newmark-agent 0.4.3 → 0.4.4

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

@@ -336405,7 +336405,7 @@ var ToolExecutor = class {
       t3("subagent_result", "Return the persisted transcript, mailbox summary, status, and latest result for a peer agent. Target by exact id (preferred) or name.", { id: { type: "string", description: "Exact peer id from subagent_list." }, name: { type: "string", description: "Convenience peer name." } }, []),
       t3("subagent_close", "Close a same-conversation peer. Root can close any peer; a peer can close only itself. Target by exact id (preferred) or name.", { id: { type: "string", description: "Exact peer id from subagent_list." }, name: { type: "string", description: "Convenience peer name." } }, []),
       t3("linked_plan", "Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.", { action: { type: "string", enum: ["get", "update"] }, markdown: { type: "string" }, expected_revision: { type: "number" } }, ["action"]),
-      t3("build_history_query", "Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." }, max_chars: { type: "number", minimum: 100, maximum: 4e3, description: "Per-event/per-guide content character bound; defaults to 2000." } }, []),
+      t3("build_history_query", "Read the concrete public work details (tool calls, results, file changes, guides) of one historical Build Block. Call it proactively when the current task continues, fixes, verifies, or depends on earlier work: reuse the returned activity instead of re-investigating from scratch. Do not call it merely to answer completion status already exposed by the prompt. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." }, max_chars: { type: "number", minimum: 100, maximum: 4e3, description: "Per-event/per-guide content character bound; defaults to 2000." } }, []),
       t3("context_compress", "Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.", { keep_recent: { type: "number", minimum: 2, maximum: 60, description: "Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages." }, force: { type: "boolean", description: "Compress even if the context is not yet over the automatic threshold. Defaults to false." } }, []),
       t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. The hot cache stays bounded; evicted folded segments remain in a conversation-isolated append-only cold archive and are loaded only by explicit search/read/restore calls. Actions: list returns a bounded index of current context entries; remove declares one long-term entry for unload (see below); summarize folds a contiguous current range; restore reinserts a folded segment when its summary marker is still present; search finds matching hot or archived segments; read returns one bounded segment without injecting the whole archive; status reports budgets, hot cache, cold archive, the protected recent zone, and pending removals. The recent context tail and last user message are protected from remove/summarize unless dangerous is true. For cache-optimization, remove ONLY targets long-term history (never the protected recent tail or last user message) and does NOT unload immediately: the declared entry stays in context for the rest of the current Build Block so the provider prefix cache stays stable, then is physically removed when the Block ends \u2014 applying to subsequent Blocks only.", {
         action: { type: "string", enum: ["list", "remove", "summarize", "restore", "search", "read", "status"], description: "list current entries; remove one; summarize a range; restore by restore_id; search hot/cold folded segments; read one bounded folded segment; status report context budgets and storage." },
@@ -340900,7 +340900,7 @@ function buildConversationTaskLedger(agent) {
     "Unfinished Continuation Queue (newest to oldest; summary fields only; use only when the current user instruction authorizes continuation and the task is relevant):",
     ...unfinishedLines.length ? unfinishedLines : ["(none)"],
     ...unfinished.length > unfinishedLines.length ? [`(${unfinished.length - unfinishedLines.length} older unfinished run(s) omitted from the bounded prompt ledger.)`] : [],
-    "When concrete work details are required, call build_history_query with history_index from this list. Do not call it merely to answer completion status already shown here."
+    "When the current task continues, fixes, verifies, or depends on earlier work in this list, proactively call build_history_query with its history_index before re-investigating. Reuse the returned tool activity and results instead of re-running commands or re-reading files this conversation already examined. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown here."
   ].join("\n");
 }
 async function shouldStopAfterTurn(agent, message) {
@@ -349168,8 +349168,13 @@ ${summary}`, segment, "local-summarize", true);
     };
   }
   resolveWindowModel(modelName) {
-    if (modelName !== "auto") return this.config.findModel(modelName);
-    return this.activeModelConfig() || this.config.findModel(this.config.getStr("models", "default_model"));
+    if (modelName === "auto" || modelName === this.model || modelName === this.activeModelName()) {
+      const active = this.activeModelConfig();
+      if (active) return active;
+    }
+    const byName = this.config.findModel(modelName);
+    if (byName) return byName;
+    return this.config.findModel(this.config.getStr("models", "default_model"));
   }
   contextMaxTokens(modelName = this.model) {
     const model = this.resolveWindowModel(modelName);
@@ -352333,7 +352338,7 @@ ${custom}`);
       "- Memory Lab is governed by an explicit Policy chain: pre-think whether memory is needed; prefer bounded memory_lab_query retrieval; then choose ADD/UPDATE/DELETE only when the user authorizes durable memory mutation.",
       "- Before memory_lab_update, inspect the target with memory_lab_query or memory_lab_read. For an existing component pass expectedUpdatedAt so concurrent/stale writes fail closed; preserve established tag parent paths.",
       "- Use memory_lab_delete only for an explicit user request to forget/remove memory. Prior revisions are retained under Memory Lab/archive and mutation decisions are appended to policy.jsonl for replay.",
-      "- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. Use build_history_query only when the current user asks what specifically happened in one Build Block; querying history is read-only and never authorizes resuming that work.",
+      "- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. When the current task continues, fixes, verifies, or depends on earlier Build Blocks, proactively call build_history_query to read the concrete tool activity and results of the relevant block, and reuse that information instead of re-investigating (re-running commands or re-reading files) from scratch. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown in the prompt.",
       "- Linked plan disclosure: a durable conversation-linked Markdown plan exists and can be inspected or updated with linked_plan when explicitly needed or required by Plan mode. Its full Markdown and revision are not injected into every model request.",
       "- A memory_lab_update, memory_lab_delete, or memory_lab_reindex call is unfinished until its awaited tool result contains rebuildReceipt.completed=true. The completion receipt is represented by the tool activity inside the current Build block and should not be repeated as a separate completion message.",
       `- Skills and subagents: skill searches enabled metadata and loads one SKILL.md body on demand; skill_download installs offline skill folders; task creates constrained subagents tracked in agent state.`,
@@ -353131,6 +353136,18 @@ var ConversationKernel = class {
     if (runtime) runtime.options.mode = mode;
     return runner.mode;
   }
+  setModel(target, model) {
+    const normalized = this.normalizeTarget(target);
+    const runtime = this.findRuntime(normalized);
+    const runner = runtime?.runner || this.createRunner(normalized);
+    if (!runtime || !runtime.activePromise) {
+      runner.setModel(model);
+    } else {
+      runtime.options.model = model;
+    }
+    runner.saveWorkspaceConversationState(true);
+    return runner.model;
+  }
   async toggleGoalPause(target) {
     const normalized = this.normalizeTarget(target);
     let runtime = this.findRuntime(normalized);
@@ -353245,14 +353262,15 @@ var ConversationKernel = class {
   }
   async prompt(message, target, options, queueMode = "followUp") {
     const normalized = this.normalizeTarget(target);
+    const active = this.findRuntime(normalized);
+    if (active?.activePromise) {
+      this.enqueueSameSession(active, message, queueMode);
+      this.activateAcceptedGoal(active, typeof message === "string" ? "" : message.goalObjective);
+      return active.activePromise;
+    }
     const runtime = this.runtime(normalized, options);
     runtime.options = { ...options };
     this.applyOptions(runtime.runner, options);
-    if (runtime.activePromise) {
-      this.enqueueSameSession(runtime, message, queueMode);
-      this.activateAcceptedGoal(runtime, typeof message === "string" ? "" : message.goalObjective);
-      return runtime.activePromise;
-    }
     runtime.generation = (this.generations.get(runtime.runtimeKey) || runtime.generation || 0) + 1;
     this.generations.set(runtime.runtimeKey, runtime.generation);
     const requestedRunId = typeof message === "string" ? "" : String(message.runId || "").trim().slice(0, 200);
@@ -353397,7 +353415,21 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.mirrorHostIfTargetActive(runtime);
     return this.result(runtime, lastTokens);
   }
+  /**
+   * Apply a model selection recorded while a Build block was running. The
+   * in-flight block never switches mid-block; the switch takes effect the next
+   * time a queued Guide/Next re-enters the block, and only when the pending
+   * selection actually differs from the runner's current selection.
+   */
+  syncPendingModel(runtime) {
+    const pending3 = String(runtime.options.model || "").trim();
+    if (!pending3) return;
+    if (pending3 === runtime.runner.model || pending3 === runtime.runner.modelSelectionValue()) return;
+    runtime.runner.setModel(pending3);
+    runtime.options.model = runtime.runner.modelSelectionValue();
+  }
   async runSingle(runtime, message, continuationMode) {
+    this.syncPendingModel(runtime);
     this.consumeQueuedMessage(runtime, typeof message === "string" ? message : message.text);
     const timeoutMs = this.processTimeoutMs(runtime);
     if (timeoutMs <= 0) {
@@ -354038,6 +354070,9 @@ async function handle(request) {
   if (request.method === "set_mode") {
     return kernel.setMode(checkedTarget(request.params.target), request.params.mode);
   }
+  if (request.method === "set_model") {
+    return kernel.setModel(checkedTarget(request.params.target), request.params.model);
+  }
   if (request.method === "set_input_mode") {
     return kernel.setInputMode(checkedTarget(request.params.target), request.params.mode);
   }

package/dist/conversation-utility-host.js

@@ -93,6 +93,9 @@ async function handle(request) {
     if (request.method === 'set_mode') {
         return kernel.setMode(checkedTarget(request.params.target), request.params.mode);
     }
+    if (request.method === 'set_model') {
+        return kernel.setModel(checkedTarget(request.params.target), request.params.model);
+    }
     if (request.method === 'set_input_mode') {
         return kernel.setInputMode(checkedTarget(request.params.target), request.params.mode);
     }

package/dist/core/agent.js

@@ -5305,9 +5305,21 @@ class Agent {
         };
     }
     resolveWindowModel(modelName) {
-        if (modelName !== 'auto')
-            return this.config.findModel(modelName);
-        return this.activeModelConfig() || this.config.findModel(this.config.getStr('models', 'default_model'));
+        // The context window (display ring, inspector, and compaction trigger) must
+        // resolve the deployment that is actually running. For the active selection
+        // — auto or a fixed model — resolve through the active deployment so a
+        // qualified selection or two same-named models across providers never fall
+        // through to the 128000 default. Only a caller-supplied foreign name (for
+        // example a validation probe against another model) resolves by bare name.
+        if (modelName === 'auto' || modelName === this.model || modelName === this.activeModelName()) {
+            const active = this.activeModelConfig();
+            if (active)
+                return active;
+        }
+        const byName = this.config.findModel(modelName);
+        if (byName)
+            return byName;
+        return this.config.findModel(this.config.getStr('models', 'default_model'));
     }
     contextMaxTokens(modelName = this.model) {
         const model = this.resolveWindowModel(modelName);
@@ -8858,7 +8870,7 @@ class Agent {
             '- Memory Lab is governed by an explicit Policy chain: pre-think whether memory is needed; prefer bounded memory_lab_query retrieval; then choose ADD/UPDATE/DELETE only when the user authorizes durable memory mutation.',
             '- Before memory_lab_update, inspect the target with memory_lab_query or memory_lab_read. For an existing component pass expectedUpdatedAt so concurrent/stale writes fail closed; preserve established tag parent paths.',
             '- Use memory_lab_delete only for an explicit user request to forget/remove memory. Prior revisions are retained under Memory Lab/archive and mutation decisions are appended to policy.jsonl for replay.',
-            '- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. Use build_history_query only when the current user asks what specifically happened in one Build Block; querying history is read-only and never authorizes resuming that work.',
+            '- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. When the current task continues, fixes, verifies, or depends on earlier Build Blocks, proactively call build_history_query to read the concrete tool activity and results of the relevant block, and reuse that information instead of re-investigating (re-running commands or re-reading files) from scratch. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown in the prompt.',
             '- Linked plan disclosure: a durable conversation-linked Markdown plan exists and can be inspected or updated with linked_plan when explicitly needed or required by Plan mode. Its full Markdown and revision are not injected into every model request.',
             '- A memory_lab_update, memory_lab_delete, or memory_lab_reindex call is unfinished until its awaited tool result contains rebuildReceipt.completed=true. The completion receipt is represented by the tool activity inside the current Build block and should not be repeated as a separate completion message.',
             `- Skills and subagents: skill searches enabled metadata and loads one SKILL.md body on demand; skill_download installs offline skill folders; task creates constrained subagents tracked in agent state.`,

package/dist/core/agentKernelRunner.js

@@ -818,7 +818,7 @@ function buildConversationTaskLedger(agent) {
         'Unfinished Continuation Queue (newest to oldest; summary fields only; use only when the current user instruction authorizes continuation and the task is relevant):',
         ...(unfinishedLines.length ? unfinishedLines : ['(none)']),
         ...(unfinished.length > unfinishedLines.length ? [`(${unfinished.length - unfinishedLines.length} older unfinished run(s) omitted from the bounded prompt ledger.)`] : []),
-        'When concrete work details are required, call build_history_query with history_index from this list. Do not call it merely to answer completion status already shown here.',
+        'When the current task continues, fixes, verifies, or depends on earlier work in this list, proactively call build_history_query with its history_index before re-investigating. Reuse the returned tool activity and results instead of re-running commands or re-reading files this conversation already examined. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown here.',
     ].join('\n');
 }
 async function shouldStopAfterTurn(agent, message) {

package/dist/core/conversationKernel.d.ts

@@ -219,6 +219,7 @@ export declare class ConversationKernel {
     setWorkRunExpanded(target: ConversationTargetInput, runId: string, expanded: boolean): boolean;
     setInputMode(target: ConversationTargetInput, mode: string): 'guide' | 'next';
     setMode(target: ConversationTargetInput, mode: AgentMode): AgentMode;
+    setModel(target: ConversationTargetInput, model: string): string;
     toggleGoalPause(target: ConversationTargetInput): Promise<boolean>;
     clearGoal(target: ConversationTargetInput): boolean;
     updateSetting(section: string, key: string, value: unknown): void;
@@ -229,6 +230,13 @@ export declare class ConversationKernel {
     prompt(message: string | AgentPromptMessage, target: ConversationTargetInput, options: ConversationKernelRunOptions, queueMode?: ConversationQueueMode): Promise<ConversationKernelRunResult>;
     private settleCooperativeStop;
     private run;
+    /**
+     * Apply a model selection recorded while a Build block was running. The
+     * in-flight block never switches mid-block; the switch takes effect the next
+     * time a queued Guide/Next re-enters the block, and only when the pending
+     * selection actually differs from the runner's current selection.
+     */
+    private syncPendingModel;
     private runSingle;
     private processTimeoutMs;
     private runtime;

package/dist/core/conversationKernel.js

@@ -365,6 +365,24 @@ class ConversationKernel {
             runtime.options.mode = mode;
         return runner.mode;
     }
+    setModel(target, model) {
+        const normalized = this.normalizeTarget(target);
+        const runtime = this.findRuntime(normalized);
+        const runner = runtime?.runner || this.createRunner(normalized);
+        if (!runtime || !runtime.activePromise) {
+            // No Build block is running: the selection applies immediately.
+            runner.setModel(model);
+        }
+        else {
+            // A Build block is running. The in-flight block keeps its current model
+            // until the next Guide/Next re-enters it; record the newly selected model
+            // as the pending choice so the next dequeue switches to it. This is the
+            // "model switch does not take effect mid-block" contract.
+            runtime.options.model = model;
+        }
+        runner.saveWorkspaceConversationState(true);
+        return runner.model;
+    }
     async toggleGoalPause(target) {
         const normalized = this.normalizeTarget(target);
         let runtime = this.findRuntime(normalized);
@@ -486,14 +504,19 @@ class ConversationKernel {
     }
     async prompt(message, target, options, queueMode = 'followUp') {
         const normalized = this.normalizeTarget(target);
+        const active = this.findRuntime(normalized);
+        if (active?.activePromise) {
+            // A Build block is already running: queue this message. Queued messages
+            // carry no send-time model/mode; the running block keeps its settings and
+            // the next dequeue follows the current conversation selection (which
+            // setModel/setMode already recorded on runtime.options).
+            this.enqueueSameSession(active, message, queueMode);
+            this.activateAcceptedGoal(active, typeof message === 'string' ? '' : message.goalObjective);
+            return active.activePromise;
+        }
         const runtime = this.runtime(normalized, options);
         runtime.options = { ...options };
         this.applyOptions(runtime.runner, options);
-        if (runtime.activePromise) {
-            this.enqueueSameSession(runtime, message, queueMode);
-            this.activateAcceptedGoal(runtime, typeof message === 'string' ? '' : message.goalObjective);
-            return runtime.activePromise;
-        }
         runtime.generation = (this.generations.get(runtime.runtimeKey) || runtime.generation || 0) + 1;
         this.generations.set(runtime.runtimeKey, runtime.generation);
         const requestedRunId = typeof message === 'string' ? '' : String(message.runId || '').trim().slice(0, 200);
@@ -657,7 +680,23 @@ class ConversationKernel {
         this.mirrorHostIfTargetActive(runtime);
         return this.result(runtime, lastTokens);
     }
+    /**
+     * Apply a model selection recorded while a Build block was running. The
+     * in-flight block never switches mid-block; the switch takes effect the next
+     * time a queued Guide/Next re-enters the block, and only when the pending
+     * selection actually differs from the runner's current selection.
+     */
+    syncPendingModel(runtime) {
+        const pending = String(runtime.options.model || '').trim();
+        if (!pending)
+            return;
+        if (pending === runtime.runner.model || pending === runtime.runner.modelSelectionValue())
+            return;
+        runtime.runner.setModel(pending);
+        runtime.options.model = runtime.runner.modelSelectionValue();
+    }
     async runSingle(runtime, message, continuationMode) {
+        this.syncPendingModel(runtime);
         this.consumeQueuedMessage(runtime, typeof message === 'string' ? message : message.text);
         const timeoutMs = this.processTimeoutMs(runtime);
         if (timeoutMs <= 0) {

package/dist/core/electronUtilityAgentClient.d.ts

@@ -102,6 +102,7 @@ export declare class ElectronUtilityAgentClient {
     rateAutoRoute(score: number, routeId?: string): Promise<UtilityAutoRouteRatingResult>;
     setWorkRunExpanded(runId: string, expanded: boolean): Promise<boolean>;
     setMode(mode: AgentMode): Promise<AgentMode>;
+    setModel(model: string): Promise<string>;
     setInputMode(mode: string): Promise<'guide' | 'next'>;
     toggleGoalPause(): Promise<boolean>;
     clearGoal(): Promise<boolean>;

package/dist/core/electronUtilityAgentClient.js

@@ -1047,6 +1047,10 @@ class ElectronUtilityAgentClient {
         await this.start();
         return await this.request('set_mode', { target: this.target, mode }, 5_000);
     }
+    async setModel(model) {
+        await this.start();
+        return await this.request('set_model', { target: this.target, model }, 5_000);
+    }
     async setInputMode(mode) {
         await this.start();
         return await this.request('set_input_mode', { target: this.target, mode }, 5_000);

package/dist/core/electronUtilityRuntimePool.d.ts

@@ -18,6 +18,7 @@ export interface ElectronTargetRuntimeClient {
     rateAutoRoute?(score: number, routeId?: string): Promise<UtilityAutoRouteRatingResult>;
     setWorkRunExpanded(runId: string, expanded: boolean): Promise<boolean>;
     setMode?(mode: AgentMode): Promise<AgentMode>;
+    setModel?(model: string): Promise<string>;
     setInputMode?(mode: string): Promise<'guide' | 'next'>;
     toggleGoalPause?(): Promise<boolean>;
     clearGoal?(): Promise<boolean>;
@@ -79,6 +80,7 @@ export declare class ElectronUtilityRuntimePool {
     setWorkRunExpanded(target: ConversationRuntimeTarget, runId: string, expanded: boolean): Promise<boolean>;
     setInputMode(target: ConversationRuntimeTarget, mode: string): Promise<'guide' | 'next' | null>;
     setMode(target: ConversationRuntimeTarget, mode: AgentMode): Promise<AgentMode | null>;
+    setModel(target: ConversationRuntimeTarget, model: string): Promise<string | null>;
     toggleGoalPause(target: ConversationRuntimeTarget): Promise<boolean | null>;
     clearGoal(target: ConversationRuntimeTarget): Promise<boolean | null>;
     updateSetting(section: string, key: string, value: unknown): Promise<void>;

package/dist/core/electronUtilityRuntimePool.js

@@ -261,6 +261,17 @@ class ElectronUtilityRuntimePool {
             this.release(entry, true);
         }
     }
+    async setModel(target, model) {
+        const entry = await this.acquireExisting((0, conversationTarget_1.normalizeConversationTarget)(target));
+        if (!entry?.client.setModel)
+            return null;
+        try {
+            return await entry.client.setModel(model);
+        }
+        finally {
+            this.release(entry, true);
+        }
+    }
     async toggleGoalPause(target) {
         const entry = await this.acquire((0, conversationTarget_1.normalizeConversationTarget)(target));
         if (!entry?.client.toggleGoalPause)

package/dist/core/utilityAgentProtocol.d.ts

@@ -132,6 +132,13 @@ export type UtilityAgentRequest = {
         target: ConversationRuntimeTarget;
         mode: AgentMode;
     };
+} | {
+    id: string;
+    method: 'set_model';
+    params: {
+        target: ConversationRuntimeTarget;
+        model: string;
+    };
 } | {
     id: string;
     method: 'set_input_mode';

package/dist/core/wslAgentClient.d.ts

@@ -79,6 +79,7 @@ export declare class WslAgentClient {
     rateAutoRoute(target: ConversationRuntimeTarget, score: number, routeId?: string): Promise<WslAutoRouteRatingResult>;
     setWorkRunExpanded(target: ConversationRuntimeTarget, runId: string, expanded: boolean): Promise<boolean>;
     setMode(target: ConversationRuntimeTarget, mode: AgentMode): Promise<AgentMode>;
+    setModel(target: ConversationRuntimeTarget, model: string): Promise<string>;
     setInputMode(target: ConversationRuntimeTarget, mode: string): Promise<'guide' | 'next'>;
     toggleGoalPause(target: ConversationRuntimeTarget): Promise<boolean>;
     clearGoal(target: ConversationRuntimeTarget): Promise<boolean>;

package/dist/core/wslAgentClient.js

@@ -337,6 +337,10 @@ class WslAgentClient {
         await this.start();
         return await this.request('set_mode', { target: await this.mapTarget(target), mode }, 5_000);
     }
+    async setModel(target, model) {
+        await this.start();
+        return await this.request('set_model', { target: await this.mapTarget(target), model }, 5_000);
+    }
     async setInputMode(target, mode) {
         await this.start();
         return await this.request('set_input_mode', { target: await this.mapTarget(target), mode }, 5_000);

package/dist/core/wslAgentProtocol.d.ts

@@ -144,6 +144,13 @@ export type WslAgentRequest = {
         target: ConversationRuntimeTarget;
         mode: AgentMode;
     };
+} | {
+    id: string;
+    method: 'set_model';
+    params: {
+        target: ConversationRuntimeTarget;
+        model: string;
+    };
 } | {
     id: string;
     method: 'set_input_mode';

package/dist/core/wslAgentRuntimePool.d.ts

@@ -18,6 +18,7 @@ export interface WslTargetRuntimeClient {
     rateAutoRoute?(target: ConversationRuntimeTarget, score: number, routeId?: string): Promise<WslAutoRouteRatingResult>;
     setWorkRunExpanded(target: ConversationRuntimeTarget, runId: string, expanded: boolean): Promise<boolean>;
     setMode?(target: ConversationRuntimeTarget, mode: AgentMode): Promise<AgentMode>;
+    setModel?(target: ConversationRuntimeTarget, model: string): Promise<string>;
     setInputMode?(target: ConversationRuntimeTarget, mode: string): Promise<'guide' | 'next'>;
     toggleGoalPause?(target: ConversationRuntimeTarget): Promise<boolean>;
     clearGoal?(target: ConversationRuntimeTarget): Promise<boolean>;
@@ -81,6 +82,7 @@ export declare class WslAgentRuntimePool {
     setWorkRunExpanded(target: ConversationRuntimeTarget, runId: string, expanded: boolean): Promise<boolean>;
     setInputMode(target: ConversationRuntimeTarget, mode: string): Promise<'guide' | 'next' | null>;
     setMode(target: ConversationRuntimeTarget, mode: AgentMode): Promise<AgentMode | null>;
+    setModel(target: ConversationRuntimeTarget, model: string): Promise<string | null>;
     toggleGoalPause(target: ConversationRuntimeTarget): Promise<boolean | null>;
     clearGoal(target: ConversationRuntimeTarget): Promise<boolean | null>;
     updateSetting(section: string, key: string, value: unknown): Promise<void>;

package/dist/core/wslAgentRuntimePool.js

@@ -276,6 +276,18 @@ class WslAgentRuntimePool {
             this.release(entry, true);
         }
     }
+    async setModel(target, model) {
+        const normalized = (0, conversationTarget_1.normalizeConversationTarget)(target);
+        const entry = await this.acquireExisting(normalized);
+        if (!entry?.client.setModel)
+            return null;
+        try {
+            return await entry.client.setModel(normalized, model);
+        }
+        finally {
+            this.release(entry, true);
+        }
+    }
     async toggleGoalPause(target) {
         const normalized = (0, conversationTarget_1.normalizeConversationTarget)(target);
         const entry = await this.acquire(normalized);

package/dist/main.js

@@ -735,12 +735,21 @@ function setWindowsConsoleMode(mode) {
     const setMode = Number.isInteger(mode);
     const script = [
         'Add-Type -TypeDefinition \'using System; using System.Runtime.InteropServices; public static class NewmarkConsoleMode { [DllImport("kernel32.dll")] public static extern IntPtr GetStdHandle(int n); [DllImport("kernel32.dll")] public static extern bool GetConsoleMode(IntPtr h, out uint mode); [DllImport("kernel32.dll")] public static extern bool SetConsoleMode(IntPtr h, uint mode); }\'',
-        '$handle = [NewmarkConsoleMode]::GetStdHandle(-10)',
-        '$current = [uint32]0',
-        'if (-not [NewmarkConsoleMode]::GetConsoleMode($handle, [ref]$current)) { exit 2 }',
+        // 输出句柄（STD_OUTPUT_HANDLE = -11）必须启用 ENABLE_VIRTUAL_TERMINAL_PROCESSING(0x4)，
+        // 否则 TUI 的备用屏 ?1049h / 清屏 2J 等 ANSI 序列在 ConHost/传统控制台下不被解析，
+        // 主屏保留滚动历史（滚轮能滚回旧帧），光标追踪也随之错位。
+        '$out = [NewmarkConsoleMode]::GetStdHandle(-11)',
+        '$outMode = [uint32]0',
+        'if (-not [NewmarkConsoleMode]::GetConsoleMode($out, [ref]$outMode)) { exit 2 }',
         setMode
-            ? `$target = [uint32]${mode}; if (-not [NewmarkConsoleMode]::SetConsoleMode($handle, $target)) { exit 3 }`
-            : '$target = [uint32](($current -band (-bnot 7)) -bor 512); if (-not [NewmarkConsoleMode]::SetConsoleMode($handle, $target)) { exit 3 }; Write-Output $current',
+            ? `$target = [uint32]${mode}; if (-not [NewmarkConsoleMode]::SetConsoleMode($out, $target)) { exit 3 }`
+            : 'if (-not [NewmarkConsoleMode]::SetConsoleMode($out, ($outMode -bor 4))) { exit 3 }',
+        // 输入句柄（STD_INPUT_HANDLE = -10）启用 ENABLE_VIRTUAL_TERMINAL_INPUT(0x200)，
+        // 清除 line/echo 让方向键等原始序列可读。
+        '$inp = [NewmarkConsoleMode]::GetStdHandle(-10)',
+        '$inpMode = [uint32]0',
+        'if ([NewmarkConsoleMode]::GetConsoleMode($inp, [ref]$inpMode)) { [NewmarkConsoleMode]::SetConsoleMode($inp, (($inpMode -band (-bnot 7)) -bor 512)) | Out-Null }',
+        'Write-Output $outMode',
     ].join('; ');
     const result = (0, child_process_1.spawnSync)('powershell.exe', ['-NoProfile', '-NonInteractive', '-ExecutionPolicy', 'Bypass', '-Command', script], {
         stdio: ['inherit', 'pipe', 'inherit'],
@@ -1206,6 +1215,11 @@ else if (isTuiArg) {
         || path.basename(process.execPath).toLowerCase() === 'newmark console runtime.exe'
         || process.env.NEWMARK_CONSOLE_WRAPPER === '1');
     if (isConsoleLauncher && process.env.NEWMARK_TUI_SIDECAR !== '1') {
+        // Console mode is shared by every process attached to the same console:
+        // enable output VT processing (and input VT) before spawning the sidecar so
+        // the TUI's ?1049h alternate-screen sequence is actually parsed by ConHost /
+        // traditional consoles. The sidecar also re-applies it defensively below.
+        setWindowsConsoleMode();
         const tuiProcess = (0, child_process_1.spawnSync)(process.execPath, [path.join(__dirname, 'launcher.js'), ...args], {
             cwd: process.cwd(),
             env: {
@@ -3162,16 +3176,31 @@ else {
             });
             electron_1.ipcMain.handle('agent:setModel', async (_event, model) => {
                 if (agent) {
-                    const before = agent.model;
+                    // Compare the resolved selection (qualified deployment or 'auto')
+                    // rather than the bare model name, so switching between two
+                    // same-named models on different providers is still recognized as a
+                    // real change.
+                    const before = agent.modelSelectionValue();
                     agent.setModel(model, true);
+                    const after = agent.modelSelectionValue();
                     // Compression and kernel reset only make sense when the model actually
                     // changed. The renderer sends this on every prompt, so an unchanged
                     // selection must not trigger a context compression round (which can
                     // run an extra model call on large histories).
-                    if (agent.model !== before) {
+                    if (after !== before) {
                         await agent.compressForModelSwitch();
                         resetConversationKernel();
                     }
+                    // Propagate the selection to the target-bound runtime. A running Build
+                    // block keeps its current model until the next Guide/Next re-enters it;
+                    // the runtime records the new selection as pending so the context
+                    // window and the next dequeue both follow the newly selected model.
+                    const target = conversationRuntimeTarget(agent.activeConversationId || 'default');
+                    ensureConversationKernel(root)?.setModel(target, model);
+                    if (wslBackendEnabled())
+                        await ensureWslConversationPool()?.setModel(target, model);
+                    else
+                        await ensureElectronUtilityPool()?.setModel(target, model);
                 }
                 return agent?.model;
             });

package/dist/tools/index.js

@@ -350,7 +350,7 @@ class ToolExecutor {
             t('subagent_result', 'Return the persisted transcript, mailbox summary, status, and latest result for a peer agent. Target by exact id (preferred) or name.', { id: { type: 'string', description: 'Exact peer id from subagent_list.' }, name: { type: 'string', description: 'Convenience peer name.' } }, []),
             t('subagent_close', 'Close a same-conversation peer. Root can close any peer; a peer can close only itself. Target by exact id (preferred) or name.', { id: { type: 'string', description: 'Exact peer id from subagent_list.' }, name: { type: 'string', description: 'Convenience peer name.' } }, []),
             t('linked_plan', 'Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.', { action: { type: 'string', enum: ['get', 'update'] }, markdown: { type: 'string' }, expected_revision: { type: 'number' } }, ['action']),
-            t('build_history_query', 'Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.', { history_index: { type: 'number', minimum: 1, description: '1-based historical Build Block index from the request ledger; 1 is the newest previous task.' }, run_id: { type: 'string', description: 'Exact run id returned by an earlier build_history_query result.' }, max_events: { type: 'number', minimum: 1, maximum: 200, description: 'Maximum trailing public work events; defaults to 80.' }, max_chars: { type: 'number', minimum: 100, maximum: 4000, description: 'Per-event/per-guide content character bound; defaults to 2000.' } }, []),
+            t('build_history_query', 'Read the concrete public work details (tool calls, results, file changes, guides) of one historical Build Block. Call it proactively when the current task continues, fixes, verifies, or depends on earlier work: reuse the returned activity instead of re-investigating from scratch. Do not call it merely to answer completion status already exposed by the prompt. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.', { history_index: { type: 'number', minimum: 1, description: '1-based historical Build Block index from the request ledger; 1 is the newest previous task.' }, run_id: { type: 'string', description: 'Exact run id returned by an earlier build_history_query result.' }, max_events: { type: 'number', minimum: 1, maximum: 200, description: 'Maximum trailing public work events; defaults to 80.' }, max_chars: { type: 'number', minimum: 100, maximum: 4000, description: 'Per-event/per-guide content character bound; defaults to 2000.' } }, []),
             t('context_compress', 'Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.', { keep_recent: { type: 'number', minimum: 2, maximum: 60, description: 'Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages.' }, force: { type: 'boolean', description: 'Compress even if the context is not yet over the automatic threshold. Defaults to false.' } }, []),
             t('context_history_manage', 'Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. The hot cache stays bounded; evicted folded segments remain in a conversation-isolated append-only cold archive and are loaded only by explicit search/read/restore calls. Actions: list returns a bounded index of current context entries; remove declares one long-term entry for unload (see below); summarize folds a contiguous current range; restore reinserts a folded segment when its summary marker is still present; search finds matching hot or archived segments; read returns one bounded segment without injecting the whole archive; status reports budgets, hot cache, cold archive, the protected recent zone, and pending removals. The recent context tail and last user message are protected from remove/summarize unless dangerous is true. For cache-optimization, remove ONLY targets long-term history (never the protected recent tail or last user message) and does NOT unload immediately: the declared entry stays in context for the rest of the current Build Block so the provider prefix cache stays stable, then is physically removed when the Block ends — applying to subsequent Blocks only.', {
                 action: { type: 'string', enum: ['list', 'remove', 'summarize', 'restore', 'search', 'read', 'status'], description: 'list current entries; remove one; summarize a range; restore by restore_id; search hot/cold folded segments; read one bounded folded segment; status report context budgets and storage.' },

package/dist/tui/src/app.js

@@ -51,8 +51,10 @@ const ESC = "\u001b[";
 function createPaintScheduler(state, output = process.stdout, renderFrame = render) {
   let pending = false;
   let lastFrame = "";
+  let cancelled = false;
   const flush = () => {
     pending = false;
+    if (cancelled) return;
     const frame = renderFrame(state);
     if (frame === lastFrame) return false;
     lastFrame = frame;
@@ -65,6 +67,8 @@ function createPaintScheduler(state, output = process.stdout, renderFrame = rend
     setImmediate(flush);
   };
   paint.flush = flush;
+  // 退出前调用：丢弃任何排队中的重绘，避免恢复主屏后又被画上一帧 TUI 画面。
+  paint.cancel = () => { cancelled = true; pending = false; };
   return paint;
 }
 
@@ -108,6 +112,28 @@ function resolveTuiWorkspacePath(args, options = {}) {
   return explicitRoot || process.cwd();
 }
 
+// 在 Windows 传统控制台（ConHost）下，输出句柄默认不启用
+// ENABLE_VIRTUAL_TERMINAL_PROCESSING，导致备用屏 ?1049h / 清屏 2J 等 ANSI
+// 序列不被解析，主屏保留滚动历史（滚轮能滚回旧帧）。Electron-as-node sidecar
+// 也不会像纯 Node 那样由 libuv 自动启用 VT。这里主动给输出句柄启用 VT 处理，
+// 给输入句柄启用原始输入，作为 main.ts 之外入口的保险。
+function enableWindowsVirtualTerminal() {
+  if (process.platform !== 'win32') return;
+  try {
+    const { spawnSync } = require('node:child_process');
+    const script = [
+      'Add-Type -TypeDefinition \'using System; using System.Runtime.InteropServices; public static class NewmarkTuiVT { [DllImport("kernel32.dll")] public static extern IntPtr GetStdHandle(int n); [DllImport("kernel32.dll")] public static extern bool GetConsoleMode(IntPtr h, out uint m); [DllImport("kernel32.dll")] public static extern bool SetConsoleMode(IntPtr h, uint m); }\'',
+      '$o = [NewmarkTuiVT]::GetStdHandle(-11); $om = [uint32]0; if ([NewmarkTuiVT]::GetConsoleMode($o, [ref]$om)) { [NewmarkTuiVT]::SetConsoleMode($o, ($om -bor 4)) | Out-Null }',
+      '$i = [NewmarkTuiVT]::GetStdHandle(-10); $im = [uint32]0; if ([NewmarkTuiVT]::GetConsoleMode($i, [ref]$im)) { [NewmarkTuiVT]::SetConsoleMode($i, (($im -band (-bnot 7)) -bor 512)) | Out-Null }',
+    ].join('; ');
+    spawnSync('powershell.exe', ['-NoProfile', '-NonInteractive', '-ExecutionPolicy', 'Bypass', '-Command', script], {
+      stdio: 'ignore',
+      windowsHide: true,
+      timeout: 3000,
+    });
+  } catch {}
+}
+
 function start(options = {}) {
   const forcedTerminal = process.env.NEWMARK_FORCE_TTY === "1";
   if ((!process.stdin.isTTY || !process.stdout.isTTY) && !forcedTerminal) {
@@ -115,6 +141,9 @@ function start(options = {}) {
     process.exitCode = 1;
     return;
   }
+  // 备用屏依赖输出句柄的 VT 处理；纯 Node 的 libuv 会自动启用，但
+  // Electron-as-node sidecar 或强制 TTY 路径不会，这里补上。
+  if (forcedTerminal || !process.stdout.isTTY) enableWindowsVirtualTerminal();
 
   const args = process.argv.slice(2);
   let adapter;
@@ -140,24 +169,38 @@ function start(options = {}) {
   let timer = null;
   let animationTimer = null;
   let closing = false;
+  // 标准 TUI 屏幕模型：进入备用屏幕缓冲（alternate screen buffer）。
+  // 备用屏没有主屏的滚动历史：每次刷新用 2J 全量重绘而不是滚动新页面，
+  // 鼠标滚轮无法滚回上一次刷新，终端滚动也不会让画面错位（否则高亮行
+  // 会随终端滚动"离开屏幕"）。
+  const enterAltScreen = () => process.stdout.write(`${ESC}?1049h${ESC}?25l${ESC}2J${ESC}H`);
+  const leaveAltScreen = () => `${ESC}?25h${ESC}?1049l${ESC}0m${ESC}2J${ESC}H`;
   const paint = createPaintScheduler(state);
   const cleanup = () => {
     if (timer) clearInterval(timer);
     if (animationTimer) clearInterval(animationTimer);
     if (typeof process.stdin.setRawMode === "function") process.stdin.setRawMode(false);
     process.stdin.pause();
-    process.stdout.write(`${ESC}?25h${ESC}0m${ESC}2J${ESC}H`);
   };
   const quit = () => {
     if (closing) return;
     closing = true;
+    paint.cancel();
     cleanup();
     if (typeof state.adapter.close === "function") state.adapter.close();
-    process.stdout.write(state.adapterKind === "mock"
-      ? "Newmark TUI demo closed. No state was saved.\n"
-      : "Newmark TUI closed. Conversation and settings state are persisted by Newmark.\n");
+    // 退出序列与关闭消息用 fs.writeSync 同步写入 fd：直接进入内核管道，
+    // 不经过异步流缓冲，process.exit() 不会丢弃任何字节（此前异步 write
+    // 在退出瞬间被 ConPTY/管道丢弃，导致恢复主屏后看不到退出消息）。
+    try {
+      const fs = require("node:fs");
+      const fd = process.stdout.fd || 1;
+      fs.writeSync(fd, `${leaveAltScreen()}\n`);
+      fs.writeSync(fd, state.adapterKind === "mock"
+        ? "Newmark TUI demo closed. No state was saved.\n"
+        : "Newmark TUI closed. Conversation and settings state are persisted by Newmark.\n");
+    } catch {}
     process.exitCode = 0;
-    setImmediate(() => process.exit(0));
+    process.exit(0);
   };
 
   function simulateReply(text) {
@@ -554,8 +597,10 @@ function start(options = {}) {
   process.stdin.resume();
   process.stdin.on("keypress", handleKey);
   process.stdout.on("resize", paint);
-  process.on("exit", () => process.stdout.write(`${ESC}?25h${ESC}0m`));
+  // 兜底：任何退出路径（包括未捕获异常后的 exit 事件）都恢复主屏缓冲与光标。
+  process.on("exit", () => process.stdout.write(`${ESC}?25h${ESC}?1049l${ESC}0m`));
   process.on("SIGTERM", quit);
+  enterAltScreen();
   paint();
 }
 

package/dist/tui/src/data.js

@@ -204,6 +204,34 @@ const providers = [
   }
 ];
 
+// Demo-only long-list injection: NEWMARK_TUI_DEMO_MODELS=<count> appends a
+// dedicated stress provider with <count> models so the featureless --demo TUI
+// can exercise long model-selection cursor-follow behavior without any runtime.
+const DEMO_MODEL_COUNT = Number(process.env.NEWMARK_TUI_DEMO_MODELS || 0);
+if (Number.isFinite(DEMO_MODEL_COUNT) && DEMO_MODEL_COUNT > 0) {
+  providers.push({
+    id: "provider-demo-stress",
+    name: "Demo Stress",
+    base_url: "https://demo.invalid/v1",
+    api_key: "",
+    has_api_key: false,
+    protocol: "openai",
+    enabled: true,
+    models: Array.from({ length: DEMO_MODEL_COUNT }, (_, index) => ({
+      name: `demo-stress-model-${String(index).padStart(3, "0")}`,
+      display: `Stress Model ${String(index).padStart(3, "0")}`,
+      description: `Long-list cursor-follow stress model ${index} description`,
+      max_tokens: 128000,
+      vision: false,
+      thinking: false,
+      enabled: true,
+      speed_rating: "unknown",
+      capability_rating: "unknown",
+      validation: { status: "unavailable", level: "discovered", checked_at: "" }
+    }))
+  });
+}
+
 const flows = [
   {
     name: "release-readiness",

package/dist/tui/src/render.js

@@ -689,29 +689,34 @@ function modelView(state, width, p) {
   const isCurrent = (selection) => selection.kind === current.kind
     && (selection.kind === "auto"
       || (selection.providerId === current.providerId && selection.modelId === current.modelId));
-  return [
+  const rows = [
     ...conversationContext(state, p, "Model and reasoning effort"),
-    `${p.bold}${tr(state, "Reasoning effort")}${p.reset}   ${p.muted}${tr(state, "Shared GUI/TUI request tier · ←/→ changes section")}${p.reset}`,
-    ...INTELLIGENCE_TIERS.map((tier, index) => {
-      const style = state.focusRegion === "content" && state.contentColumn === 0 && index === state.selected ? `${p.selected}${p.bold}` : "";
-      if (style) state.contentFocusLine = 5 + index;
-      return `${style} ${tier === currentTier ? `${p.cyan}●${p.reset}` : "○"} ${tier}${p.reset}`;
-    }),
+    `${p.bold}${tr(state, "Reasoning effort")}${p.reset}   ${p.muted}${tr(state, "Shared GUI/TUI request tier · ←/→ changes section")}${p.reset}`
+  ];
+  const tierStart = rows.length;
+  INTELLIGENCE_TIERS.forEach((tier, index) => {
+    const style = state.focusRegion === "content" && state.contentColumn === 0 && index === state.selected ? `${p.selected}${p.bold}` : "";
+    if (style) state.contentFocusLine = tierStart + index;
+    rows.push(`${style} ${tier === currentTier ? `${p.cyan}●${p.reset}` : "○"} ${tier}${p.reset}`);
+  });
+  rows.push(
     "",
     `${p.bold}${tr(state, "Deployment")}${p.reset}   ${p.muted}${tr(state, "Used by this conversation, including its Plan and Subagents")}${p.reset}`,
-    "",
-    ...options.flatMap((option, index) => {
-      const style = state.focusRegion === "content" && state.contentColumn === 1 && index === state.selected ? `${p.selected}${p.bold}` : "";
-      if (style) state.contentFocusLine = 14 + index * 2;
-      const marker = isCurrent(option.selection) ? `${p.cyan}●${p.reset}` : "○";
-      return [
-        `${style} ${marker} ${pad(option.label, Math.max(18, Math.min(32, width - 24)))} ${p.muted}${option.provider}${p.reset}`,
-        `   ${p.muted}${truncate(option.description, Math.max(20, width - 5))}${p.reset}`
-      ];
-    }),
+    ""
+  );
+  const deploymentStart = rows.length;
+  options.forEach((option, index) => {
+    const style = state.focusRegion === "content" && state.contentColumn === 1 && index === state.selected ? `${p.selected}${p.bold}` : "";
+    if (style) state.contentFocusLine = deploymentStart + index * 2;
+    const marker = isCurrent(option.selection) ? `${p.cyan}●${p.reset}` : "○";
+    rows.push(`${style} ${marker} ${pad(option.label, Math.max(18, Math.min(32, width - 24)))} ${p.muted}${option.provider}${p.reset}`);
+    rows.push(`   ${p.muted}${truncate(option.description, Math.max(20, width - 5))}${p.reset}`);
+  });
+  rows.push(
     "",
     `${p.muted}${tr(state, "Enter applies the focused tier or deployment. Effort persists globally; deployments remain per conversation.")}${p.reset}`
-  ];
+  );
+  return rows;
 }
 
 function flowBarView(state, width, p) {
@@ -1250,14 +1255,38 @@ function overlayLines(state, width, p) {
   return [];
 }
 
-function render(state, columns = process.stdout.columns || 100, rows = process.stdout.rows || 30) {
+// 实时读取终端窗口尺寸：优先 getWindowSize()（每次调用都查询底层 TTY 尺寸，
+// 不依赖 Node 缓存的 columns/rows，避免 resize 事件丢失或延迟时按旧尺寸绘制），
+// 其次回退到缓存 columns/rows，再回退 COLUMNS/LINES 环境变量与默认值。
+function readWindowSize(fallbackColumns = 100, fallbackRows = 30) {
+  const stdout = process.stdout;
+  if (stdout && typeof stdout.getWindowSize === "function") {
+    try {
+      const [width, height] = stdout.getWindowSize();
+      if (Number.isFinite(width) && Number.isFinite(height) && width > 0 && height > 0) {
+        return { columns: Math.max(1, Math.floor(width)), rows: Math.max(1, Math.floor(height)) };
+      }
+    } catch {}
+  }
+  const columns = Number(stdout?.columns) || Number(process.env.COLUMNS) || fallbackColumns;
+  const rows = Number(stdout?.rows) || Number(process.env.LINES) || fallbackRows;
+  return { columns: Math.max(1, Math.floor(columns)), rows: Math.max(1, Math.floor(rows)) };
+}
+
+function render(state, columns, rows) {
+  const size = (columns !== undefined && rows !== undefined)
+    ? { columns: Math.max(1, Math.floor(Number(columns) || 1)), rows: Math.max(1, Math.floor(Number(rows) || 1)) }
+    : readWindowSize();
   const p = palette(state);
-  const width = Math.max(52, columns);
-  const height = Math.max(20, rows);
+  // 严格约束：直接采用终端实际尺寸，绝不通过 Math.max(52/20) 强制放大。
+  // 终端窗口小于内部最小布局时，内容在输出阶段裁剪到窗口内，否则帧会
+  // 画到窗口之外（用户看不到底部行/右侧列，如小窗口下的模型长菜单）。
+  const width = size.columns;
+  const height = size.rows;
   const compact = width < 78;
   const sidebarWidth = compact ? 0 : 22;
   const contentWidth = width - sidebarWidth - 2;
-  const bodyHeight = height - 3;
+  const bodyHeight = Math.max(1, height - 3);
   const sidebar = sidebarWidth ? renderSidebar(state, bodyHeight, sidebarWidth, p) : [];
   const content = renderContent(state, contentWidth, bodyHeight, p);
   let contentLines = content;
@@ -1266,8 +1295,11 @@ function render(state, columns = process.stdout.columns || 100, rows = process.s
     let scroll = Math.max(0, Math.min(maximumScroll, Number(state.contentScroll) || 0));
     const focusLine = Number(state.contentFocusLine) || -1;
     if (focusLine >= 0) {
-      if (focusLine < scroll) scroll = focusLine;
-      else if (focusLine >= scroll + bodyHeight) scroll = focusLine - bodyHeight + 1;
+      // 居中跟随：焦点行尽量保持在视口中上部，上下都留出内容，避免选中行
+      // 贴住视口边缘或被裁掉其相邻行（如两行一组的模型选项主行+描述行）。
+      const targetLine = Math.min(Math.max(0, Math.floor(bodyHeight / 3)), Math.max(0, bodyHeight - 2));
+      if (focusLine < scroll + targetLine) scroll = Math.max(0, focusLine - targetLine);
+      else if (focusLine >= scroll + bodyHeight - 1 - targetLine) scroll = Math.min(maximumScroll, focusLine - (bodyHeight - 1 - targetLine));
     }
     scroll = Math.max(0, Math.min(maximumScroll, scroll));
     state.contentScroll = scroll;
@@ -1286,20 +1318,25 @@ function render(state, columns = process.stdout.columns || 100, rows = process.s
     : "";
   if (compact) body.unshift(pad(top, width));
   const focus = state.focusRegion === "menu" ? "MENU" : "CONTENT";
-  const footer = `${p.panel} ${p.cyan}${focus}${p.reset}${p.panel}${p.muted} · ${truncate(state.notice, width - 42)}${p.reset}${p.panel}${" ".repeat(Math.max(1, width - visibleLength(state.notice) - visibleLength(focus) - 35))}Tab back  ? help  Q quit ${p.reset}`;
-  const renderedBody = body.slice(0, height - 1);
-  while (renderedBody.length < height - 1) renderedBody.push(pad("", width));
-  let output = `${ESC}?25l${ESC}2J${ESC}H${p.paint}${renderedBody.join("\n")}\n${pad(footer, width)}`;
+  const footer = `${p.panel} ${p.cyan}${focus}${p.reset}${p.panel}${p.muted} · ${truncate(state.notice, Math.max(1, width - 42))}${p.reset}${p.panel}${" ".repeat(Math.max(1, width - visibleLength(state.notice) - visibleLength(focus) - 35))}Tab back  ? help  Q quit ${p.reset}`;
+  const bodyLines = Math.max(1, height - 1);
+  const renderedBody = body.slice(0, bodyLines);
+  while (renderedBody.length < bodyLines) renderedBody.push(pad("", width));
+  // 输出阶段严格边界：每行按可见宽度截断到窗口宽度，行数限制在窗口高度内，
+  // 确保无论内部最小布局如何，帧绝不会画到窗口之外。
+  const boundedBody = renderedBody.map((line) => truncate(line, width));
+  const boundedFooter = truncate(pad(footer, width), width);
+  let output = `${ESC}?25l${ESC}2J${ESC}H${p.paint}${boundedBody.join("\n")}\n${boundedFooter}`;
   const overlay = overlayLines(state, width, p);
   if (overlay.length) {
     const overlayWidth = Math.max(...overlay.map(visibleLength));
     const x = Math.max(1, Math.floor((width - overlayWidth) / 2) + 1);
     const y = Math.max(2, Math.floor((height - overlay.length) / 2));
     overlay.forEach((line, index) => {
-      output += `${ESC}${y + index};${x}H${line}`;
+      output += `${ESC}${y + index};${x}H${truncate(line, width)}`;
     });
   }
   return `${output}${p.final}`;
 }
 
-module.exports = { render, stripAnsi, visibleLength, wrapText };
+module.exports = { render, readWindowSize, stripAnsi, visibleLength, wrapText };

package/dist/ui/index.html

@@ -22825,6 +22825,17 @@ function schedulePostStartupUiRendering() {
         var modelResult = await api.setModel(state.model);
         if (modelResult != null) state._syncedModel = state.model;
       }
+      // The context ring/inspector must immediately reflect the newly selected
+      // model's configured context window rather than waiting for the next send.
+      if (api.getState) {
+        var ctxTarget = currentConversationTarget();
+        api.getState(ctxTarget).then(function(s) {
+          if (s && s.contextWindow !== undefined && isActiveConversationTarget(ctxTarget) && isViewingRuntimeConversationBranch(ctxTarget)) {
+            state.contextWindow = s.contextWindow;
+            window.renderContextWindow();
+          }
+        }).catch(function(){});
+      }
     });
   }
 

package/dist/wsl-agent-host.bundle.cjs

@@ -336409,7 +336409,7 @@ var ToolExecutor = class {
       t3("subagent_result", "Return the persisted transcript, mailbox summary, status, and latest result for a peer agent. Target by exact id (preferred) or name.", { id: { type: "string", description: "Exact peer id from subagent_list." }, name: { type: "string", description: "Convenience peer name." } }, []),
       t3("subagent_close", "Close a same-conversation peer. Root can close any peer; a peer can close only itself. Target by exact id (preferred) or name.", { id: { type: "string", description: "Exact peer id from subagent_list." }, name: { type: "string", description: "Convenience peer name." } }, []),
       t3("linked_plan", "Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.", { action: { type: "string", enum: ["get", "update"] }, markdown: { type: "string" }, expected_revision: { type: "number" } }, ["action"]),
-      t3("build_history_query", "Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." }, max_chars: { type: "number", minimum: 100, maximum: 4e3, description: "Per-event/per-guide content character bound; defaults to 2000." } }, []),
+      t3("build_history_query", "Read the concrete public work details (tool calls, results, file changes, guides) of one historical Build Block. Call it proactively when the current task continues, fixes, verifies, or depends on earlier work: reuse the returned activity instead of re-investigating from scratch. Do not call it merely to answer completion status already exposed by the prompt. Select by newest-to-oldest history_index, or by run_id returned from an earlier query. Every activity/guide content is bounded to max_chars (default 2000) to keep the read lean and cache-friendly.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." }, max_chars: { type: "number", minimum: 100, maximum: 4e3, description: "Per-event/per-guide content character bound; defaults to 2000." } }, []),
       t3("context_compress", "Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.", { keep_recent: { type: "number", minimum: 2, maximum: 60, description: "Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages." }, force: { type: "boolean", description: "Compress even if the context is not yet over the automatic threshold. Defaults to false." } }, []),
       t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. The hot cache stays bounded; evicted folded segments remain in a conversation-isolated append-only cold archive and are loaded only by explicit search/read/restore calls. Actions: list returns a bounded index of current context entries; remove declares one long-term entry for unload (see below); summarize folds a contiguous current range; restore reinserts a folded segment when its summary marker is still present; search finds matching hot or archived segments; read returns one bounded segment without injecting the whole archive; status reports budgets, hot cache, cold archive, the protected recent zone, and pending removals. The recent context tail and last user message are protected from remove/summarize unless dangerous is true. For cache-optimization, remove ONLY targets long-term history (never the protected recent tail or last user message) and does NOT unload immediately: the declared entry stays in context for the rest of the current Build Block so the provider prefix cache stays stable, then is physically removed when the Block ends \u2014 applying to subsequent Blocks only.", {
         action: { type: "string", enum: ["list", "remove", "summarize", "restore", "search", "read", "status"], description: "list current entries; remove one; summarize a range; restore by restore_id; search hot/cold folded segments; read one bounded folded segment; status report context budgets and storage." },
@@ -340904,7 +340904,7 @@ function buildConversationTaskLedger(agent) {
     "Unfinished Continuation Queue (newest to oldest; summary fields only; use only when the current user instruction authorizes continuation and the task is relevant):",
     ...unfinishedLines.length ? unfinishedLines : ["(none)"],
     ...unfinished.length > unfinishedLines.length ? [`(${unfinished.length - unfinishedLines.length} older unfinished run(s) omitted from the bounded prompt ledger.)`] : [],
-    "When concrete work details are required, call build_history_query with history_index from this list. Do not call it merely to answer completion status already shown here."
+    "When the current task continues, fixes, verifies, or depends on earlier work in this list, proactively call build_history_query with its history_index before re-investigating. Reuse the returned tool activity and results instead of re-running commands or re-reading files this conversation already examined. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown here."
   ].join("\n");
 }
 async function shouldStopAfterTurn(agent, message) {
@@ -349172,8 +349172,13 @@ ${summary}`, segment, "local-summarize", true);
     };
   }
   resolveWindowModel(modelName) {
-    if (modelName !== "auto") return this.config.findModel(modelName);
-    return this.activeModelConfig() || this.config.findModel(this.config.getStr("models", "default_model"));
+    if (modelName === "auto" || modelName === this.model || modelName === this.activeModelName()) {
+      const active = this.activeModelConfig();
+      if (active) return active;
+    }
+    const byName = this.config.findModel(modelName);
+    if (byName) return byName;
+    return this.config.findModel(this.config.getStr("models", "default_model"));
   }
   contextMaxTokens(modelName = this.model) {
     const model = this.resolveWindowModel(modelName);
@@ -352337,7 +352342,7 @@ ${custom}`);
       "- Memory Lab is governed by an explicit Policy chain: pre-think whether memory is needed; prefer bounded memory_lab_query retrieval; then choose ADD/UPDATE/DELETE only when the user authorizes durable memory mutation.",
       "- Before memory_lab_update, inspect the target with memory_lab_query or memory_lab_read. For an existing component pass expectedUpdatedAt so concurrent/stale writes fail closed; preserve established tag parent paths.",
       "- Use memory_lab_delete only for an explicit user request to forget/remove memory. Prior revisions are retained under Memory Lab/archive and mutation decisions are appended to policy.jsonl for replay.",
-      "- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. Use build_history_query only when the current user asks what specifically happened in one Build Block; querying history is read-only and never authorizes resuming that work.",
+      "- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. When the current task continues, fixes, verifies, or depends on earlier Build Blocks, proactively call build_history_query to read the concrete tool activity and results of the relevant block, and reuse that information instead of re-investigating (re-running commands or re-reading files) from scratch. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown in the prompt.",
       "- Linked plan disclosure: a durable conversation-linked Markdown plan exists and can be inspected or updated with linked_plan when explicitly needed or required by Plan mode. Its full Markdown and revision are not injected into every model request.",
       "- A memory_lab_update, memory_lab_delete, or memory_lab_reindex call is unfinished until its awaited tool result contains rebuildReceipt.completed=true. The completion receipt is represented by the tool activity inside the current Build block and should not be repeated as a separate completion message.",
       `- Skills and subagents: skill searches enabled metadata and loads one SKILL.md body on demand; skill_download installs offline skill folders; task creates constrained subagents tracked in agent state.`,
@@ -353135,6 +353140,18 @@ var ConversationKernel = class {
     if (runtime) runtime.options.mode = mode;
     return runner.mode;
   }
+  setModel(target, model) {
+    const normalized = this.normalizeTarget(target);
+    const runtime = this.findRuntime(normalized);
+    const runner = runtime?.runner || this.createRunner(normalized);
+    if (!runtime || !runtime.activePromise) {
+      runner.setModel(model);
+    } else {
+      runtime.options.model = model;
+    }
+    runner.saveWorkspaceConversationState(true);
+    return runner.model;
+  }
   async toggleGoalPause(target) {
     const normalized = this.normalizeTarget(target);
     let runtime = this.findRuntime(normalized);
@@ -353249,14 +353266,15 @@ var ConversationKernel = class {
   }
   async prompt(message, target, options, queueMode = "followUp") {
     const normalized = this.normalizeTarget(target);
+    const active = this.findRuntime(normalized);
+    if (active?.activePromise) {
+      this.enqueueSameSession(active, message, queueMode);
+      this.activateAcceptedGoal(active, typeof message === "string" ? "" : message.goalObjective);
+      return active.activePromise;
+    }
     const runtime = this.runtime(normalized, options);
     runtime.options = { ...options };
     this.applyOptions(runtime.runner, options);
-    if (runtime.activePromise) {
-      this.enqueueSameSession(runtime, message, queueMode);
-      this.activateAcceptedGoal(runtime, typeof message === "string" ? "" : message.goalObjective);
-      return runtime.activePromise;
-    }
     runtime.generation = (this.generations.get(runtime.runtimeKey) || runtime.generation || 0) + 1;
     this.generations.set(runtime.runtimeKey, runtime.generation);
     const requestedRunId = typeof message === "string" ? "" : String(message.runId || "").trim().slice(0, 200);
@@ -353401,7 +353419,21 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.mirrorHostIfTargetActive(runtime);
     return this.result(runtime, lastTokens);
   }
+  /**
+   * Apply a model selection recorded while a Build block was running. The
+   * in-flight block never switches mid-block; the switch takes effect the next
+   * time a queued Guide/Next re-enters the block, and only when the pending
+   * selection actually differs from the runner's current selection.
+   */
+  syncPendingModel(runtime) {
+    const pending3 = String(runtime.options.model || "").trim();
+    if (!pending3) return;
+    if (pending3 === runtime.runner.model || pending3 === runtime.runner.modelSelectionValue()) return;
+    runtime.runner.setModel(pending3);
+    runtime.options.model = runtime.runner.modelSelectionValue();
+  }
   async runSingle(runtime, message, continuationMode) {
+    this.syncPendingModel(runtime);
     this.consumeQueuedMessage(runtime, typeof message === "string" ? message : message.text);
     const timeoutMs = this.processTimeoutMs(runtime);
     if (timeoutMs <= 0) {
@@ -354100,6 +354132,9 @@ async function handle(request) {
   if (request.method === "set_mode") {
     return kernel.setMode(requestTarget(request.params), request.params.mode);
   }
+  if (request.method === "set_model") {
+    return kernel.setModel(requestTarget(request.params), request.params.model);
+  }
   if (request.method === "set_input_mode") {
     return kernel.setInputMode(requestTarget(request.params), request.params.mode);
   }

package/dist/wsl-agent-host.js

@@ -206,6 +206,9 @@ async function handle(request) {
     if (request.method === 'set_mode') {
         return kernel.setMode(requestTarget(request.params), request.params.mode);
     }
+    if (request.method === 'set_model') {
+        return kernel.setModel(requestTarget(request.params), request.params.model);
+    }
     if (request.method === 'set_input_mode') {
         return kernel.setInputMode(requestTarget(request.params), request.params.mode);
     }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "newmark-agent",
   "productName": "Newmark Agent",
-  "version": "0.4.3",
+  "version": "0.4.4",
   "description": "Newmark Agent — Portable AI coding agent with rich GUI and CLI (TypeScript)",
   "homepage": "https://github.com/positer/Newmark-Agent",
   "repository": {
@@ -33,7 +33,7 @@
     "start:dev": "npm run build && electron .",
     "start:cli": "node dist/launcher.js --cli",
     "test:tui": "npm run build && npm run test:tui:built",
-    "test:tui:built": "node ../TUI/test/demo.test.js && node ../TUI/test/cursor-follow-stress.test.js && node dist/tests/tuiLauncherVerify.js && node dist/tests/tuiStopRaceVerify.js",
+    "test:tui:built": "node ../TUI/test/demo.test.js && node ../TUI/test/cursor-follow-stress.test.js && node ../TUI/test/cursor-follow-pty-gate.js && node dist/tests/tuiLauncherVerify.js && node dist/tests/tuiStopRaceVerify.js",
     "test:ssh-tui-stress": "npm run build && npm run test:ssh-tui-stress:built",
     "test:ssh-tui-stress:built": "node scripts/release-ssh-tui-stress.cjs",
     "test:wsl-tui-stress:built": "node scripts/release-wsl-tui-stress.cjs",
@@ -53,7 +53,7 @@
     "test:desktop": "npm run build && npm run test:desktop:built",
     "test:deletion-safety": "npm run build && npm run test:deletion-safety:built",
     "test:deletion-safety:built": "node scripts/deletion-safety-stress.cjs",
-  "test:desktop:built": "node dist/tests/verify.js && node dist/tests/thinkingTierMapCacheStressVerify.js && node dist/tests/computerUseSessionVerify.js && node dist/tests/contextSystemV2Verify.js && node dist/tests/contextSystemV2StressVerify.js && node dist/tests/providerAdapterV2Verify.js && node dist/tests/providerTimeoutRecoveryVerify.js && node dist/tests/agentRuntimeV2Verify.js && node dist/tests/toolchainExposureV2Verify.js && node dist/tests/localOcrFallbackVerify.js && node dist/tests/dev018ModeSubagentStressVerify.js && node dist/tests/conversationBranchStressVerify.js && node dist/tests/conversationArchiveConcurrencyVerify.js && node dist/tests/conversationArchiveRuntimeVerify.js && node dist/tests/dev040ComprehensiveStressVerify.js && node dist/tests/memoryPolicyVerify.js && node dist/tests/newmarkSelectVerify.js && node dist/tests/mcpManagerVerify.js && node dist/tests/dshCompatibilityVerify.js && node dist/tests/performanceOptimizationVerify.js && node scripts/compression-pressure-stress.cjs && node dist/tests/compressionFidelityVerify.js && node dist/tests/responseTrajectoryVerify.js && node dist/tests/normalChatRegressionVerify.js && node dist/tests/dev008-subagent.js && node dist/tests/dev009Verify.js && node dist/tests/runtimeIsolationVerify.js && node dist/tests/runtimePoolCapacityVerify.js && node dist/tests/guideWorkRunVerify.js && node scripts/flow-pause-stop-draft-stress.cjs && node dist/tests/guideUiReconcileVerify.js && node dist/tests/queueAttachmentIsolationVerify.js && node dist/tests/workspaceMenuVerify.js && node dist/tests/userImagePersistenceVerify.js && node dist/tests/displayImageVerify.js && node dist/tests/visualPreferencesVerify.js && node dist/tests/browserUseVerify.js && node dist/tests/toolProcessVerify.js && node dist/tests/toolProvisioningVerify.js && node dist/tests/taskToolsVerify.js && node dist/tests/contextCacheHitStressVerify.js && node dist/tests/startupPrewarmVerify.js && node dist/tests/pdfPreviewServerVerify.js && node dist/tests/editorLifecycleVerify.js && node dist/tests/terminalTakeoverVerify.js && node dist/tests/workspaceFileRouterVerify.js && node dist/tests/computerUsePerformanceVerify.js && node dist/tests/autoRouterVerify.js && node dist/tests/autoAgentIntegrationVerify.js && node dist/tests/autoRouteRatingVerify.js && node dist/tests/providerIdentityVerify.js && node dist/tests/modelValidationVerify.js && node dist/tests/modelValidationAgentIntegrationVerify.js && node scripts/deletion-safety-stress.cjs",
+  "test:desktop:built": "node dist/tests/verify.js && node dist/tests/thinkingTierMapCacheStressVerify.js && node dist/tests/computerUseSessionVerify.js && node dist/tests/contextSystemV2Verify.js && node dist/tests/contextSystemV2StressVerify.js && node dist/tests/providerAdapterV2Verify.js && node dist/tests/providerTimeoutRecoveryVerify.js && node dist/tests/agentRuntimeV2Verify.js && node dist/tests/toolchainExposureV2Verify.js && node dist/tests/localOcrFallbackVerify.js && node dist/tests/dev018ModeSubagentStressVerify.js && node dist/tests/conversationBranchStressVerify.js && node dist/tests/conversationArchiveConcurrencyVerify.js && node dist/tests/conversationArchiveRuntimeVerify.js && node dist/tests/dev040ComprehensiveStressVerify.js && node dist/tests/memoryPolicyVerify.js && node dist/tests/newmarkSelectVerify.js && node dist/tests/mcpManagerVerify.js && node dist/tests/dshCompatibilityVerify.js && node dist/tests/performanceOptimizationVerify.js && node scripts/compression-pressure-stress.cjs && node dist/tests/compressionFidelityVerify.js && node dist/tests/responseTrajectoryVerify.js && node dist/tests/normalChatRegressionVerify.js && node dist/tests/dev008-subagent.js && node dist/tests/dev009Verify.js && node dist/tests/runtimeIsolationVerify.js && node dist/tests/runtimePoolCapacityVerify.js && node dist/tests/guideWorkRunVerify.js && node dist/tests/modelSwitchBehaviorVerify.js && node scripts/flow-pause-stop-draft-stress.cjs && node dist/tests/guideUiReconcileVerify.js && node dist/tests/queueAttachmentIsolationVerify.js && node dist/tests/workspaceMenuVerify.js && node dist/tests/userImagePersistenceVerify.js && node dist/tests/displayImageVerify.js && node dist/tests/visualPreferencesVerify.js && node dist/tests/browserUseVerify.js && node dist/tests/toolProcessVerify.js && node dist/tests/toolProvisioningVerify.js && node dist/tests/taskToolsVerify.js && node dist/tests/contextCacheHitStressVerify.js && node dist/tests/startupPrewarmVerify.js && node dist/tests/pdfPreviewServerVerify.js && node dist/tests/editorLifecycleVerify.js && node dist/tests/terminalTakeoverVerify.js && node dist/tests/workspaceFileRouterVerify.js && node dist/tests/computerUsePerformanceVerify.js && node dist/tests/autoRouterVerify.js && node dist/tests/autoAgentIntegrationVerify.js && node dist/tests/autoRouteRatingVerify.js && node dist/tests/providerIdentityVerify.js && node dist/tests/modelValidationVerify.js && node dist/tests/modelValidationAgentIntegrationVerify.js && node scripts/deletion-safety-stress.cjs",
     "test:conversation-branch-stress": "npm run build && node dist/tests/conversationBranchStressVerify.js",
     "test:conversation-archive-concurrency": "npm run build && node dist/tests/conversationArchiveConcurrencyVerify.js",
     "test:memory-policy": "npm run build && node dist/tests/memoryPolicyVerify.js",