newmark-agent 0.6.2 → 0.6.3

package/dist/core/requestContextEstimate.d.ts

@@ -2,6 +2,10 @@
 export interface RequestContextEstimate {
     requestId: string;
     runId: string;
+    /** Immutable runtime branch that owned this model request. */
+    branchId: string;
+    /** Content hash of the actual submitted system/messages/tools payload. */
+    contextHash: string;
     model: string;
     at: string;
     estimatedTokens: number;
@@ -14,5 +18,5 @@ export interface RequestContextEstimate {
     inputTokens?: number;
     cacheReadTokens?: number;
 }
-export declare function estimateSubmittedContext(messages: Array<Record<string, unknown>>, systemPrompt: string, tools: unknown[], buildBlockStart: number): Omit<RequestContextEstimate, 'requestId' | 'runId' | 'model' | 'at'>;
+export declare function estimateSubmittedContext(messages: Array<Record<string, unknown>>, systemPrompt: string, tools: unknown[], buildBlockStart: number): Omit<RequestContextEstimate, 'requestId' | 'runId' | 'model' | 'at' | 'branchId' | 'contextHash'>;
 //# sourceMappingURL=requestContextEstimate.d.ts.map

package/dist/llm/provider.js

@@ -1311,12 +1311,19 @@ class LLMProvider {
     async *chatStreamWithTools(model, messages, systemPrompt, temperature, maxTokens, tools, signal, reasoningTier, sessionId) {
         if (signal?.aborted)
             throw abortFailure(signal);
+        // A non-positive value means "provider owns the output limit". Protocols
+        // whose wire format requires a positive max_tokens (Anthropic, GitHub
+        // Models) fall back to the selected reasoning tier's budget; OpenAI-style
+        // adapters omit the field entirely so the endpoint decides.
+        const requiredMaxTokens = maxTokens > 0
+            ? maxTokens
+            : this.intelligenceConfig(reasoningTier || 'medium').maxTokens;
         if (this.protocol() === 'anthropic') {
-            yield* this.anthropicChatWithTools(model, messages, systemPrompt, temperature, maxTokens, tools, signal);
+            yield* this.anthropicChatWithTools(model, messages, systemPrompt, temperature, requiredMaxTokens, tools, signal);
             return;
         }
         if (this.protocol() === 'github_models') {
-            yield* this.githubModelsChatStreamWithTools(model, messages, systemPrompt, temperature, maxTokens, tools, signal);
+            yield* this.githubModelsChatStreamWithTools(model, messages, systemPrompt, temperature, requiredMaxTokens, tools, signal);
             return;
         }
         if (this.useProviderAdaptersV2) {

package/dist/main.js

@@ -3292,6 +3292,10 @@ else {
                     // Other conversations' paused Flows are never touched by this send.
                     const flow = activeFlowStateFor(target);
                     const snapshot = await runtimeSnapshotForTarget(target);
+                    // The renderer's visible composer selection is authoritative for this
+                    // send. It is provider-qualified (deployment:providerId:modelId), so it
+                    // can never fall back to a same-named model on another provider.
+                    const requestedModelSelection = String(requested.model || '').trim();
                     const structured = typeof message === 'string' ? { text: message } : { ...message };
                     const requestedMode = String(requested.requestedMode || structured.visibleMode || snapshot.mode || 'build');
                     const inputMode = String(requested.inputMode || snapshot.inputMode || 'guide') === 'next' ? 'next' : 'guide';
@@ -3319,9 +3323,18 @@ else {
                             });
                             return { ok: receipt.status !== 'rejected', accepted: receipt.status !== 'rejected', receipt, ...await runtimeSnapshotForTarget(target) };
                         }
+                        const requestedBranchPath = Array.isArray(structured.branchPath)
+                            ? (structured.branchPath || []).map(id => String(id || '')).filter(Boolean)
+                            : String(structured.branchPath || '').split('>').map(id => id.trim()).filter(Boolean);
+                        const requestedBranchNodeId = String(structured.branchNodeId
+                            || requestedBranchPath[requestedBranchPath.length - 1]
+                            || '');
                         return await applyConversationAction(target, 'queue_enqueue', '', {
                             id, text: structured.visibleUserInput || structured.text, requestedMode: requestedMode === 'flow' ? 'build' : requestedMode,
                             goalObjective: objective, images: structured.images, createdAt: new Date().toISOString(),
+                            branchNodeId: requestedBranchNodeId || undefined,
+                            branchPath: requestedBranchPath.length ? requestedBranchPath : undefined,
+                            modelSelection: requestedModelSelection || undefined,
                         });
                     }
                     await selectConversationMode(target, requestedMode);
@@ -3335,10 +3348,10 @@ else {
                     // 发送命令时锁定输入框选择的模型：接受命令后的整个运行过程都以该
                     // 模型为准（唯一例外是显式的不可用回退，且回退会以结构化事件同步
                     // 到前端输入框下方的选择区，而不是隐藏的参数回退）。
-                    const requestedModel = String(snapshot.model || agent.model);
+                    const requestedModel = String(requestedModelSelection || snapshot.model || agent.model);
                     const options = {
                         mode: requestedMode === 'goal' ? 'goal' : requestedMode,
-                        model: String(snapshot.model || agent.ensureUsableModelSelection()),
+                        model: String(requestedModelSelection || snapshot.model || agent.ensureUsableModelSelection()),
                         intelligence: String(snapshot.intelligence || agent.intelligence),
                         inputMode: inputMode,
                         engine: agent.engine,
@@ -3460,6 +3473,16 @@ else {
                     settlement: new Promise(resolve => { resolvePendingSettlement = resolve; }),
                 };
                 pendingFlowStarts.set(flowKey, pending);
+                // Admit the Flow build at command-acceptance time. It may wait behind a
+                // running Build, but its queue position fixes the parent of every later
+                // Next item; the actual Flow start only claims this existing build.
+                const flowCommandId = `flow-${flowKey}-${(0, crypto_1.randomUUID)()}`;
+                const flowBuildId = ensureConversationKernel(root).admitExternalBuild(flowTarget, {
+                    commandId: flowCommandId,
+                    text: String(input || ''),
+                    requestedMode: 'flow',
+                    modelSelection: agent.modelSelectionValue(),
+                });
                 let previousQueuePaused = false;
                 try {
                     let state = await runtimeSnapshotForTarget(flowTarget);
@@ -3485,7 +3508,13 @@ else {
                 }
                 const flowAgent = ensureConversationKernel(root).beginExternalRun(flowTarget, {
                     mode: 'flow', model: agent.model, intelligence: agent.intelligence, inputMode: agent.inputMode, engine: agent.engine,
-                }, () => isolatedConversationAgent(flowTarget), previousQueuePaused);
+                }, () => isolatedConversationAgent(flowTarget), previousQueuePaused, {
+                    commandId: flowCommandId,
+                    text: String(input || ''),
+                    requestedMode: 'flow',
+                    modelSelection: agent.modelSelectionValue(),
+                    existingBuildId: flowBuildId || undefined,
+                });
                 mainConversationOwners.add(flowKey);
                 conversationSelections.set(flowKey, { ...conversationSelections.get(flowKey), mode: 'flow' });
                 const previousMode = flowAgent.mode;
@@ -3655,7 +3684,13 @@ else {
                 const flowAgent = ensureConversationKernel(root).beginExternalRun(flowTarget, {
                     mode: 'flow', model: suspension.flowAgent?.model || agent.model, intelligence: suspension.flowAgent?.intelligence || agent.intelligence,
                     inputMode: suspension.flowAgent?.inputMode || agent.inputMode, engine: agent.engine,
-                }, () => suspension.flowAgent || isolatedConversationAgent(flowTarget), suspension.queueWasPaused ?? true);
+                }, () => suspension.flowAgent || isolatedConversationAgent(flowTarget), suspension.queueWasPaused ?? true, {
+                    commandId: `flow-resume-${flowKey}-${(0, crypto_1.randomUUID)()}`,
+                    text: String(response || ''),
+                    requestedMode: 'flow',
+                    modelSelection: agent.modelSelectionValue(),
+                    resume: true,
+                });
                 mainConversationOwners.add(flowKey);
                 suspension.flowAgent = flowAgent;
                 // A previous Flow may have been interrupted just before its isolated

package/dist/providers/chat-completions.adapter.js

@@ -44,8 +44,13 @@ class ChatCompletionsAdapter {
             model: request.model,
             messages,
             temperature: request.temperature,
-            max_tokens: request.maxOutputTokens,
         };
+        // Omit the cap entirely for provider-owned output. Only auxiliary
+        // requests that explicitly need a bounded completion send a positive
+        // max_tokens value.
+        if (Number.isFinite(request.maxOutputTokens) && request.maxOutputTokens > 0) {
+            body.max_tokens = request.maxOutputTokens;
+        }
         if (tools.length) {
             body.tools = tools;
             body.tool_choice = 'auto';

package/dist/providers/responses.adapter.js

@@ -69,8 +69,11 @@ class ResponsesAdapter {
             model: request.model,
             input: input.length ? input : [{ role: 'user', content: '' }],
             temperature: request.temperature,
-            max_output_tokens: request.maxOutputTokens,
         };
+        // Omit the cap entirely for provider-owned output.
+        if (Number.isFinite(request.maxOutputTokens) && request.maxOutputTokens > 0) {
+            body.max_output_tokens = request.maxOutputTokens;
+        }
         if (request.reasoningEffort)
             body.reasoning = { effort: request.reasoningEffort, summary: 'auto' };
         if (request.systemPrompt)

package/dist/server.d.ts

@@ -32,6 +32,7 @@ export interface HostedMobileServerOptions {
         clientMessageId?: string;
         flowName?: string;
         flowStart?: number;
+        model?: string;
     }) => Promise<Record<string, unknown>>;
 }
 export interface HostedMobileServerStatus {

package/dist/server.js

@@ -153,6 +153,7 @@ function hostedCommandOptions(params) {
         clientMessageId: String(params.clientMessageId || ''),
         flowName: String(params.flowName || params.defaultFlow || ''),
         flowStart: typeof params.flowStart === 'number' && Number.isFinite(params.flowStart) ? params.flowStart : undefined,
+        model: String(params.model || ''),
     };
 }
 const MOBILE_EDITABLE_EXTENSIONS = new Set([

package/dist/ui/index.html

@@ -7143,6 +7143,10 @@ var state = {
   _conversationLoadingMsgEl: null,
   _syncedMode: null,
   _syncedModel: null,
+  // Monotonic revision for explicit renderer model changes. Activation
+  // snapshots may restore the conversation's remembered model only when no
+  // newer local selection was made while the snapshot was in flight.
+  modelSelectionRevision: 0,
   _syncedInputMode: null,
   _editingQueueIndex: -1,
   _editingGoalText: '',
@@ -9353,7 +9357,7 @@ var api = window.api || {
         conversationId: target.conversationId,
         action: action
       };
-      ['id', 'text', 'value', 'requestedMode', 'goalObjective', 'createdAt', 'images', 'paused', 'orderedIds'].forEach(function(field) {
+      ['id', 'text', 'value', 'requestedMode', 'goalObjective', 'createdAt', 'images', 'paused', 'orderedIds', 'branchNodeId', 'branchPath'].forEach(function(field) {
         if (input && Object.prototype.hasOwnProperty.call(input, field)) body[field] = input[field];
       });
       var r = await fetch('/api/queue-action', {
@@ -13287,6 +13291,35 @@ function hydrateConversationBranchState(snapshot) {
   rememberRuntimeConversationBranchPath(currentConversationTarget(s.conversationId || activeConversationId()));
 }
 
+// Apply the conversation's remembered provider-qualified model to the
+// composer selector. The backend snapshot is authoritative; a newer explicit
+// user selection made while the snapshot was in flight wins via the revision
+// guard passed by the activation caller.
+function applyConversationModelSelection(snapshot) {
+  var nextModel = String(snapshot && snapshot.model || '').trim();
+  if (!nextModel) return;
+  state.model = nextModel;
+  if (typeof snapshot.modelLabel === 'string' && snapshot.modelLabel) state.modelLabel = snapshot.modelLabel;
+  state._syncedModel = nextModel;
+  var sel = els['model-select'] || document.getElementById('model-select');
+  if (!sel) {
+    if (window.syncModelSelectSurface) window.syncModelSelectSurface();
+    return;
+  }
+  var hasOption = Array.prototype.some.call(sel.options, function(option) {
+    return String(option.value || '') === nextModel;
+  });
+  if (hasOption) {
+    sel.value = nextModel;
+    window.syncModelSelectSurface();
+  } else if (window.refreshModelSelect) {
+    // The remembered deployment is not in the current enabled catalog (for
+    // example its provider was explicitly disabled). Reuse the existing
+    // visibility fallback instead of displaying a model that cannot run.
+    window.refreshModelSelect();
+  }
+}
+
 function applyConversationSnapshot(s, requestedConversationId, options) {
   if (!s) return;
   if (s.computerUse) {
@@ -13309,6 +13342,11 @@ function applyConversationSnapshot(s, requestedConversationId, options) {
     if (s && s.runtimeKey) registerRuntimeKey(snapshotTarget, s.runtimeKey);
     hydrateConversationDraftSnapshot(s, snapshotTarget);
     hydrateConversationBranchState(s);
+    var modelRevisionMatches = !options || options.modelRevision === undefined
+      || Number(state.modelSelectionRevision || 0) === Number(options.modelRevision || 0);
+    if (modelRevisionMatches && isActiveConversationTarget(snapshotTarget)) {
+      applyConversationModelSelection(s);
+    }
     rebindQueueToRuntimeBranch(snapshotTarget);
     if (s && Array.isArray(s.workRuns) && !preserveReadableHistory) {
       var viewedBranchIdForSync = Array.isArray(s.viewedBranchNodePath) && s.viewedBranchNodePath.length
@@ -14443,6 +14481,7 @@ function applyModelFallbackToInputSurface(event) {
   var nextValue = modelDeploymentValueForName(f.to, f.providerId || '');
   if (!nextValue || String(state.model || '') === nextValue) return;
   state.model = nextValue;
+  state.modelSelectionRevision = Number(state.modelSelectionRevision || 0) + 1;
   state.modelLabel = '';
   // 不预置 _syncedModel：下一次发送前的 syncConversationExecutionState 会把
   // 回退模型正式同步到后端 host agent（并持久化为该对话的首选模型）。
@@ -17208,6 +17247,11 @@ function bindBackendQueuedRequestToTarget(text, target, occurrence) {
     request.createdAt = queueItem.createdAt;
     request.runId = queueItem.runId;
     request.images = (queueItem.images || []).map(function(image) { return Object.assign({}, image); });
+    if (Array.isArray(queueItem.branchPath) && queueItem.branchPath.length) {
+      request.branchPath = queueItem.branchPath.join('>');
+    } else if (queueItem.branchNodeId) {
+      request.branchPath = String(queueItem.branchNodeId);
+    }
   }
   return request;
 }
@@ -17920,7 +17964,15 @@ window.sendMessage = async function(modeOverride, queuedText, opts) {
       queuedModeRequest.goalDeclaration = true;
       queuedModeRequest.goalObjective = rawText;
     }
-    var queueResult = await api.sendMessage({ text: rawText, images: queuedModeRequest.images }, lockedTarget, {
+    var queuedBranchPath = Array.isArray(queuedModeRequest.branchPath)
+      ? queuedModeRequest.branchPath
+      : String(queuedModeRequest.branchPath || '').split('>').filter(Boolean);
+    var queueResult = await api.sendMessage({
+      text: rawText,
+      images: queuedModeRequest.images,
+      branchPath: queuedBranchPath,
+      branchNodeId: String(queuedBranchPath[queuedBranchPath.length - 1] || '')
+    }, lockedTarget, {
       clientMessageId: queuedModeRequest.clientMessageId,
       inputMode: 'next',
       requestedMode: requestedMode,
@@ -18117,12 +18169,6 @@ window.sendMessage = async function(modeOverride, queuedText, opts) {
   updateSubmitButtonState();
   var executionMode = opts.goalDeclaration || requestedMode === 'goal' ? 'build'
     : ((opts.forceBuild || idleBuildNextImmediate) && requestedMode === 'build' ? 'build' : requestedMode);
-  // The command freezes its mode and delivery policy; selector persistence
-  // must not rewrite the mode of an already running target before acceptance.
-  if (api.setModel && state.model !== state._syncedModel) {
-    var selectedModelResult = await api.setModel(state.model);
-    if (selectedModelResult != null) state._syncedModel = state.model;
-  }
   if (consumeProvisionalStop(lockedTarget, localRunId)) {
     // Stop/Escape won before the backend accepted the prompt. Do not send the
     // provider request after the user has already cancelled it.
@@ -18166,10 +18212,20 @@ window.sendMessage = async function(modeOverride, queuedText, opts) {
   try {
     var activationBeforeSend = pendingConversationActivation(lockedTarget);
     if (activationBeforeSend) await activationBeforeSend;
+    // Bind the model actually shown in the composer to the exact target
+    // conversation after its activation snapshot restored the remembered
+    // selection. This makes the visible provider/model authoritative for the
+    // send and prevents a stale renderer selection from reaching another
+    // conversation's remembered deployment.
+    if (api.setModel && state.model && state.model !== state._syncedModel) {
+      var selectedModelResult = await api.setModel(state.model);
+      if (selectedModelResult != null) state._syncedModel = state.model;
+    }
     var sendPromise = api.sendMessage(requestMessage, lockedTarget, {
       requestedMode: requestedMode, inputMode: effectiveInputMode,
       goalObjective: opts.goalObjective || '',
-      clientMessageId: String(requestMessage && requestMessage.clientMessageId || localRunId)
+      clientMessageId: String(requestMessage && requestMessage.clientMessageId || localRunId),
+      model: state.model
     });
     if (requestedMode === 'goal') activateSubmittedGoal(opts.goalObjective || rawText);
     var r = await sendPromise;
@@ -23965,6 +24021,7 @@ function syncBackendConversation() {
   }
   if (conv && api.ensureConversation) {
     var requestedTarget = currentConversationTarget(conv.id);
+    var modelRevisionAtActivation = Number(state.modelSelectionRevision || 0);
     state.activeBackendConversationId = conv.id;
     var historyPromise = loadActiveConversationMessages(conv.id);
     requestedLoadGeneration = state.conversationLoadGeneration;
@@ -23973,7 +24030,7 @@ function syncBackendConversation() {
       var id = (s && s.conversationId) || conv.id;
       if (!requestIsCurrent() || String(id) !== requestedConversationId) return;
       state.activeBackendConversationId = id;
-      applyConversationSnapshot(s, id, { preserveReadableHistory: true });
+      applyConversationSnapshot(s, id, { preserveReadableHistory: true, modelRevision: modelRevisionAtActivation });
     }).catch(function(error) {
       if (!requestIsCurrent()) return;
       showUiNotice(
@@ -24505,12 +24562,13 @@ window.newConversation = function(workspaceReference, branchCommunication) {
   renderConversations();
   els['chat-area'].innerHTML = '';
   var target = currentConversationTarget(id);
+  var modelRevisionAtActivation = Number(state.modelSelectionRevision || 0);
   var activation = api.activateConversation ? api.activateConversation(target) : (api.ensureConversation ? api.ensureConversation(target) : Promise.resolve(null));
   var activationKey = runtimeKeyFor(target.workspaceId, target.conversationId);
   var activationReady = activation.then(function(s) {
     if (createdWorkspaceKey !== currentWorkspaceKey() || id !== String(activeConversationId() || 'default')) return;
     state.activeBackendConversationId = String((s && s.conversationId) || id);
-    if (s) applyConversationSnapshot(s, id, { preserveReadableHistory: true });
+    if (s) applyConversationSnapshot(s, id, { preserveReadableHistory: true, modelRevision: modelRevisionAtActivation });
     if (branchCommunication && api.setConversationBranchCommunication) {
       api.setConversationBranchCommunication(target, true).then(function() {
         var convsNow = currentWorkspaceConversations();
@@ -28307,6 +28365,7 @@ function schedulePostStartupUiRendering() {
       var selectedContextGeneration = state.conversationLoadGeneration;
       var selectedContextModel = this.value;
       state.model = selectedContextModel;
+      state.modelSelectionRevision = Number(state.modelSelectionRevision || 0) + 1;
       window.syncModelSelectSurface();
       if (state.model !== state._syncedModel) {
         var modelResult = await api.setModel(selectedContextModel);