newmark-agent 0.6.2 → 0.6.4

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

@@ -3123,7 +3123,7 @@ else {
             };
             const mutateConversationQueue = async (target, rawAction, input = {}) => {
                 const action = rawAction.replace(/^queue_/, '');
-                if (!['enqueue', 'update', 'delete', 'reorder', 'toggle_pause', 'set_pause', 'guide'].includes(action))
+                if (!['enqueue', 'update', 'delete', 'reorder', 'toggle_pause', 'set_pause', 'repair_blocked', 'guide'].includes(action))
                     throw new Error('Unknown queue action');
                 const key = activeFlowStateKey(target);
                 const flow = activeFlowStateFor(target);
@@ -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([