newmark-agent 0.5.4 → 0.5.8

package/dist/conversation-utility-host.bundle.cjs

@@ -328670,7 +328670,7 @@ function providerStreamTimeoutError(timeoutMs) {
   error.message = `Stream read timeout after ${timeoutMs}ms`;
   return error;
 }
-async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
+async function readProviderStreamChunk(reader, signal, timeoutMs = 0) {
   if (signal.aborted) throw providerAbortError(signal);
   let timer;
   let onAbort;
@@ -328678,9 +328678,9 @@ async function readProviderStreamChunk(reader, signal, timeoutMs = 3e4) {
     onAbort = () => reject(providerAbortError(signal));
     signal.addEventListener("abort", onAbort, { once: true });
   });
-  const timeoutPromise = new Promise((_3, reject) => {
+  const timeoutPromise = timeoutMs > 0 ? new Promise((_3, reject) => {
     timer = setTimeout(() => reject(providerStreamTimeoutError(timeoutMs)), timeoutMs);
-  });
+  }) : new Promise(() => void 0);
   try {
     return await Promise.race([reader.read(), abortPromise, timeoutPromise]);
   } catch (error) {
@@ -328709,6 +328709,30 @@ function parseProviderSse(raw) {
   }
   return events;
 }
+function assembleCompatibleToolArguments(parts) {
+  const nonEmpty = (parts || []).map(String).filter((part) => part && part !== "null");
+  if (!nonEmpty.length) return "{}";
+  const isJsonObject = (value) => {
+    try {
+      const parsed = JSON.parse(value);
+      return !!parsed && typeof parsed === "object" && !Array.isArray(parsed);
+    } catch {
+      return false;
+    }
+  };
+  const incremental = nonEmpty.join("");
+  if (isJsonObject(incremental)) return incremental;
+  let compatible = "";
+  for (const incoming of nonEmpty) {
+    if (!compatible) compatible = incoming;
+    else if (incoming === compatible) continue;
+    else if (incoming.startsWith(compatible)) compatible = incoming;
+    else if (compatible.startsWith(incoming)) continue;
+    else compatible += incoming;
+  }
+  if (isJsonObject(compatible)) return compatible;
+  return [...nonEmpty].reverse().find(isJsonObject) || compatible;
+}
 function isContentPolicyBlocked(json) {
   const choices = Array.isArray(json.choices) ? json.choices : [];
   const choice = choices[0] || {};
@@ -328937,6 +328961,8 @@ var ChatCompletionsAdapter = class {
     let contentPolicyBlocked = false;
     let emittedContent = false;
     let emittedTool = false;
+    let emittedReasoning = false;
+    let explicitCompletion = false;
     try {
       while (true) {
         const { done, value } = await readProviderStreamChunk(reader, signal);
@@ -328948,7 +328974,10 @@ var ChatCompletionsAdapter = class {
           const trimmed = line.trim();
           if (!trimmed.startsWith("data: ")) continue;
           const data = trimmed.slice(6);
-          if (data === "[DONE]") continue;
+          if (data === "[DONE]") {
+            explicitCompletion = true;
+            continue;
+          }
           let json;
           try {
             json = JSON.parse(data);
@@ -328961,11 +328990,16 @@ var ChatCompletionsAdapter = class {
           }
           if (isContentPolicyBlocked(json)) contentPolicyBlocked = true;
           const choices = Array.isArray(json.choices) ? json.choices : [];
-          const delta = choices[0]?.delta;
+          const choice = choices[0];
+          if (choice?.finish_reason !== void 0 && choice.finish_reason !== null) explicitCompletion = true;
+          const delta = choice?.delta;
           if (!delta) continue;
           if (delta.reasoning_content) {
             const reasoning = this.extractText(delta.reasoning_content);
-            if (reasoning) yield { type: "reasoning.summary.delta", delta: reasoning };
+            if (reasoning) {
+              emittedReasoning = true;
+              yield { type: "reasoning.summary.delta", delta: reasoning };
+            }
           }
           const textDelta = this.extractText(delta.content);
           if (textDelta) {
@@ -329010,13 +329044,17 @@ var ChatCompletionsAdapter = class {
             type: "tool_call.completed",
             id: currentToolCall.id,
             name: currentToolCall.name,
-            arguments: currentToolCall.argumentParts.join("")
+            arguments: assembleCompatibleToolArguments(currentToolCall.argumentParts)
           };
         }
       } else if (!emittedContent && !emittedTool && contentPolicyBlocked) {
         yield { type: "response.failed", error: "[Error] Content policy refusal (content_filter)." };
         return;
       }
+      if (!explicitCompletion && !emittedContent && !emittedTool && !emittedReasoning) {
+        yield { type: "response.failed", error: "[LLM Error] Chat stream ended before an explicit completion." };
+        return;
+      }
       yield { type: "response.completed" };
     } finally {
       reader.releaseLock();
@@ -329213,6 +329251,7 @@ var ResponsesAdapter = class {
               const key3 = `${String(payload.item_id || "")}:${String(payload.summary_index || 0)}`;
               const delta = this.extractText(payload.delta);
               if (delta) {
+                emittedContent = true;
                 reasoningSummaries.set(key3, (reasoningSummaries.get(key3) || "") + delta);
                 yield { type: "reasoning.summary.delta", delta };
               }
@@ -329243,7 +329282,7 @@ var ResponsesAdapter = class {
                 calls.set(key3, {
                   id: String(item.call_id || item.id || key3),
                   name: String(item.name || ""),
-                  arguments: String(item.arguments || ""),
+                  argumentParts: item.arguments ? [String(item.arguments)] : [],
                   emitted: false
                 });
               }
@@ -329251,9 +329290,9 @@ var ResponsesAdapter = class {
             }
             if (eventType === "response.function_call_arguments.delta") {
               const key3 = String(payload.item_id || payload.call_id || payload.output_index || "");
-              const call = calls.get(key3) || { id: String(payload.call_id || key3), name: String(payload.name || ""), arguments: "", emitted: false };
+              const call = calls.get(key3) || { id: String(payload.call_id || key3), name: String(payload.name || ""), argumentParts: [], emitted: false };
               const delta = String(payload.delta || "");
-              call.arguments += delta;
+              if (delta) call.argumentParts.push(delta);
               calls.set(key3, call);
               yield { type: "tool_call.arguments.delta", id: call.id, delta };
               continue;
@@ -329265,19 +329304,20 @@ var ResponsesAdapter = class {
                 const call = calls.get(key3) || {
                   id: String(item.call_id || item.id || key3),
                   name: String(item.name || ""),
-                  arguments: String(item.arguments || ""),
+                  argumentParts: item.arguments ? [String(item.arguments)] : [],
                   emitted: false
                 };
                 call.id = String(item.call_id || call.id);
                 call.name = String(item.name || call.name);
-                call.arguments = typeof item.arguments === "string" ? item.arguments : call.arguments;
+                if (typeof item.arguments === "string" && item.arguments) call.argumentParts.push(item.arguments);
                 if (!call.emitted) {
                   call.emitted = true;
+                  const argumentsJson = assembleCompatibleToolArguments(call.argumentParts);
                   yield { type: "tool_call.started", id: call.id, name: call.name };
-                  if (call.arguments && call.arguments !== "{}") {
-                    yield { type: "tool_call.arguments.delta", id: call.id, delta: call.arguments };
+                  if (argumentsJson !== "{}") {
+                    yield { type: "tool_call.arguments.delta", id: call.id, delta: argumentsJson };
                   }
-                  yield { type: "tool_call.completed", id: call.id, name: call.name, arguments: call.arguments };
+                  yield { type: "tool_call.completed", id: call.id, name: call.name, arguments: argumentsJson };
                 }
                 calls.set(key3, call);
               }
@@ -329302,8 +329342,14 @@ var ResponsesAdapter = class {
       } else if (!completed) {
         yield { type: "response.failed", error: "[LLM Error] Responses stream ended before response.completed." };
       } else if (!emittedContent && calls.size === 0) {
-        yield { type: "response.failed", error: "[Error] Empty Responses stream." };
+        yield { type: "response.failed", error: "[Error] Provider returned an empty response." };
       } else {
+        for (const call of calls.values()) {
+          if (call.emitted) continue;
+          const argumentsJson = assembleCompatibleToolArguments(call.argumentParts);
+          yield { type: "tool_call.started", id: call.id, name: call.name };
+          yield { type: "tool_call.completed", id: call.id, name: call.name, arguments: argumentsJson };
+        }
         yield { type: "response.completed" };
       }
     } finally {
@@ -330182,7 +330228,7 @@ ${responsePath}
    * `provider_adapters_v2` context flag. Request serialization and SSE
    * normalization are delegated to the shared provider adapters while the
    * transport orchestration (loopback node-http, fetch -> node-http fallback,
-   * 120s/30s timeouts) and the 4xx Chat -> Responses downgrade stay here.
+   * cancellation-only streaming and the 4xx Chat -> Responses downgrade) stay here.
    * The emitted request body and StreamToken stream are byte-equivalent to
    * the legacy inlined path.
    */
@@ -330319,9 +330365,9 @@ ${responsePath}
   }
   /**
    * Loopback-aware transport injected into adapter `execute`. Streaming
-   * requests retain the fetch-to-node fallback for transport failures, while
-   * a local deadline is returned directly so one request cannot become a
-   * second Windows fallback request.
+   * requests retain the fetch-to-node fallback for transport failures. They
+   * have no response deadline; only caller cancellation or a concrete
+   * transport/provider failure may end the request.
    */
   buildProviderAdapterTransport() {
     return async (request, signal) => {
@@ -330331,7 +330377,7 @@ ${responsePath}
         const forwardAbort = () => abort.abort(signal?.reason);
         if (signal?.aborted) forwardAbort();
         else signal?.addEventListener("abort", forwardAbort, { once: true });
-        const effectiveTimeout = this.effectiveRequestTimeout(12e4);
+        const effectiveTimeout = 0;
         const timer = effectiveTimeout > 0 ? setTimeout(() => abort.abort(providerTimeoutError(effectiveTimeout)), effectiveTimeout) : void 0;
         try {
           try {
@@ -330476,7 +330522,7 @@ ${responsePath}
     const forwardAbort = () => abort.abort(signal?.reason);
     if (signal?.aborted) forwardAbort();
     else signal?.addEventListener("abort", forwardAbort, { once: true });
-    const effectiveTimeout = this.effectiveRequestTimeout(12e4);
+    const effectiveTimeout = 0;
     const timeout = effectiveTimeout > 0 ? setTimeout(() => abort.abort(providerTimeoutError(effectiveTimeout)), effectiveTimeout) : void 0;
     let reader = null;
     try {
@@ -330508,10 +330554,14 @@ ${responsePath}
       }
       const decoder = new TextDecoder();
       let buffer = "";
-      let currentToolCall = null;
+      const toolCalls = /* @__PURE__ */ new Map();
+      const toolCallOrder = [];
+      let syntheticToolIndex = 0;
+      let lastToolIndex = 0;
       let currentReasoningContent = "";
       let contentPolicyBlocked = false;
       let emittedContent = false;
+      let explicitCompletion = false;
       const streamSignal = signal || new AbortController().signal;
       while (true) {
         const { done, value } = await readProviderStreamChunk(reader, streamSignal);
@@ -330523,7 +330573,10 @@ ${responsePath}
           const trimmed = line.trim();
           if (!trimmed.startsWith("data: ")) continue;
           const data = trimmed.slice(6);
-          if (data === "[DONE]") continue;
+          if (data === "[DONE]") {
+            explicitCompletion = true;
+            continue;
+          }
           try {
             const json = JSON.parse(data);
             if (json.usage) yield { type: "usage", text: "", usage: extractProviderUsage(json) };
@@ -330541,24 +330594,44 @@ ${responsePath}
             }
             if (delta.tool_calls) {
               for (const tc of delta.tool_calls) {
-                if (tc.id) {
-                  if (currentToolCall) {
-                    yield { type: "tool_call", text: "", toolCall: currentToolCall, reasoningContent: currentReasoningContent || void 0 };
-                  }
-                  currentToolCall = { id: tc.id, name: tc.function?.name || "", arguments: tc.function?.arguments || "" };
-                } else if (tc.function?.arguments && currentToolCall) {
-                  currentToolCall.arguments += tc.function.arguments;
+                const rawIndex = Number(tc.index);
+                const index = Number.isInteger(rawIndex) && rawIndex >= 0 ? rawIndex : tc.id ? syntheticToolIndex++ : lastToolIndex;
+                lastToolIndex = index;
+                let call = toolCalls.get(index);
+                if (!call && (tc.id || tc.function?.name)) {
+                  call = { id: tc.id || "", name: tc.function?.name || "", argumentParts: [] };
+                  toolCalls.set(index, call);
+                  toolCallOrder.push(index);
                 }
+                if (!call) continue;
+                if (tc.id && !call.id) call.id = tc.id;
+                if (tc.function?.name && !call.name) call.name = tc.function.name;
+                if (tc.function?.arguments) call.argumentParts.push(tc.function.arguments);
               }
             }
           } catch {
           }
         }
       }
-      if (currentToolCall && currentToolCall.arguments) {
-        yield { type: "tool_call", text: "", toolCall: currentToolCall, reasoningContent: currentReasoningContent || void 0 };
+      if (toolCallOrder.length) {
+        for (const index of toolCallOrder) {
+          const call = toolCalls.get(index);
+          if (!call) continue;
+          yield {
+            type: "tool_call",
+            text: "",
+            toolCall: {
+              id: call.id,
+              name: call.name,
+              arguments: assembleCompatibleToolArguments(call.argumentParts)
+            },
+            reasoningContent: currentReasoningContent || void 0
+          };
+        }
       } else if (!emittedContent && contentPolicyBlocked) {
         yield { type: "text", text: "[Error] Content policy refusal (content_filter)." };
+      } else if (!explicitCompletion && !emittedContent && !currentReasoningContent) {
+        yield { type: "text", text: "[LLM Error] GitHub Models stream ended before an explicit completion." };
       }
     } finally {
       reader?.releaseLock();
@@ -340511,6 +340584,22 @@ function createToolchainCore() {
   return { registry: new ToolRegistry(), catalog: new CapabilityCatalog() };
 }
 
+// src/core/emptyResponseRetry.ts
+var EMPTY_RESPONSE_RETRY_DELAYS_MS = [200, 800, 2e3, 1e4, 6e4];
+var MAX_EMPTY_RESPONSE_RETRIES = EMPTY_RESPONSE_RETRY_DELAYS_MS.length;
+var MAX_CONSECUTIVE_EMPTY_RESPONSES = MAX_EMPTY_RESPONSE_RETRIES + 1;
+function emptyResponseRetryDelayMs(consecutiveEmptyResponses) {
+  return EMPTY_RESPONSE_RETRY_DELAYS_MS[Math.max(0, consecutiveEmptyResponses - 1)] ?? 0;
+}
+function observeEmptyResponseOutcome(consecutiveEmptyResponses, emptyResponse) {
+  const nextCount = emptyResponse ? consecutiveEmptyResponses + 1 : 0;
+  return {
+    consecutiveEmptyResponses: nextCount,
+    retry: emptyResponse && nextCount <= MAX_EMPTY_RESPONSE_RETRIES,
+    terminate: emptyResponse && nextCount > MAX_EMPTY_RESPONSE_RETRIES
+  };
+}
+
 // src/core/agentKernelRunner.ts
 var publicStreamFilters = /* @__PURE__ */ new WeakMap();
 var brokerOnlyAssistantBuffers = /* @__PURE__ */ new WeakMap();
@@ -340662,7 +340751,7 @@ function kernelTurnFailed(agent, turn) {
   return turn.stopReason === "error" || agent.isLlmErrorText(turn.text);
 }
 function providerTurnIsEmpty(turn) {
-  return /provider returned an empty response/i.test(`${turn.errorMessage}
+  return !turn.activity && /provider returned an empty response/i.test(`${turn.errorMessage}
 ${turn.text}`);
 }
 function removeTrailingFailedAssistant(agent, messages) {
@@ -340671,6 +340760,12 @@ function removeTrailingFailedAssistant(agent, messages) {
   const text = KernelMessageText(last);
   if (last.stopReason === "error" || agent.isLlmErrorText(text)) messages.pop();
 }
+function removeTrailingThoughtOnlyAssistant(messages) {
+  const last = messages[messages.length - 1];
+  if (last?.role !== "assistant") return;
+  const hasToolCall = last.content.some((content) => content.type === "toolCall");
+  if (!KernelMessageText(last).trim() && !hasToolCall) messages.pop();
+}
 function normalizePublicProviderError(error, secrets = []) {
   let raw = "";
   if (error instanceof Error) {
@@ -340796,8 +340891,17 @@ async function runAgentKernel(agent) {
   const tokens = [];
   const runOnce = async (promptMessages, appendPromptToAgentHistory) => {
     let lastAssistant = null;
+    let observedActivity = false;
+    let observedThought = false;
     const unsubscribe = kernel2.subscribe(async (event) => {
       await handleKernelEvent(agent, event, tokens);
+      if (event.type === "message_update") {
+        const delta = event.assistantMessageEvent;
+        const deltaText = typeof delta.delta === "string" ? delta.delta : "";
+        const thoughtDelta = delta.type === "thinking_delta" && !!deltaText.trim();
+        observedThought = observedThought || thoughtDelta;
+        observedActivity = observedActivity || thoughtDelta || delta.type === "text_delta" && !!deltaText.trim() || delta.type === "toolcall_end";
+      }
       if (event.type === "message_end" && event.message.role === "assistant") {
         lastAssistant = event.message;
       }
@@ -340815,11 +340919,13 @@ async function runAgentKernel(agent) {
       const assistant = lastAssistant;
       const text = assistant ? KernelMessageText(assistant) : "";
       const hasToolCall = !!assistant?.content?.some((content) => content.type === "toolCall");
-      const emptyResponse = !assistant || !text.trim() && !hasToolCall && String(assistant?.stopReason || "") !== "aborted";
+      const emptyResponse = !assistant || !text.trim() && !hasToolCall && !observedActivity && String(assistant?.stopReason || "") !== "aborted";
       return {
         text: emptyResponse ? "[Error] Provider returned an empty response." : text,
         stopReason: String(assistant?.stopReason || ""),
-        errorMessage: String(assistant?.errorMessage || (emptyResponse ? "Provider returned an empty response." : ""))
+        errorMessage: String(assistant?.errorMessage || (emptyResponse ? "Provider returned an empty response." : "")),
+        activity: observedActivity || !!text.trim() || hasToolCall,
+        thoughtOnly: observedThought && !text.trim() && !hasToolCall && !["error", "aborted"].includes(String(assistant?.stopReason || ""))
       };
     } finally {
       unsubscribe();
@@ -340857,16 +340963,30 @@ async function runAgentKernel(agent) {
       agent.recordWorkStatus("Final visual fallback used: local mini OCR plus conservative text correction.");
     }
     if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some((t3) => t3.text?.includes("[Model fallback]"))) {
-      tokens.unshift({ type: "text", text: `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.` });
+      const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
+      tokens.unshift({ type: "text", text: notice });
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: modelBeforeKernelRun, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
     }
-    let emptyResponseRetries = 0;
-    while (providerTurnIsEmpty(lastTurn) && emptyResponseRetries < 2) {
+    let consecutiveEmptyResponses = 0;
+    for (; ; ) {
+      const emptyResponseState = observeEmptyResponseOutcome(consecutiveEmptyResponses, providerTurnIsEmpty(lastTurn));
+      consecutiveEmptyResponses = emptyResponseState.consecutiveEmptyResponses;
+      if (lastTurn.thoughtOnly) {
+        removeTrailingThoughtOnlyAssistant(kernel2.state.messages);
+        lastTurn = await runWithCompressionResume([], false);
+        continue;
+      }
+      if (!emptyResponseState.retry) break;
       removeTrailingFailedAssistant(agent, kernel2.state.messages);
-      emptyResponseRetries += 1;
-      const notice = `[Model retry] Provider returned an empty response; retrying the same deployment (${emptyResponseRetries}/2).`;
+      const retryNumber = consecutiveEmptyResponses;
+      const notice = `[Model retry] Provider returned an empty response; retrying the same deployment (${retryNumber}/${MAX_EMPTY_RESPONSE_RETRIES}) after ${emptyResponseRetryDelayMs(consecutiveEmptyResponses)}ms.`;
       tokens.push({ type: "text", text: notice });
       agent.recordWorkStatus(notice);
-      await agent.waitForPlannedRouteRetry();
+      await agent.waitForPlannedRouteRetry(emptyResponseRetryDelayMs(consecutiveEmptyResponses));
       lastTurn = await runWithCompressionResume([], false);
     }
     let routeRetries = 0;
@@ -340878,6 +340998,11 @@ async function runAgentKernel(agent) {
       const notice = routeTransitionNotice(agent, previous);
       tokens.push({ type: "text", text: notice });
       agent.recordWorkStatus(notice);
+      agent.emitWorkEvent({
+        type: "status",
+        content: notice,
+        fallback: { from: previous, to: agent.model, providerId: agent.activeDeployment()?.providerId }
+      });
       kernel2.state.model = toKernelModel(agent);
       const fallbackToolSurface = refreshToolSurface(true);
       kernel2.state.systemPrompt = [agent.buildSystemPrompt(), fallbackToolSurface.systemPromptNotice].filter(Boolean).join("\n\n");
@@ -341057,7 +341182,7 @@ async function runAgentKernel(agent) {
             return;
           }
           if (textStarted) finalContent.push({ type: "text", text });
-          if (!finalContent.length) {
+          if (!finalContent.length && !thinking.trim()) {
             text = "[Error] Provider returned an empty response.";
             finalContent.push({ type: "text", text });
           }
@@ -343302,14 +343427,16 @@ var AutoRouter = class {
         retryDelayMs
       });
     }
+    if (failure.allowModelFallback === false) return attempts.slice(0, remainingAttempts);
     const selection = decision.requestedSelection;
     if (selection.kind === "fixed") return attempts.slice(0, remainingAttempts);
     const scope = selection.kind === "auto" ? selection.scope : { kind: "provider", providerId: current.providerId };
     const subset = selection.kind === "auto" ? selection.subset : void 0;
     const currentGroup = current.logicalModelGroupId;
+    const currentProviderId = current.providerId;
     const now2 = this.now();
     const attemptedDeployments = decision.attempts.map((attempt) => attempt.deployment);
-    const eligible = candidates.filter((candidate) => candidate.enabled && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
+    const eligible = candidates.filter((candidate) => candidate.enabled && candidate.deployment.providerId === currentProviderId && inScope(candidate.deployment, scope) && inSubset(candidate.deployment, subset) && !sameDeployment(candidate.deployment, current) && !attemptedDeployments.some((attempted) => sameDeployment(candidate.deployment, attempted)) && validationEligible(candidate, now2).length === 0 && this.passedInitialHardFilters(decision, candidate) && this.circuitState(candidate.deployment, now2, false) !== "open");
     const equivalent = currentGroup ? eligible.find((candidate) => candidate.deployment.logicalModelGroupId === currentGroup && !candidate.fallbackOnly) : void 0;
     const fallback = eligible.find((candidate) => candidate.fallbackOnly);
     const rankedAlternates = decision.rankedCandidates.map((ranked) => eligible.find((candidate) => sameDeployment(candidate.deployment, ranked.deployment))).filter((candidate) => !!candidate && candidate !== equivalent && candidate !== fallback && !candidate.fallbackOnly);
@@ -345639,7 +345766,12 @@ var Agent4 = class _Agent {
   beginRouteAttempt() {
     this.routeAttemptStartedAt = Date.now();
   }
-  async waitForPlannedRouteRetry() {
+  async waitForPlannedRouteRetry(explicitDelayMs) {
+    if (explicitDelayMs !== void 0) {
+      if (explicitDelayMs <= 0) return;
+      await new Promise((resolve16) => setTimeout(resolve16, explicitDelayMs));
+      return;
+    }
     const waitBudgetMs = Math.max(0, Math.min(15e3, this.lastRouteDecision?.retryBudgetMs ?? 5e3));
     const delay = Math.max(0, Math.min(waitBudgetMs, this.lastRouteRetryDelayMs));
     this.lastRouteRetryDelayMs = 0;
@@ -346747,7 +346879,8 @@ ${String(event.toolArgs || "")}`;
       sequence,
       status: input.status,
       guide: !isToolEvent && input.guide ? this.normalizeGuideReceipt(input.guide) : void 0,
-      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0
+      displayImage: isToolEvent ? this.hydrateDisplayImage(input.displayImage) : void 0,
+      fallback: input.fallback
     };
     if (activeRun && this.isPersistablePublicWorkEvent(event)) {
       activeRun.sequence = Number(sequence || activeRun.sequence + 1);
@@ -349666,9 +349799,15 @@ ${summary}`, segment, "local-summarize", true);
   }
   compressionBuildBlockStart(messages) {
     const activeRunId = this.currentWorkRunId();
-    if (!activeRunId) return 0;
-    const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
-    return index >= 0 ? index : 0;
+    if (activeRunId) {
+      const index = messages.findIndex((message) => String(message.run_id || message.runId || "") === activeRunId);
+      if (index >= 0) return index;
+    }
+    let lastRunBoundary = -1;
+    for (let index = 0; index < messages.length; index += 1) {
+      if (String(messages[index]?.run_id || messages[index]?.runId || "")) lastRunBoundary = index + 1;
+    }
+    return lastRunBoundary >= 0 ? lastRunBoundary : messages.length;
   }
   recentContextSuffix(messages, maxMessages, tokenBudget) {
     if (!messages.length) return [];
@@ -350597,17 +350736,12 @@ ${msg.content}
         currentAttempt.sideEffectBoundary = this.routeSideEffectCommitted;
         if (this.routeAttemptStartedAt > 0) currentAttempt.durationMs = Math.max(0, Date.now() - this.routeAttemptStartedAt);
       }
-      if (!fallbackEnabled) {
-        this.lastRouteDecision.finalStatus = "failed";
-        this.routeAttemptStartedAt = 0;
-        this.persistRouteDecision(this.lastRouteDecision);
-        return null;
-      }
       if (!this.pendingAutoAttempts.length) {
         this.pendingAutoAttempts = this.autoRouter.planAttempts(this.lastRouteDecision, this.autoRouteCandidates(), {
           error: failure,
           streamCommitted: this.routeStreamCommitted,
-          sideEffectCommitted: this.routeSideEffectCommitted
+          sideEffectCommitted: this.routeSideEffectCommitted,
+          allowModelFallback: fallbackEnabled
         });
       }
       const next2 = this.pendingAutoAttempts.shift();
@@ -350642,7 +350776,6 @@ ${msg.content}
     if (!next?.name) return null;
     this.model = next.name;
     this.fixedDeployment = this.deploymentRef(next);
-    if (next.provider) this.config.set("models", "auto_switch_anchor_provider", next.provider);
     return current;
   }
   isLlmErrorText(text) {
@@ -350650,12 +350783,10 @@ ${msg.content}
   }
   scopedSwitchModels(currentModelName) {
     const all = this.config.allModels();
-    if (this.config.autoSwitchScope() !== "provider") return all;
-    const current = currentModelName === "auto" ? void 0 : this.config.findModel(currentModelName);
-    const providerId = current?.provider_id || this.config.autoSwitchAnchorProvider() || this.config.findModel(this.config.getStr("models", "default_model"))?.provider_id || all[0]?.provider_id || "";
-    if (!providerId) return all;
-    const provider = this.config.findProvider(providerId);
-    return all.filter((m2) => m2.provider_id === (provider?.id || providerId));
+    const current = currentModelName === "auto" ? this.activeModelConfig() : this.config.findModel(currentModelName);
+    const providerId = current?.provider_id || this.fixedDeployment?.providerId || (currentModelName !== "auto" ? this.activeDeployment()?.providerId : void 0);
+    if (!providerId) return [];
+    return all.filter((m2) => m2.provider_id === providerId);
   }
   async validateModels(selectedNames, options = {}) {
     if (this.modelValidationPromise) return this.modelValidationPromise;
@@ -352037,7 +352168,10 @@ ${items.map((item) => this.formatAutomation(item)).join("\n")}`;
     if (!sa) return "[Subagent] Not found.";
     const requestedModel = this.normalizeSubagentModelSelection(sa.model && sa.model !== "default" ? sa.model : this.modelSelectionValue() || this.model);
     const requestedDeployment = parseDeploymentSelectionValue2(requestedModel);
-    const assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    let assignedModel = requestedModel === "auto" ? this.activeModelConfig() : requestedDeployment ? this.config.findDeployment(requestedDeployment) : this.config.findModel(requestedModel);
+    if (!assignedModel && !requestedDeployment && requestedModel !== "auto") {
+      assignedModel = this.activeModelConfig();
+    }
     const model = assignedModel?.name || (requestedModel === "auto" ? this.activeModelName() : requestedModel);
     const activeModel = this.activeModelConfig();
     const activeProvider = this.engineModel();
@@ -353466,6 +353600,32 @@ var ConversationKernel = class {
       followUp: queued?.followUp.slice() || []
     };
   }
+  /**
+   * dev-0.5.6: 排队 followUp 消息 drain 时转成普通用户消息载荷。
+   *
+   * 入队时（enqueueSameSession）为去重/编辑保留了 clientMessageId，并把
+   * runId 固定为入队时运行的 runId；若原样传给 process，写入 chatMessages
+   * 的用户消息会携带旧 runId + clientMessageId，渲染端（PC renderTranscript /
+   * 移动端 projectRemoteConversationItems）会把「clientMessageId + runId 命中
+   * workRuns」的消息误判为 guide 消息而跳过气泡，导致排队自动发送的消息不
+   * 显示在用户输入时间线。清除这两个字段后 process 走普通用户消息分支
+   * （chatMessages.push，mode 取 visibleMode，runId 取当前 run）。
+   */
+  drainQueuedFollowUpMessage(message) {
+    if (typeof message === "string") return message;
+    const text = String(message.text || "");
+    const images = message.images;
+    const attachments = message.attachments;
+    const visible = message.visibleUserInput;
+    const visibleMode = message.visibleMode;
+    return {
+      text,
+      ...images?.length ? { images } : {},
+      ...attachments?.length ? { attachments } : {},
+      ...visible ? { visibleUserInput: visible } : {},
+      ...visibleMode ? { visibleMode } : {}
+    };
+  }
   queueItems(target) {
     const runtime = this.findRuntime(target);
     if (!runtime) return [];
@@ -354207,7 +354367,7 @@ ${batchText}`,
           };
           lastTokens = await this.runSingle(runtime, batchMessage, "steer");
         } else {
-          lastTokens = await this.runSingle(runtime, next.message, next.queueMode);
+          lastTokens = await this.runSingle(runtime, this.drainQueuedFollowUpMessage(next.message), next.queueMode);
         }
       }
       const rootMessage = runtime.runner.subagents.readRootInbox()[0];
@@ -354666,7 +354826,11 @@ ${text}`;
       type: "queue_update",
       content: "Conversation queue updated.",
       conversationId: runtime.id,
-      queue: this.queued(runtime.target)
+      queue: this.queued(runtime.target),
+      // Structured rows with stable kernel ids so every consumer (PC UI and
+      // the paired mobile client) can render/edit/delete the same items.
+      queueItems: this.queueItems(runtime.target),
+      queuePaused: runtime.queuePaused === true
     });
   }
   clearQueued(runtime) {

package/dist/core/agent.d.ts

@@ -464,7 +464,7 @@ export declare class Agent {
     markRouteToolExecuted(name: string, rawArgs?: string): void;
     routeTransitionKind(): PlannedRouteAttempt['kind'] | '';
     beginRouteAttempt(): void;
-    waitForPlannedRouteRetry(): Promise<void>;
+    waitForPlannedRouteRetry(explicitDelayMs?: number): Promise<void>;
     recordRouteSuccess(latencyMs?: number, throughput?: number): void;
     recordRouteToolOutcome(valid: boolean): void;
     updateProviders(value: unknown): void;