInt(process.env["OMNIUS_MODE_COLLAPSE_MAX_TOKENS"] || "1024", 10) || 1024,
-          timeoutMs: parseInt(process.env["OMNIUS_MODE_COLLAPSE_TIMEOUT_MS"] || "30000", 10) || 3e4
-        }, `model:mode_collapse:${turn}`);
-        const c = r?.choices?.[0]?.message?.content;
-        return typeof c === "string" ? c : "";
-      } catch {
-        return "";
-      }
-    };
-    runModeCollapseAnalysis({
-      inputs: {
-        thinkingSample,
-        totalThinkingChars,
-        thinkingDurationMs,
-        goal,
-        lastToolNames: []
-      },
-      callable
-    }).then((result) => {
-      if (result.verdict.category !== "healthy" && result.verdict.confidence >= 0.7) {
-        backend.recordModeCollapse();
-        this.emit({
-          type: "status",
-          content: `Mode-collapse analysis: ${result.verdict.category} (conf=${result.verdict.confidence.toFixed(2)}) — ${result.verdict.diagnosis.slice(0, 120)}. Collapse #${backend.getModeCollapseCount()}, scrutiny level ${backend.getElevatedScrutinyLevel()}`,
-          timestamp: (/* @__PURE__ */ new Date()).toISOString()
-        });
-      } else {
-        this.emit({
-          type: "status",
-          content: `Thinking analysis: ${result.verdict.category} (conf=${result.verdict.confidence.toFixed(2)}) — thinking was healthy, continuing`,
-          timestamp: (/* @__PURE__ */ new Date()).toISOString()
-        });
-      }
-    }).catch(() => {
-    });
-  }
 };
 function stripThinkBlocks(s) {
   if (!s)
@@ -429363,12 +429133,62 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
     });
   }
   async chatCompletion(request) {
+    if (this._provider.protocol === "ollama")
+      request.logicalRequestId = normalizeOllamaLogicalRequestId(request.logicalRequestId ?? crypto.randomUUID());
     return this.remoteFallback().complete(this.requestWithRunSignal(request), async (localRequest) => {
       if (localRequest.localAdmissionFailure !== void 0)
         throw localRequest.localAdmissionFailure;
+      if ((this._provider.protocol === "ollama" || this._provider.protocol === "openai-chat") && isLocalInferenceEndpoint(this.baseUrl) && localRequest.maxTokens > 0)
+        return this.aggregateLocalCompletion(localRequest);
       return this.chatCompletionDirect(localRequest);
     });
   }
+  async aggregateLocalCompletion(request, emptyRecovery) {
+    let content = "", thinking = "";
+    let usage, responseDiagnostics;
+    let finishReason2;
+    const calls = /* @__PURE__ */ new Map();
+    for await (const chunk of this.chatCompletionStreamDirect(request, emptyRecovery)) {
+      request.signal?.throwIfAborted();
+      if (chunk.type === "content" && chunk.content) {
+        if (chunk.thinking)
+          thinking += chunk.content;
+        else
+          content += chunk.content;
+      } else if (chunk.type === "tool_call_delta") {
+        const index = chunk.toolCallIndex ?? 0;
+        const call2 = calls.get(index) ?? { id: chunk.toolCallId ?? (request.configuredEndpointOnly === true ? "" : crypto.randomUUID()), name: "", args: "" };
+        if (chunk.toolCallId)
+          call2.id = chunk.toolCallId;
+        if (chunk.toolCallName)
+          call2.name = chunk.toolCallName;
+        if (chunk.toolCallArgs)
+          call2.args += chunk.toolCallArgs;
+        calls.set(index, call2);
+      } else if (chunk.type === "usage")
+        usage = chunk.usage;
+      else if (chunk.type === "finish") {
+        finishReason2 = chunk.finishReason;
+        if (chunk.responseDiagnostics)
+          responseDiagnostics = chunk.responseDiagnostics;
+      }
+    }
+    request.signal?.throwIfAborted();
+    const message = normalizeProviderToolMessage({ content, tool_calls: [...calls.values()].map((call2) => ({
+      id: call2.id,
+      type: "function",
+      function: { name: call2.name, arguments: call2.args || "{}" }
+    })) }, this.model, request.configuredEndpointOnly === true);
+    return {
+      choices: [{ message: {
+        content: thinking ? `<think>${thinking}</think>${message.content ?? ""}` : message.content,
+        ...message.toolCalls?.length ? { toolCalls: message.toolCalls } : {}
+      } }],
+      ...responseDiagnostics ? { responseDiagnostics } : {},
+      ...finishReason2 ? { outputTruncated: finishReason2 === "length" } : {},
+      ...usage ? { usage: buildAgenticUsage(usage) } : {}
+    };
+  }
   async chatCompletionDirect(request) {
     if (this._provider.protocol === "ollama") {
       request.logicalRequestId = normalizeOllamaLogicalRequestId(request.logicalRequestId ?? crypto.randomUUID());
@@ -429817,6 +429637,8 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
    * Ollama pool routing as non-stream completions.
    */
   async *chatCompletionStream(request) {
+    if (this._provider.protocol === "ollama")
+      request.logicalRequestId = normalizeOllamaLogicalRequestId(request.logicalRequestId ?? crypto.randomUUID());
     const backend = this;
     yield* this.remoteFallback().stream(this.requestWithRunSignal(request), async function* (localRequest) {
       if (localRequest.localAdmissionFailure !== void 0)
@@ -429824,7 +429646,7 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
       yield* backend.chatCompletionStreamDirect(localRequest);
     });
   }
-  async *chatCompletionStreamDirect(request) {
+  async *chatCompletionStreamDirect(request, emptyRecovery) {
     if (this._provider.protocol === "ollama") {
       request.logicalRequestId = normalizeOllamaLogicalRequestId(request.logicalRequestId ?? crypto.randomUUID());
     }
@@ -429909,7 +429731,7 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
       stream_options: { include_usage: true }
     };
     const isOllamaTransport = this._provider.protocol === "ollama";
-    const useOllamaPool = request.configuredEndpointOnly !== true && isOllamaTransport && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
+    const useOllamaPool = !emptyRecovery && request.configuredEndpointOnly !== true && isOllamaTransport && shouldUseOllamaPoolForBaseUrl(this.baseUrl);
     if (isOllamaTransport)
       body["think"] = effectiveThink;
     if (!isOllamaTransport && request.chatTemplateKwargs)
@@ -429931,7 +429753,7 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
       queuePolicy: request.poolQueuePolicy,
       queueTimeoutMs: request.poolQueueTimeoutMs
     }) : null;
-    let requestBaseUrl = poolSlot?.baseUrl ?? this.baseUrl;
+    let requestBaseUrl = emptyRecovery?.baseUrl ?? poolSlot?.baseUrl ?? this.baseUrl;
     let poolSuccess = false;
     const releasePoolSlot = (success) => {
       if (!poolSlot)
@@ -429948,10 +429770,12 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
     const armStreamTimeout = () => {
       const waiting = localInitialDeadline !== void 0 && !localSemanticProgress;
       const timeout = waiting ? Math.max(1, localInitialDeadline - Date.now()) : streamTimeoutMs;
+      if (streamTimeoutHandle) {
+        clearTimeout(streamTimeoutHandle);
+        streamTimeoutHandle = void 0;
+      }
       if (timeout === null)
         return;
-      if (streamTimeoutHandle)
-        clearTimeout(streamTimeoutHandle);
       streamTimeoutHandle = setTimeout(() => {
         streamAbort.abort(waiting ? new LocalInferenceNotReadyError("initial_response_timeout") : new Error(`stream timeout: no response or chunk within ${(timeout / 1e3).toFixed(0)}s`));
       }, timeout);
@@ -430235,16 +430059,25 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
           yield { type: "finish", finishReason: finishReason2 };
         }
       }
-      if (isOllamaTransport && effectiveThink === false && sawReasoningTokens && !accumulatedContent.trim() && !sawToolCallDeltas) {
+      const emptyVisibleOutcome = classifyThinkOutcome(accumulatedContent);
+      const literalThinkingOnly = accumulatedContent.includes("<think>") && (emptyVisibleOutcome === "empty_after_strip" || emptyVisibleOutcome === "unclosed_think");
+      if (!emptyRecovery && isOllamaTransport && effectiveThink === false && (sawReasoningTokens && !accumulatedContent.trim() || literalThinkingOnly) && !sawToolCallDeltas) {
         const nativeRecoveryRequest = withLogicalRequestVariant(request, "native-empty-recovery");
-        const nativeRecovery = await this.nativeOllamaChatCompletion(nativeRecoveryRequest, {
+        const nativeRecovery = isLocalInferenceEndpoint(this.baseUrl) && request.maxTokens > 0 ? await this.aggregateLocalCompletion({
+          ...nativeRecoveryRequest,
+          preferNativeOllamaChat: true,
+          think: false,
+          signal: streamAbort.signal,
+          onLocalInferenceSemanticProgress: markLocalSemanticProgress
+        }, { baseUrl: requestBaseUrl }) : await this.nativeOllamaChatCompletion(nativeRecoveryRequest, {
           baseUrl: requestBaseUrl,
           signal: streamAbort.signal,
           noThinkDirectives: true
         });
         const recoveredMessage = nativeRecovery.choices[0]?.message;
         const recoveredContent = recoveredMessage?.content ?? "";
-        if (recoveredContent) {
+        const usableRecoveryContent = recoveredContent.trim().length >= 2 && classifyThinkOutcome(recoveredContent) === null;
+        if (usableRecoveryContent) {
           accumulatedContent = recoveredContent;
           yield { type: "content", content: recoveredContent };
         }
@@ -430273,6 +430106,17 @@ var OllamaAgenticBackend = class _OllamaAgenticBackend {
           type: "finish",
           finishReason: sawToolCallDeltas ? "tool_calls" : "stop"
         };
+        if (!usableRecoveryContent && !sawToolCallDeltas && request.disableEmptyContentRecovery !== true) {
+          const changedRetry = withLogicalRequestVariant(request, "openai-empty-retry");
+          yield* this.chatCompletionStreamDirect({
+            ...changedRetry,
+            responseFormat: void 0,
+            response_format: void 0,
+            think: false,
+            signal: streamAbort.signal,
+            onLocalInferenceSemanticProgress: markLocalSemanticProgress
+          }, { baseUrl: requestBaseUrl });
+        }
       }
       this._finalizeStreamGuard(effectiveThink, accumulatedContent, accumulatedThinking, sawReasoningTokens);
       poolSuccess = true;