增强游戏创作智能体生成过程可见性

在聊天中实时展示 Planner、Orchestrator、Generator 和 Evaluator 进度
将生成完成摘要切换为完整 run trace 和 LLM 对话证据
补齐质量评审任务、测试覆盖和实施计划文档

apps/ai-game-creator-shell/src-tauri/src/main.rs

@@ -31,6 +31,7 @@ use shared_contracts::game_creation_app::{
     GAME_CREATION_AGENT_RUN_SCHEMA_VERSION, GAME_CREATION_AGENT_TOOL_CALL_MAX,
     GAME_CREATION_APP_COMMANDS, GAME_CREATION_APP_LIMITED_RUN_COMMANDS,
 };
+use tauri::Emitter;
 use tauri_plugin_opener::OpenerExt;
 
 #[derive(Debug, Eq, PartialEq, Serialize)]
@@ -69,6 +70,14 @@ struct GenerateLocalGameDraftResult {
     manifest: GameCreationAppManifest,
 }
 
+#[derive(Clone, Debug, Eq, PartialEq, Serialize)]
+#[serde(rename_all = "camelCase")]
+struct GameCreatorAgentProgressEvent {
+    project_path: String,
+    stage: String,
+    message: String,
+}
+
 #[derive(Debug, Eq, PartialEq, Serialize)]
 #[serde(rename_all = "camelCase")]
 struct GameCreatorLlmConfigStatus {
@@ -460,7 +469,7 @@ static AUDIO_AGENT_ROLES: [AgentRoleDefinition; 2] = [
     },
 ];
 
-static CODE_AGENT_ROLES: [AgentRoleDefinition; 4] = [
+static CODE_AGENT_ROLES: [AgentRoleDefinition; 5] = [
     AgentRoleDefinition {
         id: "director",
         role: "Director",
@@ -475,6 +484,13 @@ static CODE_AGENT_ROLES: [AgentRoleDefinition; 4] = [
         tool_id: "agent.role.brief.code.code",
         brief_path_name: "code.md",
     },
+    AgentRoleDefinition {
+        id: "review",
+        role: "Review",
+        task_id: "quality-review",
+        tool_id: "agent.role.brief.code.review",
+        brief_path_name: "review.md",
+    },
     AgentRoleDefinition {
         id: "preview",
         role: "Preview",
@@ -540,7 +556,7 @@ const GAME_CREATOR_AGENT_GROUP_DEFINITIONS: [AgentGroupDefinition; 6] = [
     AgentGroupDefinition {
         id: "code",
         label: "程序组",
-        role: "Director + Code + Preview + Playtest",
+        role: "Director + Code + Review + Preview + Playtest",
         brief_path_name: "code.md",
         roles: &CODE_AGENT_ROLES,
     },
@@ -740,10 +756,16 @@ fn get_local_game_manifest(project_path: String) -> Result<GameCreationAppManife
 
 #[tauri::command]
 async fn generate_local_game_draft(
+    app: tauri::AppHandle,
     project_path: String,
     prompt: String,
 ) -> Result<GenerateLocalGameDraftResult, String> {
-    generate_local_game_draft_at(Path::new(project_path.trim()), prompt.trim()).await
+    generate_local_game_draft_at(
+        Path::new(project_path.trim()),
+        prompt.trim(),
+        Some(&AgentProgressEmitter::new(&app, project_path.trim())),
+    )
+    .await
 }
 
 #[tauri::command]
@@ -1049,6 +1071,7 @@ fn append_agent_db_record(root: &Path, mut record: serde_json::Value) -> Result<
 async fn generate_local_game_draft_at(
     root: &Path,
     prompt: &str,
+    progress: Option<&AgentProgressEmitter<'_>>,
 ) -> Result<GenerateLocalGameDraftResult, String> {
     let prompt = prompt.trim();
     if prompt.is_empty() {
@@ -1060,14 +1083,41 @@ async fn generate_local_game_draft_at(
     let long_memory = read_optional_text(&root.join("memory/project.md"))?;
     let asset_context = render_local_asset_prompt_context(root)?;
     let long_memory = append_prompt_context(&asset_context, &long_memory);
+    emit_agent_progress(
+        progress,
+        "llm.planner",
+        "Planner 正在调用 LLM 整理规格和专业组分工",
+    );
     let client = build_game_creator_llm_client_from_env()?;
-    let loop_result =
-        run_game_creator_agent_loop_at(root, &client, prompt, &short_memory, &long_memory).await?;
+    let loop_result = run_game_creator_agent_loop_at(
+        root,
+        &client,
+        prompt,
+        &short_memory,
+        &long_memory,
+        progress,
+    )
+    .await?;
     let mut loop_result = loop_result;
+    emit_agent_progress(
+        progress,
+        "artifact.write",
+        "ArtifactWriter 正在写入本地代码、数值、美术、音乐和发布草案",
+    );
     let mut result = write_local_game_draft_at(root, prompt, &loop_result.draft)?;
     append_local_artifact_write_step(root, prompt, &mut loop_result)?;
+    emit_agent_progress(
+        progress,
+        "playtest.static_smoke",
+        "Playtest 正在运行 game.static_smoke 自检",
+    );
     append_static_smoke_step(root, prompt, &mut loop_result)?;
     append_agent_loop_log(root, &loop_result)?;
+    emit_agent_progress(
+        progress,
+        "agent.complete",
+        "Agent loop 已通过 Evaluator 和静态自检，正在启动本地预览",
+    );
     result.manifest = read_manifest_for_project(root)?;
     Ok(result)
 }
@@ -1342,6 +1392,37 @@ fn check_game_creator_llm_config_values(
     }
 }
 
+struct AgentProgressEmitter<'a> {
+    app: &'a tauri::AppHandle,
+    project_path: String,
+}
+
+impl<'a> AgentProgressEmitter<'a> {
+    fn new(app: &'a tauri::AppHandle, project_path: &str) -> Self {
+        Self {
+            app,
+            project_path: project_path.to_string(),
+        }
+    }
+
+    fn emit(&self, stage: &str, message: &str) {
+        let _ = self.app.emit(
+            "game-creator-agent-progress",
+            GameCreatorAgentProgressEvent {
+                project_path: self.project_path.clone(),
+                stage: stage.to_string(),
+                message: message.to_string(),
+            },
+        );
+    }
+}
+
+fn emit_agent_progress(progress: Option<&AgentProgressEmitter<'_>>, stage: &str, message: &str) {
+    if let Some(progress) = progress {
+        progress.emit(stage, message);
+    }
+}
+
 #[cfg(test)]
 async fn request_llm_game_draft_with_client(
     client: &LlmClient,
@@ -1368,6 +1449,7 @@ async fn run_game_creator_agent_loop_at(
     prompt: &str,
     short_memory: &str,
     long_memory: &str,
+    progress: Option<&AgentProgressEmitter<'_>>,
 ) -> Result<GameCreatorAgentLoopResult, String> {
     let spec_path = root.join(".agent/spec.md");
     let findings_path = root.join(".agent/findings.md");
@@ -1378,6 +1460,7 @@ async fn run_game_creator_agent_loop_at(
         .map(|spec| render_planner_spec(prompt, &spec))?;
     fs::write(&spec_path, &planner_spec)
         .map_err(|error| format!("写入 Planner 规格失败：{}: {error}", spec_path.display()))?;
+    emit_agent_progress(progress, "llm.planner.done", "Planner 规格已生成");
     steps.push(with_task_context(
         agent_trace_step(
             0,
@@ -1419,6 +1502,11 @@ async fn run_game_creator_agent_loop_at(
 
     let mut last_error = "Evaluator 未产出可用结果".to_string();
     for pass in 1..=GAME_CREATOR_AGENT_LOOP_MAX_PASSES {
+        emit_agent_progress(
+            progress,
+            "agent.orchestrator",
+            &format!("Orchestrator 正在规划第 {pass} 轮任务图"),
+        );
         let spec_markdown = read_optional_text(&spec_path)?;
         let findings_markdown = read_optional_text(&findings_path)?;
         let agenda = write_agent_pass_agenda(root, pass, &findings_markdown)?;
@@ -1477,22 +1565,38 @@ async fn run_game_creator_agent_loop_at(
                         findings_path.display()
                     )
                 })?;
-                steps.push(agent_trace_step(
-                    pass,
-                    "Evaluator",
-                    "needs-revision",
-                    &[],
-                    &[".agent/findings.md"],
-                    "记录专业组协作失败原因",
-                    "file.write.findings",
+                steps.push(with_task_context(
+                    agent_trace_step(
+                        pass,
+                        "Evaluator",
+                        "needs-revision",
+                        &[],
+                        &[".agent/findings.md"],
+                        "记录专业组协作失败原因",
+                        "file.write.findings",
+                    ),
+                    "code",
+                    "Review",
+                    Some("quality-review"),
+                    "evaluation",
                 ));
                 last_error = issues.join("；");
                 write_agent_run_trace(root, &run_id, prompt, "needs-revision", pass, &steps, None)?;
                 continue;
             }
         };
+        emit_agent_progress(
+            progress,
+            "agent.role_briefs",
+            &format!("6 组角色 brief 已完成，第 {pass} 轮交给 Generator"),
+        );
         append_group_brief_steps(root, pass, &agenda.relative_path, &group_briefs, &mut steps);
         let group_briefs_markdown = render_agent_group_briefs_context(&group_briefs);
+        emit_agent_progress(
+            progress,
+            "llm.generator",
+            &format!("Generator 正在调用 LLM 生成第 {pass} 轮可运行草案"),
+        );
         match request_generator_game_draft_with_client(
             client,
             prompt,
@@ -1506,6 +1610,11 @@ async fn run_game_creator_agent_loop_at(
         .await
         {
             Ok(draft) => {
+                emit_agent_progress(
+                    progress,
+                    "llm.generator.done",
+                    &format!("Generator 第 {pass} 轮草案已返回，Evaluator 开始质量评审"),
+                );
                 let pass_artifacts = write_agent_pass_artifacts(root, pass, &draft)?;
                 let mut generator_input_paths = vec![
                     "memory/session.md".to_string(),
@@ -1536,30 +1645,42 @@ async fn run_game_creator_agent_loop_at(
                         findings_path.display()
                     )
                 })?;
-                steps.push(agent_trace_step_owned(
-                    pass,
-                    "Evaluator",
-                    if issues.is_empty() {
-                        "passed"
-                    } else {
-                        "needs-revision"
-                    },
-                    vec![
-                        pass_artifacts.game_html.clone(),
-                        pass_artifacts.design_markdown.clone(),
-                        pass_artifacts.balance_json.clone(),
-                        pass_artifacts.art_manifest_json.clone(),
-                        pass_artifacts.audio_manifest_json.clone(),
-                    ],
-                    vec![".agent/findings.md".to_string()],
-                    if issues.is_empty() {
-                        "静态验收通过"
-                    } else {
-                        "发现问题，要求下一轮 Generator 修复"
-                    },
-                    "evaluator.static_html_check",
+                steps.push(with_task_context(
+                    agent_trace_step_owned(
+                        pass,
+                        "Evaluator",
+                        if issues.is_empty() {
+                            "passed"
+                        } else {
+                            "needs-revision"
+                        },
+                        vec![
+                            pass_artifacts.game_html.clone(),
+                            pass_artifacts.design_markdown.clone(),
+                            pass_artifacts.balance_json.clone(),
+                            pass_artifacts.art_manifest_json.clone(),
+                            pass_artifacts.audio_manifest_json.clone(),
+                            pass_artifacts.publish_readme.clone(),
+                        ],
+                        vec![".agent/findings.md".to_string()],
+                        if issues.is_empty() {
+                            "质量评审通过：玩法、资产、数值、程序和发布包装可进入预览试玩"
+                        } else {
+                            "质量评审发现问题，要求下一轮 Generator 修复"
+                        },
+                        "evaluator.quality_review",
+                    ),
+                    "code",
+                    "Review",
+                    Some("quality-review"),
+                    "evaluation",
                 ));
                 if issues.is_empty() {
+                    emit_agent_progress(
+                        progress,
+                        "evaluator.passed",
+                        &format!("Evaluator 第 {pass} 轮质量评审通过"),
+                    );
                     write_agent_run_trace(root, &run_id, prompt, "passed", pass, &steps, None)?;
                     return Ok(GameCreatorAgentLoopResult {
                         run_id,
@@ -1571,6 +1692,11 @@ async fn run_game_creator_agent_loop_at(
                     });
                 }
                 last_error = issues.join("；");
+                emit_agent_progress(
+                    progress,
+                    "evaluator.needs_revision",
+                    &format!("Evaluator 第 {pass} 轮要求返工：{last_error}"),
+                );
                 write_agent_run_trace(root, &run_id, prompt, "needs-revision", pass, &steps, None)?;
             }
             Err(error) => {
@@ -1591,14 +1717,20 @@ async fn run_game_creator_agent_loop_at(
                         findings_path.display()
                     )
                 })?;
-                steps.push(agent_trace_step(
-                    pass,
-                    "Evaluator",
-                    "needs-revision",
-                    &[],
-                    &[".agent/findings.md"],
-                    "记录 Generator 失败原因",
-                    "file.write.findings",
+                steps.push(with_task_context(
+                    agent_trace_step(
+                        pass,
+                        "Evaluator",
+                        "needs-revision",
+                        &[],
+                        &[".agent/findings.md"],
+                        "记录 Generator 失败原因",
+                        "file.write.findings",
+                    ),
+                    "code",
+                    "Review",
+                    Some("quality-review"),
+                    "evaluation",
                 ));
                 last_error = issues.join("；");
                 write_agent_run_trace(root, &run_id, prompt, "needs-revision", pass, &steps, None)?;
@@ -3119,11 +3251,13 @@ fn build_agent_run_task_graph_trace(
         &tasks,
         "preview-readiness",
         GameCreationAppTaskStatus::Completed,
-    ) && !steps.iter().any(|step| {
-        step.task_id.as_deref() == Some("preview-playtest")
-            && step.phase == "preview"
-            && step.status == "running"
-    }) {
+    )
+        && !steps.iter().any(|step| {
+            step.task_id.as_deref() == Some("preview-playtest")
+                && step.phase == "preview"
+                && step.status == "running"
+        })
+    {
         set_task_status_if_current(
             &mut tasks,
             "preview-playtest",
@@ -3271,7 +3405,9 @@ fn task_status_from_agent_step(
     }
 
     match step.phase.as_str() {
-        "planning" | "handoff" | "playtest" => Some(GameCreationAppTaskStatus::Completed),
+        "planning" | "handoff" | "playtest" | "evaluation" => {
+            Some(GameCreationAppTaskStatus::Completed)
+        }
         "preview" => Some(GameCreationAppTaskStatus::Running),
         "role-brief" if role_brief_completes_task(task_id) => {
             Some(GameCreationAppTaskStatus::Completed)
@@ -3284,12 +3420,12 @@ fn task_status_from_agent_step(
 fn role_brief_completes_task(task_id: &str) -> bool {
     matches!(
         task_id,
-        "balance-director"
-            | "art-director"
-            | "art-polish"
-            | "audio-director"
-            | "code-director"
-            | "publish-strategy"
+            "balance-director"
+                | "art-director"
+                | "art-polish"
+                | "audio-director"
+                | "code-director"
+                | "publish-strategy"
     )
 }
 
@@ -5300,6 +5436,7 @@ fn record_draft_task_progress(
         "audio-asset-plan",
         "code-director",
         "code-prototype",
+        "quality-review",
     ] {
         set_task_status(
             &mut manifest,
@@ -5734,7 +5871,8 @@ fn run_cli_command(command: CliCommand) -> Result<(), String> {
                 .enable_all()
                 .build()
                 .map_err(|error| format!("创建 CLI runtime 失败：{error}"))?;
-            let result = runtime.block_on(generate_local_game_draft_at(&project_path, &prompt))?;
+            let result =
+                runtime.block_on(generate_local_game_draft_at(&project_path, &prompt, None))?;
             let (preview, stop) = start_local_game_preview_for_project(&project_path)?;
             record_preview_state(
                 &project_path,
@@ -6324,7 +6462,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         std::env::set_var("GENARRATIVE_GAME_CREATOR_LLM_BASE_URL", base_url);
         std::env::set_var("GENARRATIVE_GAME_CREATOR_LLM_MODEL", "mock-game-model");
 
-        let result = generate_local_game_draft_at(&root, "用上传角色图做主角").await;
+        let result = generate_local_game_draft_at(&root, "用上传角色图做主角", None).await;
 
         restore_env("GENARRATIVE_GAME_CREATOR_LLM_API_KEY", previous_api_key);
         restore_env("GENARRATIVE_GAME_CREATOR_LLM_BASE_URL", previous_base_url);
@@ -6407,6 +6545,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
             "做一个月光厨房弹幕游戏",
             "",
             "# 项目长期记忆\n",
+            None,
         )
         .await
         .expect("agent loop");
@@ -6471,7 +6610,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         assert!(first_agenda.contains("mode: initial"));
         assert!(first_agenda.contains("activeTasks: design-director"));
         assert!(first_agenda.contains("wave 1: design-director"));
-        assert!(first_agenda.contains("wave 11: publish-package"));
+        assert!(first_agenda.contains("wave 12: publish-package"));
         let first_task_graph: Value = serde_json::from_str(
             &fs::read_to_string(root.join(".agent/passes/pass-1/task-graph.json"))
                 .expect("task graph 1"),
@@ -6620,7 +6759,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
                 .iter()
                 .any(|task_id| task_id == "code-prototype")));
         let steps = trace["steps"].as_array().unwrap();
-        assert_eq!(steps.len(), 64);
+        assert_eq!(steps.len(), 66);
         assert_eq!(steps[0]["agent"], "Planner");
         assert_eq!(steps[0]["phase"], "planning");
         assert_eq!(steps[0]["taskId"], "design-director");
@@ -6793,7 +6932,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         assert_eq!(run_history["runId"], trace["runId"]);
         assert_eq!(run_history["status"], "passed");
         assert_eq!(run_history["stopReason"], "evaluator-passed");
-        assert_eq!(run_history["steps"].as_array().unwrap().len(), 64);
+        assert_eq!(run_history["steps"].as_array().unwrap().len(), 66);
         let manifest: Value =
             serde_json::from_str(&fs::read_to_string(root.join(".agent/manifest.json")).unwrap())
                 .expect("manifest json");
@@ -6838,7 +6977,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         std::env::set_var("GENARRATIVE_GAME_CREATOR_LLM_BASE_URL", base_url);
         std::env::set_var("GENARRATIVE_GAME_CREATOR_LLM_MODEL", "mock-game-model");
 
-        let error = generate_local_game_draft_at(&root, "做一个会失败三轮的厨房游戏")
+        let error = generate_local_game_draft_at(&root, "做一个会失败三轮的厨房游戏", None)
             .await
             .expect_err("max-pass failure should bubble out");
 
@@ -6990,7 +7129,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         assert_eq!(manifest["projectId"], "project-1");
         assert_eq!(manifest["name"], "像素动作原型");
         assert_eq!(manifest["assets"].as_array().unwrap().len(), 0);
-        assert_eq!(manifest["tasks"].as_array().unwrap().len(), 15);
+        assert_eq!(manifest["tasks"].as_array().unwrap().len(), 16);
         assert_eq!(manifest["tasks"][0]["id"], "design-director");
 
         fs::remove_dir_all(root).ok();
@@ -7058,6 +7197,7 @@ GENARRATIVE_GAME_CREATOR_LLM_STREAM=true
         assert_task_status(&manifest, "audio-asset-plan", "completed");
         assert_task_status(&manifest, "code-director", "completed");
         assert_task_status(&manifest, "code-prototype", "completed");
+        assert_task_status(&manifest, "quality-review", "completed");
         assert_task_status(&manifest, "preview-readiness", "waiting-for-confirmation");
         assert_task_status(&manifest, "preview-playtest", "pending");
         assert_task_status(&manifest, "publish-strategy", "pending");

apps/ai-game-creator-shell/src/App.tsx

@@ -1,4 +1,4 @@
-import { type ChangeEvent, type FormEvent, useState } from 'react';
+import { type ChangeEvent, type FormEvent, useEffect, useState } from 'react';
 
 import {
   createGameCreationAppManifest,
@@ -133,6 +133,12 @@ interface ChatMessage {
   text: string;
 }
 
+interface AgentProgressEvent {
+  projectPath: string;
+  stage: string;
+  message: string;
+}
+
 export type PendingCommand =
   | {
       id: 'game.generate_draft';
@@ -767,7 +773,24 @@ function summarizeSuggestedToolCalls(trace: GameCreationAgentRunTrace) {
     .join('\n');
 }
 
+function summarizeLlmConversation(trace: GameCreationAgentRunTrace) {
+  return trace.steps
+    .filter((step) =>
+      step.toolCalls.some((toolCall) => toolCall.toolId.startsWith('llm.')),
+    )
+    .slice(-6)
+    .map((step) => {
+      const toolIds = step.toolCalls
+        .filter((toolCall) => toolCall.toolId.startsWith('llm.'))
+        .map((toolCall) => toolCall.toolId)
+        .join(', ');
+      return `- ${step.agent} #${step.pass} · ${step.status} · ${step.phase} · ${toolIds}`;
+    })
+    .join('\n');
+}
+
 export function summarizeAgentRunTrace(trace: GameCreationAgentRunTrace) {
+  const llmConversation = summarizeLlmConversation(trace);
   const recentSteps = trace.steps
     .slice(-5)
     .map(
@@ -833,6 +856,7 @@ export function summarizeAgentRunTrace(trace: GameCreationAgentRunTrace) {
     `状态：${trace.status} · ${trace.passes}/${trace.maxPasses} 轮 · ${trace.stopReason}`,
     `工具调用：${trace.toolCallCount}/${trace.maxToolCalls}`,
     `下一步：${trace.nextStep}`,
+    llmConversation ? `LLM 对话：\n${llmConversation}` : null,
     taskSummary ? `任务：${taskSummary}` : null,
     `active 任务：${formatTraceTaskIds(
       trace.taskGraph.activeTaskIds,
@@ -855,43 +879,18 @@ export function summarizeAgentRunTrace(trace: GameCreationAgentRunTrace) {
     .join('\n');
 }
 
-export function summarizeAgentRunCompletion(trace: GameCreationAgentRunTrace) {
-  const repair =
-    trace.taskGraph.repairFocus.length > 0
-      ? `\n返工焦点：${trace.taskGraph.repairFocus.join('；')}`
-      : '';
-  const artifacts = trace.artifacts
-    .filter((artifact) =>
-      [
-        'game/index.html',
-        'game/game_design.md',
-        'game/balance.json',
-        'assets/manifest.art.json',
-        'assets/manifest.audio.json',
-        'exports/README.md',
-      ].includes(artifact.path),
-    )
-    .map((artifact) => artifact.path)
-    .join('，');
-  const suggestedTools = summarizeSuggestedToolCalls(trace);
+function summarizeAgentRunCompletionForChat(
+  trace: GameCreationAgentRunTrace,
+) {
+  return `${summarizeAgentRunTrace(trace)}\n完整 trace：/trace`;
+}
 
+function gameDraftStartedMessage() {
   return [
-    `Agent loop：${trace.status} · ${trace.passes}/${trace.maxPasses} 轮 · ${trace.stopReason}`,
-    `工具调用：${trace.toolCallCount}/${trace.maxToolCalls}`,
-    `active 任务：${formatTraceTaskIds(
-      trace.taskGraph.activeTaskIds,
-      trace.taskGraph.tasks,
-    )}`,
-    `carry-over 任务：${formatTraceTaskIds(
-      trace.taskGraph.carriedTaskIds,
-      trace.taskGraph.tasks,
-    )}${repair}`,
-    suggestedTools ? `建议命令：\n${suggestedTools}` : null,
-    artifacts ? `本地产物：${artifacts}` : null,
-    '完整 trace：/trace',
-  ]
-    .filter(Boolean)
-    .join('\n');
+    '开始调用 LLM：Planner 正在整理规格。',
+    '随后 Orchestrator 编排 6 组角色 brief，Generator 生成代码和资产清单，Evaluator 做质量评审。',
+    '完成后会把 run trace 和本地产物摘要发回这里。',
+  ].join('\n');
 }
 
 function pendingCommandTitle(command: PendingCommand) {
@@ -942,7 +941,7 @@ export function pendingCommandDetail(
     return `运行自检并启动 ${projectPath}/game/`;
   }
   if (command.id === 'game.generate_draft') {
-    return `调用 LLM，写入 ${projectPath}/game、assets、memory、exports，通过自检后启动本地 HTTP 预览并交给外部浏览器`;
+    return `调用 LLM Planner / Generator，编排 6 组角色 brief，写入 ${projectPath}/game、assets、memory、exports，通过 Evaluator 和自检后启动本地 HTTP 预览并交给外部浏览器`;
   }
   if (command.id === 'project.create') {
     return `创建 ${command.projectPath}`;
@@ -1028,6 +1027,37 @@ export function App() {
     `${GAME_CREATION_APP_COMMANDS.length} 个命令已登记权限。`,
   ]);
 
+  useEffect(() => {
+    const listen = window.__TAURI__?.event?.listen;
+    if (!listen) {
+      return;
+    }
+    let cleanup: (() => void) | null = null;
+    let disposed = false;
+    void listen<AgentProgressEvent>('game-creator-agent-progress', (event) => {
+      if (
+        localProject &&
+        event.payload.projectPath !== localProject.projectPath
+      ) {
+        return;
+      }
+      setMessages((current) => [
+        ...current,
+        { role: 'assistant', text: event.payload.message },
+      ]);
+    }).then((unlisten) => {
+      if (disposed) {
+        unlisten();
+        return;
+      }
+      cleanup = unlisten;
+    });
+    return () => {
+      disposed = true;
+      cleanup?.();
+    };
+  }, [localProject?.projectPath]);
+
   function appendLocalPermissionLog(
     projectPath: string | null,
     event: 'permission.pending' | 'permission.confirm' | 'permission.cancel',
@@ -1534,6 +1564,11 @@ export function App() {
 
     try {
       setCommandLog((current) => [...current, 'game.generate_draft']);
+      setProjectStatus('正在调用 LLM');
+      setMessages((current) => [
+        ...current,
+        { role: 'assistant', text: gameDraftStartedMessage() },
+      ]);
       const result = await invoke<GenerateLocalGameDraftResult>(
         'generate_local_game_draft',
         { projectPath: nextProjectPath, prompt },
@@ -1544,7 +1579,7 @@ export function App() {
         manifest: result.manifest,
       });
       setManifest(result.manifest);
-      setProjectStatus('已生成草案');
+      setProjectStatus('已生成可试玩原型');
       setCommandLog((current) => [
         ...current,
         'memory.write',
@@ -3253,7 +3288,7 @@ export function App() {
         `${trace.status} · ${trace.passes}/${trace.maxPasses} 轮 · ${trace.stopReason}`,
       );
       setCommandLog((current) => [...current, `file.read ${relativePath}`]);
-      return summarizeAgentRunCompletion(trace);
+      return summarizeAgentRunCompletionForChat(trace);
     } catch (error) {
       setAgentRunTrace(null);
       setAgentRunStatus(error instanceof Error ? error.message : String(error));

apps/ai-game-creator-shell/src/vite-env.d.ts

@@ -8,5 +8,11 @@ interface Window {
         args?: Record<string, unknown>,
       ) => Promise<Result>;
     };
+    event?: {
+      listen?: <Payload = unknown>(
+        event: string,
+        handler: (event: { payload: Payload }) => void,
+      ) => Promise<() => void>;
+    };
   };
 }

apps/ai-game-creator-shell/tests/agentTraceSummary.test.ts

@@ -79,13 +79,18 @@ describe('AI 游戏创作 Agent loop 摘要', () => {
       taskGraph: {
         goal: '做一个弹幕厨房游戏',
         readyTaskIds: ['code-director'],
-        activeTaskIds: ['code-director', 'publish-package'],
+        activeTaskIds: ['code-director', 'quality-review', 'publish-package'],
         carriedTaskIds: ['design-director'],
         repairFocus: ['gameHtml 缺少 canvas'],
         repairRoutes: [
           {
             issue: 'gameHtml 缺少 canvas',
-            taskIds: ['code-director', 'preview-readiness', 'publish-package'],
+            taskIds: [
+              'code-director',
+              'quality-review',
+              'preview-readiness',
+              'publish-package',
+            ],
             reason: 'code-repair+dependency-impact',
           },
         ],
@@ -106,14 +111,23 @@ describe('AI 游戏创作 Agent loop 摘要', () => {
           pass: 2,
           mode: 'repair',
           summary: '第 2 轮重跑程序链路及下游发布包装',
-          activeTaskIds: ['code-director', 'publish-package'],
+          activeTaskIds: ['code-director', 'quality-review', 'publish-package'],
           carriedTaskIds: ['design-director'],
-          dependencyWaves: [['code-director'], ['publish-package']],
+          dependencyWaves: [
+            ['code-director'],
+            ['quality-review'],
+            ['publish-package'],
+          ],
           repairFocus: ['gameHtml 缺少 canvas'],
           repairRoutes: [
             {
               issue: 'gameHtml 缺少 canvas',
-              taskIds: ['code-director', 'preview-readiness', 'publish-package'],
+              taskIds: [
+                'code-director',
+                'quality-review',
+                'preview-readiness',
+                'publish-package',
+              ],
               reason: 'code-repair+dependency-impact',
             },
           ],
@@ -130,21 +144,21 @@ describe('AI 游戏创作 Agent loop 摘要', () => {
     expect(summary).toContain('工具调用：12/128');
     expect(summary).toContain('任务：已完成 2');
     expect(summary).toContain(
-      'active 任务：程序组 / Director 拆解程序实现(code-director), 运营组 / Publish 整理发布包装(publish-package)',
+      'active 任务：程序组 / Director 拆解程序实现(code-director), 程序组 / Review 执行质量评审(quality-review), 运营组 / Publish 整理发布包装(publish-package)',
     );
     expect(summary).toContain(
       'carry-over 任务：策划组 / Director 拆解创作方向(design-director)',
     );
     expect(summary).toContain('返工焦点：gameHtml 缺少 canvas');
     expect(summary).toContain(
-      '返工路线：code-repair+dependency-impact: 程序组 / Director 拆解程序实现(code-director), 程序组 / Preview 执行静态自检(preview-readiness), 运营组 / Publish 整理发布包装(publish-package)',
+      '返工路线：code-repair+dependency-impact: 程序组 / Director 拆解程序实现(code-director), 程序组 / Review 执行质量评审(quality-review), 程序组 / Preview 执行静态自检(preview-readiness), 运营组 / Publish 整理发布包装(publish-package)',
     );
     expect(summary).toContain('建议命令：');
     expect(summary).toContain('agent.tool.suggest.canvas.project_sync');
     expect(summary).toContain('/sync-canvas-project <画板项目ID>');
     expect(summary).toContain('编排轮次：');
     expect(summary).toContain(
-      'pass 2 · repair · active 2 · carry 1 · waves 程序组 / Director 拆解程序实现(code-director) / 运营组 / Publish 整理发布包装(publish-package)',
+      'pass 2 · repair · active 3 · carry 1 · waves 程序组 / Director 拆解程序实现(code-director) / 程序组 / Review 执行质量评审(quality-review) / 运营组 / Publish 整理发布包装(publish-package)',
     );
     expect(summary).toContain('.agent/passes/pass-2/task-graph.json');
     expect(summary).toContain('Orchestrator #2 · completed · plan');

apps/ai-game-creator-shell/tests/appSurface.test.ts

@@ -1,5 +1,5 @@
 /** @vitest-environment jsdom */
-import { cleanup, fireEvent, render, screen } from '@testing-library/react';
+import { act, cleanup, fireEvent, render, screen } from '@testing-library/react';
 import { readFileSync } from 'node:fs';
 import { resolve } from 'node:path';
 import React from 'react';
@@ -232,9 +232,9 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 2,
           agent: 'Evaluator',
           phase: 'evaluate',
-          taskId: 'preview-readiness',
+          taskId: 'quality-review',
           group: 'code',
-          role: 'Preview',
+          role: 'Review',
           status: 'passed',
           inputPaths: ['.agent/passes/pass-2/draft.json'],
           outputPaths: ['.agent/findings.md'],
@@ -296,7 +296,7 @@ describe('AI 游戏创作 App 界面边界', () => {
         repairRoutes: [
           {
             issue: '缺少输入监听',
-            taskIds: ['code-prototype', 'preview-readiness'],
+            taskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
             reason: 'code-runtime',
           },
         ],
@@ -317,14 +317,18 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 2,
           mode: 'repair',
           summary: '第 2 轮返工',
-          activeTaskIds: ['code-prototype', 'preview-readiness'],
+          activeTaskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
           carriedTaskIds: ['design-director'],
-          dependencyWaves: [['code-prototype'], ['preview-readiness']],
+          dependencyWaves: [
+            ['code-prototype'],
+            ['quality-review'],
+            ['preview-readiness'],
+          ],
           repairFocus: ['缺少输入监听'],
           repairRoutes: [
             {
               issue: '缺少输入监听',
-              taskIds: ['code-prototype', 'preview-readiness'],
+              taskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
               reason: 'code-runtime',
             },
           ],
@@ -394,7 +398,7 @@ describe('AI 游戏创作 App 界面边界', () => {
     expect(screen.getByText(/能力\/命令契约：通过/)).not.toBeNull();
     expect(screen.getByText(/6 组任务配置：通过/)).not.toBeNull();
     expect(screen.getByText(/6 组协作证据：通过/)).not.toBeNull();
-    expect(screen.getByText(/角色任务 15 个/)).not.toBeNull();
+    expect(screen.getByText(/角色任务 16 个/)).not.toBeNull();
     expect(
       screen.getByText(/Loop trace：通过 · run run-chat-audit/),
     ).not.toBeNull();
@@ -646,9 +650,9 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 3,
           agent: 'Evaluator',
           phase: 'evaluate',
-          taskId: 'preview-readiness',
+          taskId: 'quality-review',
           group: 'code',
-          role: 'Preview',
+          role: 'Review',
           status: 'needs-revision',
           inputPaths: ['.agent/passes/pass-3/draft.json'],
           outputPaths: ['.agent/findings.md'],
@@ -679,13 +683,13 @@ describe('AI 游戏创作 App 界面边界', () => {
       taskGraph: {
         goal: '做一个三轮仍失败的厨房弹幕游戏',
         readyTaskIds: [],
-        activeTaskIds: ['code-prototype', 'preview-readiness'],
+        activeTaskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
         carriedTaskIds: ['design-director'],
         repairFocus: ['缺少输入监听'],
         repairRoutes: [
           {
             issue: '缺少输入监听',
-            taskIds: ['code-prototype', 'preview-readiness'],
+            taskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
             reason: 'code-runtime',
           },
         ],
@@ -696,14 +700,18 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 3,
           mode: 'repair',
           summary: '第 3 轮返工仍未通过',
-          activeTaskIds: ['code-prototype', 'preview-readiness'],
+          activeTaskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
           carriedTaskIds: ['design-director'],
-          dependencyWaves: [['code-prototype'], ['preview-readiness']],
+          dependencyWaves: [
+            ['code-prototype'],
+            ['quality-review'],
+            ['preview-readiness'],
+          ],
           repairFocus: ['缺少输入监听'],
           repairRoutes: [
             {
               issue: '缺少输入监听',
-              taskIds: ['code-prototype', 'preview-readiness'],
+              taskIds: ['code-prototype', 'quality-review', 'preview-readiness'],
               reason: 'code-runtime',
             },
           ],
@@ -1294,7 +1302,7 @@ describe('AI 游戏创作 App 界面边界', () => {
 
     expect(await screen.findByText(/项目：未命名游戏原型/)).not.toBeNull();
     expect(screen.getByText(/目录：\/tmp\/authorized-game/)).not.toBeNull();
-    expect(screen.getByText(/任务：已完成 1，待处理 14/)).not.toBeNull();
+    expect(screen.getByText(/任务：已完成 1，待处理 15/)).not.toBeNull();
     expect(screen.getByText(/资产：1 个/)).not.toBeNull();
     expect(
       screen.getByText(/预览：运行中 http:\/\/127\.0\.0\.1:3210\//),
@@ -1420,11 +1428,77 @@ describe('AI 游戏创作 App 界面边界', () => {
     expect(screen.getByText('game.generate_draft')).not.toBeNull();
     expect(
       screen.getByText(
-        '调用 LLM，写入 /tmp/authorized-game/game、assets、memory、exports，通过自检后启动本地 HTTP 预览并交给外部浏览器',
+        '调用 LLM Planner / Generator，编排 6 组角色 brief，写入 /tmp/authorized-game/game、assets、memory、exports，通过 Evaluator 和自检后启动本地 HTTP 预览并交给外部浏览器',
       ),
     ).not.toBeNull();
   });
 
+  it('streams agent generation progress into the user chat', async () => {
+    const manifest = createGameCreationAppManifest(
+      'local-project-draft',
+      '未命名游戏原型',
+    );
+    let progressHandler:
+      | ((event: {
+          payload: { projectPath: string; stage: string; message: string };
+        }) => void)
+      | null = null;
+    const invoke = vi.fn(
+      async (command: string, args?: Record<string, unknown>) => {
+        if (command === 'append_local_permission_log') {
+          return {};
+        }
+        if (command !== 'init_local_game_project') {
+          throw new Error(`unexpected invoke ${command}`);
+        }
+        const projectPath = String(args?.projectPath ?? '');
+        return {
+          projectPath,
+          manifestPath: `${projectPath}/.agent/manifest.json`,
+          manifest,
+        };
+      },
+    );
+    const listen = vi.fn(async (_event: string, handler: typeof progressHandler) => {
+      progressHandler = handler;
+      return vi.fn();
+    });
+    window.__TAURI__ = { core: { invoke }, event: { listen } };
+    renderAppAt('/');
+
+    submitChat('/project /tmp/authorized-game');
+    fireEvent.click(screen.getByRole('button', { name: '确认' }));
+    expect(
+      await screen.findByText('已设置本地项目：/tmp/authorized-game'),
+    ).not.toBeNull();
+
+    await act(async () => {
+      progressHandler?.({
+        payload: {
+          projectPath: '/tmp/other-game',
+          stage: 'llm.planner',
+          message: '不应该显示',
+        },
+      });
+      progressHandler?.({
+        payload: {
+          projectPath: '/tmp/authorized-game',
+          stage: 'llm.planner',
+          message: 'Planner 正在调用 LLM 整理规格和专业组分工',
+        },
+      });
+    });
+
+    expect(listen).toHaveBeenCalledWith(
+      'game-creator-agent-progress',
+      expect.any(Function),
+    );
+    expect(screen.queryByText('不应该显示')).toBeNull();
+    expect(
+      screen.getByText('Planner 正在调用 LLM 整理规格和专业组分工'),
+    ).not.toBeNull();
+  });
+
   it('shows agent loop evidence in chat after generation completes', async () => {
     const manifest = createGameCreationAppManifest(
       'local-project-draft',
@@ -1449,6 +1523,77 @@ describe('AI 游戏创作 App 界面边界', () => {
       goal: '做一个反弹弹幕厨房游戏',
       coordination: 'filesystem',
       steps: [
+        {
+          pass: 0,
+          agent: 'Planner',
+          phase: 'planning',
+          taskId: 'design-director',
+          group: 'design',
+          role: 'Director',
+          status: 'completed',
+          inputPaths: [
+            'memory/session.md',
+            'memory/project.md',
+            '.agent/manifest.json',
+          ],
+          outputPaths: ['.agent/spec.md'],
+          summary: '完成玩法规格和专业组分工',
+          toolCalls: [
+            {
+              toolId: 'llm.chat.planner',
+              status: 'completed',
+              inputPaths: ['memory/session.md', 'memory/project.md'],
+              outputPaths: ['.agent/spec.md'],
+              summary: 'Planner 规格生成',
+            },
+          ],
+        },
+        {
+          pass: 2,
+          agent: 'Generator',
+          phase: 'generate',
+          taskId: 'code-prototype',
+          group: 'code',
+          role: 'Code',
+          status: 'completed',
+          inputPaths: [
+            '.agent/spec.md',
+            '.agent/findings.md',
+            '.agent/passes/pass-2/agenda.md',
+          ],
+          outputPaths: ['.agent/passes/pass-2/draft.json'],
+          summary: '生成结构化游戏草案',
+          toolCalls: [
+            {
+              toolId: 'llm.chat.generator',
+              status: 'completed',
+              inputPaths: ['.agent/spec.md', '.agent/findings.md'],
+              outputPaths: ['.agent/passes/pass-2/draft.json'],
+              summary: 'Generator 草案生成',
+            },
+          ],
+        },
+        {
+          pass: 2,
+          agent: 'Evaluator',
+          phase: 'evaluation',
+          taskId: 'quality-review',
+          group: 'code',
+          role: 'Review',
+          status: 'passed',
+          inputPaths: ['.agent/passes/pass-2/draft.json'],
+          outputPaths: ['.agent/findings.md'],
+          summary: '质量评审通过',
+          toolCalls: [
+            {
+              toolId: 'evaluator.quality_review',
+              status: 'completed',
+              inputPaths: ['.agent/passes/pass-2/draft.json'],
+              outputPaths: ['.agent/findings.md'],
+              summary: 'Evaluator 质量评审',
+            },
+          ],
+        },
         {
           pass: 2,
           agent: '美术组 / Asset',
@@ -1492,13 +1637,13 @@ describe('AI 游戏创作 App 界面边界', () => {
       taskGraph: {
         goal: '做一个反弹弹幕厨房游戏',
         readyTaskIds: ['preview-playtest'],
-        activeTaskIds: ['code-director', 'preview-readiness'],
+        activeTaskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
         carriedTaskIds: ['design-director', 'design-foundation'],
         repairFocus: ['缺少输入监听'],
         repairRoutes: [
           {
             issue: '缺少输入监听',
-            taskIds: ['code-director', 'preview-readiness'],
+            taskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
             reason: 'code-runtime',
           },
         ],
@@ -1509,14 +1654,19 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 2,
           mode: 'repair',
           summary: '第 2 轮按 Evaluator 反馈返工',
-          activeTaskIds: ['code-director', 'preview-readiness'],
+          activeTaskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
           carriedTaskIds: ['design-director', 'design-foundation'],
-          dependencyWaves: [['code-director'], ['preview-readiness']],
+          dependencyWaves: [
+            ['code-director'],
+            ['code-prototype'],
+            ['quality-review'],
+            ['preview-readiness'],
+          ],
           repairFocus: ['缺少输入监听'],
           repairRoutes: [
             {
               issue: '缺少输入监听',
-              taskIds: ['code-director', 'preview-readiness'],
+              taskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
               reason: 'code-runtime',
             },
           ],
@@ -1609,17 +1759,23 @@ describe('AI 游戏创作 App 界面边界', () => {
     submitChat('做一个反弹弹幕厨房游戏');
     fireEvent.click(screen.getByRole('button', { name: '确认' }));
 
+    expect(screen.getByText(/开始调用 LLM：Planner 正在整理规格。/)).not.toBeNull();
+    expect(screen.getByText(/Generator 生成代码和资产清单/)).not.toBeNull();
     expect(
       await screen.findByText(/已保存并启动本地预览：http:\/\/127\.0\.0\.1:3210\//),
     ).not.toBeNull();
     expect(screen.getByText(/已交给外部浏览器打开。/)).not.toBeNull();
+    expect(screen.getByText(/Run：run-chat-generate/)).not.toBeNull();
     expect(
-      screen.getByText(/Agent loop：passed · 2\/3 轮 · evaluator-passed/),
+      screen.getByText(/状态：passed · 2\/3 轮 · evaluator-passed/),
     ).not.toBeNull();
     expect(screen.getByText(/工具调用：36\/128/)).not.toBeNull();
+    expect(screen.getByText(/LLM 对话：/)).not.toBeNull();
+    expect(screen.getByText(/Planner #0 · completed · planning · llm\.chat\.planner/)).not.toBeNull();
+    expect(screen.getByText(/Generator #2 · completed · generate · llm\.chat\.generator/)).not.toBeNull();
     expect(
       screen.getByText(
-        /active 任务：程序组 \/ Director 拆解程序实现\(code-director\), 程序组 \/ Preview 执行静态自检\(preview-readiness\)/,
+        /active 任务：程序组 \/ Director 拆解程序实现\(code-director\), 程序组 \/ Code 生成可运行原型\(code-prototype\), 程序组 \/ Review 执行质量评审\(quality-review\), 程序组 \/ Preview 执行静态自检\(preview-readiness\)/,
       ),
     ).not.toBeNull();
     expect(screen.getByText(/carry-over 任务：策划组/)).not.toBeNull();
@@ -1630,11 +1786,11 @@ describe('AI 游戏创作 App 界面边界', () => {
     expect(
       screen.getByText(/\/sync-canvas-project <画板项目ID>/),
     ).not.toBeNull();
-    expect(
-      screen.getByText(
-        /本地产物：game\/index\.html，game\/game_design\.md，assets\/manifest\.art\.json/,
-      ),
-    ).not.toBeNull();
+    expect(screen.getByText(/编排轮次：/)).not.toBeNull();
+    expect(screen.getByText(/产物快照：/)).not.toBeNull();
+    expect(screen.getByText(/- game\/index\.html · fnv1a64:game/)).not.toBeNull();
+    expect(screen.getByText(/最近步骤：/)).not.toBeNull();
+    expect(screen.getByText(/Evaluator #2 · passed · evaluation/)).not.toBeNull();
     expect(screen.getByText(/完整 trace：\/trace/)).not.toBeNull();
     expect(screen.queryByLabelText('开发环境')).toBeNull();
     expect(screen.queryByText('编排 Trace')).toBeNull();
@@ -1896,13 +2052,13 @@ describe('AI 游戏创作 App 界面边界', () => {
       taskGraph: {
         goal: '做一个反弹弹幕厨房游戏',
         readyTaskIds: ['code-director'],
-        activeTaskIds: ['code-director', 'preview-readiness'],
+        activeTaskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
         carriedTaskIds: ['design-director'],
         repairFocus: ['缺少输入监听'],
         repairRoutes: [
           {
             issue: '缺少输入监听',
-            taskIds: ['code-director', 'preview-readiness'],
+            taskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
             reason: 'code-repair',
           },
         ],
@@ -1913,14 +2069,19 @@ describe('AI 游戏创作 App 界面边界', () => {
           pass: 2,
           mode: 'repair',
           summary: '第 2 轮按 Evaluator 反馈返工',
-          activeTaskIds: ['code-director', 'preview-readiness'],
+          activeTaskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
           carriedTaskIds: ['design-director'],
-          dependencyWaves: [['code-director'], ['preview-readiness']],
+          dependencyWaves: [
+            ['code-director'],
+            ['code-prototype'],
+            ['quality-review'],
+            ['preview-readiness'],
+          ],
           repairFocus: ['缺少输入监听'],
           repairRoutes: [
             {
               issue: '缺少输入监听',
-              taskIds: ['code-director', 'preview-readiness'],
+              taskIds: ['code-director', 'code-prototype', 'quality-review', 'preview-readiness'],
               reason: 'code-repair',
             },
           ],
@@ -1970,7 +2131,7 @@ describe('AI 游戏创作 App 界面边界', () => {
     ).not.toBeNull();
     expect(
       screen.getByText(
-        /active 任务：程序组 \/ Director 拆解程序实现\(code-director\), 程序组 \/ Preview 执行静态自检\(preview-readiness\)/,
+        /active 任务：程序组 \/ Director 拆解程序实现\(code-director\), 程序组 \/ Code 生成可运行原型\(code-prototype\), 程序组 \/ Review 执行质量评审\(quality-review\), 程序组 \/ Preview 执行静态自检\(preview-readiness\)/,
       ),
     ).not.toBeNull();
     expect(screen.getByText(/返工路线：code-repair/)).not.toBeNull();

apps/ai-game-creator-shell/tests/rememberCommand.test.ts

@@ -95,7 +95,7 @@ describe('AI 游戏创作聊天记忆命令', () => {
         '/tmp/game',
       ),
     ).toBe(
-      '调用 LLM，写入 /tmp/game/game、assets、memory、exports，通过自检后启动本地 HTTP 预览并交给外部浏览器',
+      '调用 LLM Planner / Generator，编排 6 组角色 brief，写入 /tmp/game/game、assets、memory、exports，通过 Evaluator 和自检后启动本地 HTTP 预览并交给外部浏览器',
     );
   });
 

docs/project-memory/shared-memory/decision-log.md

@@ -16,6 +16,22 @@
 
 ---
 
+## 2026-06-26 AI 游戏创作 App 生成过程必须在聊天可见
+
+- 背景：普通用户窗口只保留聊天入口，但如果生成确认后只显示“已生成草案”和本地产物路径，真实 LLM / Agent loop 会被误解成固定模板落盘。
+- 决策：`game.generate_draft` 保持正式用户窗口不展示开发面板，但必须通过聊天实时显示 Planner LLM、Orchestrator、6 组角色 brief、Generator LLM、Evaluator、ArtifactWriter 和自检进度；生成完成后普通聊天消息直接展示 `.agent/run.latest.json` 的 Run、LLM 对话、loop 轮次、active / carry-over 任务、编排轮次、最近步骤、建议命令和本地产物快照，`/trace` 继续读取同一份完整证据。
+- 影响范围：`apps/ai-game-creator-shell/src/App.tsx`、`apps/ai-game-creator-shell/src-tauri/src/main.rs`、AI 游戏创作 App 聊天体验和实施计划文档。
+- 验证方式：运行 `npm run ai-game-creator-shell:check`、`npm run check:encoding` 和 `git diff --check`。
+- 关联文档：`docs/technical/【技术方案】AI游戏创作智能体App实施计划-2026-06-24.md`。
+
+## 2026-06-26 AI 游戏创作 App 增加显式质量评审 Gate
+
+- 背景：AI 游戏创作 App 已有 Evaluator loop 和静态 smoke，但任务图、能力清单和 trace 中没有单独的质检 / 评审任务，用户无法从 `/tasks`、`/trace` 或 `/audit` 看出质量评审是明确环节。
+- 决策：保持策划、美术、程序、数值、音乐、运营 6 个专业组不变，在程序组内新增 `quality-review` / `Review` 角色任务；Evaluator 的评审 step 绑定到该任务，依赖顺序为 `code-prototype -> quality-review -> preview-readiness -> preview-playtest -> publish-strategy -> publish-package`。`game.static_smoke` 只完成 `preview-readiness`，不代替质量评审。
+- 影响范围：AI 游戏创作 App 任务图、共享契约、Tauri trace / manifest 状态推导、聊天 `/capabilities` `/tasks` `/trace` `/audit` 摘要和实施计划文档。
+- 验证方式：运行 `npm run ai-game-creator-shell:check`、`npm run check:encoding` 和 `git diff --check`。
+- 关联文档：`docs/technical/【技术方案】AI游戏创作智能体App实施计划-2026-06-24.md`。
+
 ## 2026-06-25 AI 游戏创作 App 真实 LLM 联调用流式请求
 
 - 背景：AI 游戏创作 App 的真实 OpenAI-compatible provider 验收中，小请求可返回，但 Planner 等稍长非流式请求会在上游响应前被网关空闲连接切断，表现为 TLS record 解密失败；本地无密钥 provider smoke 不能覆盖该真实网关行为。

docs/technical/【技术方案】AI游戏创作智能体App实施计划-2026-06-24.md

@@ -86,7 +86,7 @@ game-project/
 
 ## v1 验收证据矩阵
 
-- `npm run ai-game-creator-shell:check`：覆盖壳 typecheck、聊天命令单测、用户 / 开发窗口 UI 边界 smoke、聊天侧 `/capabilities` 展示标准 Agent 能力清单且不打开开发面板、聊天侧 `/audit` 从 manifest / 本地文件 / `.agent/run.latest.json` 分别汇总用户面、6 组任务配置、6 组协作证据、任务编排、loop、记忆、本地产物、HTTP 预览、画板回流和权限日志证据且不打开开发面板；未生成 `.agent/run.latest.json` 前，`/audit` 只能标记任务配置通过，不能把 6 组协作证据误判为通过；trace 已存在但状态为 `failed`、`needs-revision`、`running`、`max-passes-exhausted` 或缺少 `Evaluator passed` 步骤时，`/audit` 不能把 loop 误判为通过。聊天侧 `/llm-status` 只显示 base_url / model / API Key 已读取状态且不泄露密钥本体、聊天侧长期记忆查看 / 追加 / 覆盖 / 删除的授权本地项目路径、上传资产写入后的 manifest 刷新和 `/assets` 聊天可见性、`/smoke` 聊天侧确认后只通过授权本地项目路径执行白名单 `game.static_smoke`、`/run` 聊天侧确认后通过授权本地项目路径执行 `game.static_smoke`、启动 `127.0.0.1` 本地预览并交给外部浏览器、`/preview` 聊天侧确认后通过授权本地项目路径启动 `127.0.0.1` 本地预览并交给外部浏览器、`/status` 聊天侧项目 / 任务 / 资产 / 预览 / 最近命令汇总、`/files` 聊天侧本地文件列表、`/read` 聊天侧文件读取的授权本地项目路径、`/tasks` 聊天侧任务拆分与下一步专业组展示的授权本地项目路径、聊天确认生成后自动读取 `.agent/run.latest.json` 并在普通聊天消息里展示 loop 轮次 / 工具调用 / active 任务 / carry-over / 返工焦点 / 画板同步建议命令 / 本地产物、`/trace` 聊天侧读取 `.agent/run.latest.json` 并展示 loop 轮次 / active 任务 / 返工路线 / dependency waves 的授权本地项目路径、`platform-agent` 编排测试、共享契约测试、Tauri 本地能力测试和无密钥本地 provider 端到端 smoke；用于证明独立 App、真实 LLM-compatible loop、本地落盘、自检和 HTTP 预览闭环，并覆盖 loop 跑满 3 轮失败时不会写入最终游戏产物。
+- `npm run ai-game-creator-shell:check`：覆盖壳 typecheck、聊天命令单测、用户 / 开发窗口 UI 边界 smoke、聊天侧 `/capabilities` 展示标准 Agent 能力清单且不打开开发面板、聊天侧 `/audit` 从 manifest / 本地文件 / `.agent/run.latest.json` 分别汇总用户面、6 组任务配置、6 组协作证据、任务编排、loop、记忆、本地产物、HTTP 预览、画板回流和权限日志证据且不打开开发面板；未生成 `.agent/run.latest.json` 前，`/audit` 只能标记任务配置通过，不能把 6 组协作证据误判为通过；trace 已存在但状态为 `failed`、`needs-revision`、`running`、`max-passes-exhausted` 或缺少 `Evaluator passed` 步骤时，`/audit` 不能把 loop 误判为通过。聊天侧 `/llm-status` 只显示 base_url / model / API Key 已读取状态且不泄露密钥本体、聊天侧长期记忆查看 / 追加 / 覆盖 / 删除的授权本地项目路径、上传资产写入后的 manifest 刷新和 `/assets` 聊天可见性、`/smoke` 聊天侧确认后只通过授权本地项目路径执行白名单 `game.static_smoke`、`/run` 聊天侧确认后通过授权本地项目路径执行 `game.static_smoke`、启动 `127.0.0.1` 本地预览并交给外部浏览器、`/preview` 聊天侧确认后通过授权本地项目路径启动 `127.0.0.1` 本地预览并交给外部浏览器、`/status` 聊天侧项目 / 任务 / 资产 / 预览 / 最近命令汇总、`/files` 聊天侧本地文件列表、`/read` 聊天侧文件读取的授权本地项目路径、`/tasks` 聊天侧任务拆分与下一步专业组展示的授权本地项目路径、聊天确认生成后实时展示 Planner / Orchestrator / 角色 brief / Generator / Evaluator / 写盘 / 自检进度，并自动读取 `.agent/run.latest.json` 在普通聊天消息里展示 Run、LLM 对话、loop 轮次、工具调用、active / carry-over 任务、返工焦点、编排轮次、最近步骤、画板同步建议命令和本地产物快照、`/trace` 聊天侧读取 `.agent/run.latest.json` 并展示 loop 轮次 / active 任务 / 返工路线 / dependency waves 的授权本地项目路径、`platform-agent` 编排测试、共享契约测试、Tauri 本地能力测试和无密钥本地 provider 端到端 smoke；用于证明独立 App、真实 LLM-compatible loop、本地落盘、自检和 HTTP 预览闭环，并覆盖 loop 跑满 3 轮失败时不会写入最终游戏产物。
 - `npm run check:native-shells`：覆盖 AI 游戏创作壳的 release/dev 窗口边界、正式用户 App 不嵌入游戏预览 iframe、用户侧预览命令交给外部浏览器和 Tauri release `--no-bundle` 构建 smoke；用于证明正式用户窗口只登记 `main` 聊天窗口，开发面板只在 debug/dev 路径打开，独立壳能完成 release 编译。
 - `npm run check:encoding` 与 `git diff --check`：覆盖中文文档、中文命令文案和补丁空白；用于避免乱码、尾随空白和无关格式漂移。
 - `npm run ai-game-creator-shell:llm-status`：只检查 LLM 环境变量是否就绪，不请求上游、不显示 API Key；用于本机联调前确认配置。CLI 和桌面 App 内的 `/llm-status` / 生成入口都会先读取仓库根目录或 `apps/ai-game-creator-shell/` 下 gitignored 的 `.env.secrets.local`，再检查当前进程环境。
@@ -100,17 +100,17 @@ game-project/
 - 普通用户侧的生成、上传、运行、自检、预览状态 / 启动 / 打开 / 停止、记忆写入和画板资产导入都必须先完成 `/project` 初始化；未初始化时只提示设置本地项目，不落到默认临时目录。
 - 终端可用 `npm run ai-game-creator-shell:llm-status` 检查 LLM 环境变量是否就绪；CLI 和桌面 App 内的 `/llm-status` / 生成入口都会先读取 gitignored 的 `.env.secrets.local`，不请求上游、不显示 API Key，缺配置时以非零状态退出或在聊天里提示未就绪。
 - 终端可用 `npm run ai-game-creator-shell:check` 跑 v1 开发验收：壳 typecheck、`platform-agent` 编排测试、共享契约测试、Tauri Rust 测试和无密钥本地 provider 端到端 smoke。
-- 终端可用 `npm run ai-game-creator-shell:agent-run -- /绝对项目路径 "游戏创作需求"` 跑一次真实 LLM 生成、落盘、`game.static_smoke` 和本地 HTTP 预览；该入口读取当前环境和 gitignored 的 `.env.secrets.local`，不把 API Key 写入仓库或项目文件。自动验证可加 `--no-wait`，例如 `npm run ai-game-creator-shell:agent-run -- --no-wait /tmp/genarrative-ai-game-test "像素风反弹弹幕厨房"`，生成预览 trace 后立即停止本地预览，避免终端卡在回车等待。真实 OpenAI-compatible 网关建议设置 `GENARRATIVE_GAME_CREATOR_LLM_STREAM=true` 跑 Planner、组内角色和 Generator，避免长请求非流式空闲断连。
+- 终端可用 `npm run ai-game-creator-shell:agent-run -- /绝对项目路径 "游戏创作需求"` 跑一次真实 LLM 生成、落盘、`game.static_smoke` 和本地 HTTP 预览；该入口读取当前环境和 gitignored 的 `.env.secrets.local`，不把 API Key 写入仓库或项目文件。自动验证可加 `--no-wait`，例如 `npm run ai-game-creator-shell:agent-run -- --no-wait /tmp/genarrative-ai-game-test "像素风反弹弹幕厨房"`，生成预览 trace 后立即停止本地预览，避免终端卡在回车等待。真实 OpenAI-compatible 网关建议设置 `GENARRATIVE_GAME_CREATOR_LLM_STREAM=true` 跑 Planner 和 Generator，避免长请求非流式空闲断连。
 - 终端可用 `npm run ai-game-creator-shell:agent-run:smoke` 跑一次无密钥本地端到端 smoke：脚本启动本机 OpenAI-compatible SSE 流式测试 provider，预置一个本地上传图片和一个本地上传音频，复用真实 `--agent-run`、Planner / Orchestrator / 角色 agent / Generator / Evaluator loop、本地落盘、`game.static_smoke` 和本地 HTTP 预览，并断言每次 provider 请求都使用 `stream: true`、provider prompt 收到图片与音频资产上下文、生成 HTML 引用这些资产、预览服务能用 `GET` 读取 `/assets/...`、用 `HEAD` 返回真实资源长度和对应 MIME、headless Chrome 打开预览后至少执行一帧游戏 JS，且通过确定性亮色探针采样证明 canvas 不是空白画布、`.agent/run.latest.json` 的 step group 覆盖 design / balance / art / audio / code / publishing 六组、第二轮会重跑 Evaluator 命中任务及其下游影响任务，未受影响角色 carry-over；随后脚本自动给 CLI 发送回车停止预览。该脚本只用于开发验证，不进入产品生成路径。
 - `npm run ai-game-creator-shell:dev` 的 Tauri `devUrl` 固定为 `http://127.0.0.1:3080/`，Vite 必须 `strictPort` 对齐；`beforeDevCommand` 先复用已经跑在 3080 且页面标题为 `AI 游戏创作` 的本 app Vite server，否则才启动新的 Vite，若端口被其它服务占用则直接失败并提示释放端口。
-- `.agent/manifest.json` 会保存 6 个专业组下 15 个组内角色任务状态，当前覆盖 `Director`、`Gameplay`、`Difficulty`、`Asset`、`Polish`、`SFX`、`Code`、`Preview`、`Playtest`、`Publish`；开发窗口的专业组面板读取 manifest，而不是前端硬编码。
+- `.agent/manifest.json` 会保存 6 个专业组下 16 个组内角色任务状态，当前覆盖 `Director`、`Gameplay`、`Difficulty`、`Asset`、`Polish`、`SFX`、`Code`、`Review`、`Preview`、`Playtest`、`Publish`；程序组内显式包含 `quality-review` 质量评审 gate，由 Evaluator trace 标记完成；开发窗口的专业组面板读取 manifest，而不是前端硬编码。
 - 共享契约和 `platform-agent` 会按任务依赖与 `completed` 状态计算当前可执行任务，作为 v1 的最小编排选择器；每轮 `Orchestrator` 的 activeTaskIds、carriedTaskIds、repairRoutes 和 dependencyWaves 由 `platform-agent` 纯编排内核产出，`apps/ai-game-creator-shell` 只负责写入 `.agent/passes/pass-N/` 和执行本地工具；`Evaluator` 会在 `.agent/findings.md` 写出 `## Repair Routes` JSON，下一轮编排优先采用该结构化 taskIds，解析不到时才退回关键词路由；返工路由会按任务图自动扩展下游影响任务，例如美术资产变化会继续触发程序预览和运营包装重算。
-- `game.generate_draft` 使用 OpenAI-compatible LLM 配置生成结构化 JSON 草案，读取 `GENARRATIVE_GAME_CREATOR_LLM_API_KEY` / `GENARRATIVE_LLM_API_KEY` / `LLM_API_KEY` / `OPENAI_API_KEY`、`GENARRATIVE_GAME_CREATOR_LLM_BASE_URL` / `GENARRATIVE_LLM_BASE_URL` / `LLM_BASE_URL` / `OPENAI_BASE_URL`、`GENARRATIVE_GAME_CREATOR_LLM_MODEL` / `GENARRATIVE_LLM_MODEL` / `LLM_MODEL` / `OPENAI_MODEL`；`GENARRATIVE_GAME_CREATOR_LLM_STREAM=true` 时 Planner、组内角色和 Generator 使用流式请求；缺少配置或模型返回非法 JSON 时直接失败，不静默回退固定模板。
+- `game.generate_draft` 使用 OpenAI-compatible LLM 配置生成结构化 JSON 草案，读取 `GENARRATIVE_GAME_CREATOR_LLM_API_KEY` / `GENARRATIVE_LLM_API_KEY` / `LLM_API_KEY` / `OPENAI_API_KEY`、`GENARRATIVE_GAME_CREATOR_LLM_BASE_URL` / `GENARRATIVE_LLM_BASE_URL` / `LLM_BASE_URL` / `OPENAI_BASE_URL`、`GENARRATIVE_GAME_CREATOR_LLM_MODEL` / `GENARRATIVE_LLM_MODEL` / `LLM_MODEL` / `OPENAI_MODEL`；`GENARRATIVE_GAME_CREATOR_LLM_STREAM=true` 时 Planner 和 Generator 使用流式请求；缺少配置或模型返回非法 JSON 时直接失败，不静默回退固定模板。
 - 聊天输入 `/llm-status` 会触发只读 `llm.config_check`，确认 LLM base_url、model 和 API Key 是否已从环境变量读取；状态消息不会显示或保存 API Key。
 - `game.generate_draft` 的 LLM JSON 必须包含 `handoffs` 数组，覆盖 `design`、`balance`、`art`、`audio`、`code`、`publishing` 6 个专业组；每组必须给出 role、summary、outputs 和 next，缺组或交接内容不完整会判定为模型输出无效并进入返工。
-- `game.generate_draft` 的真实生成路径使用最小 Planner / Orchestrator / 组内角色 agent / Generator / Evaluator loop：Planner 写 `.agent/spec.md`；每轮 Orchestrator 先写 `.agent/passes/pass-N/agenda.md` 和 `.agent/passes/pass-N/task-graph.json`，首轮全量调度 15 个角色任务，返工轮按 `.agent/findings.md` 生成结构化 `repairRoutes`，重跑命中问题的角色任务及其下游依赖任务，其余角色 brief 从上一轮 carry-over；`task-graph.json` 记录 activeTaskIds、carriedTaskIds、repairFocus、repairRoutes 和按依赖排序的 dependencyWaves；角色 brief 写入 `.agent/passes/pass-N/groups/<group>/*.md`，再汇总为 `.agent/passes/pass-N/groups/*.md`；Generator 必须读取用户需求、记忆、`.agent/spec.md`、本轮 `agenda.md`、`task-graph.json`、`.agent/findings.md` 和 6 组汇总 brief 后返回结构化 JSON；每轮会把 Generator 草案拆成 6 组交接快照，写入 `.agent/passes/pass-N/`；Evaluator 做本地静态验收并写 `.agent/findings.md`。
+- `game.generate_draft` 的真实生成路径使用最小 Planner / Orchestrator / 组内角色 agent / Generator / Evaluator loop：Planner 写 `.agent/spec.md`；每轮 Orchestrator 先写 `.agent/passes/pass-N/agenda.md` 和 `.agent/passes/pass-N/task-graph.json`，首轮全量调度 16 个角色任务，返工轮按 `.agent/findings.md` 生成结构化 `repairRoutes`，重跑命中问题的角色任务及其下游依赖任务，其余角色 brief 从上一轮 carry-over；`task-graph.json` 记录 activeTaskIds、carriedTaskIds、repairFocus、repairRoutes 和按依赖排序的 dependencyWaves；角色 brief 写入 `.agent/passes/pass-N/groups/<group>/*.md`，再汇总为 `.agent/passes/pass-N/groups/*.md`；Generator 必须读取用户需求、记忆、`.agent/spec.md`、本轮 `agenda.md`、`task-graph.json`、`.agent/findings.md` 和 6 组汇总 brief 后返回结构化 JSON；每轮会把 Generator 草案拆成 6 组交接快照，写入 `.agent/passes/pass-N/`；Evaluator 做质量评审并写 `.agent/findings.md`，通过后才进入 `game.static_smoke` 静态自检和预览试玩。
 - loop 最多执行 3 轮；Evaluator 发现 HTML 非自包含、缺少 `canvas`、缺少 `requestAnimationFrame`、缺少输入监听或用户输入未转义时，把问题写入 `.agent/findings.md` 并让下一轮 Generator 修复。3 轮仍失败则 `game.generate_draft` 失败，不写最终游戏产物。
-- loop 每次运行会写 `.agent/run.latest.json` 和 `.agent/runs/<runId>.json`，记录 `Planner` / `Orchestrator` agenda / 15 个组内角色 brief 或 carry-over / 6 个 `GroupCoordinator` 汇总 / `Generator` / 6 个专业组交接 / `Evaluator` / `ArtifactWriter` / `Playtest` step、每步 `toolCalls`、输入文件、输出文件、状态、轮次、maxPasses、toolCallCount、maxToolCalls、stopReason、nextStep 和错误摘要；Planner、角色 agent 和 Generator 的 `inputPaths` 必须包含对应记忆文件、`.agent/manifest.json` 和 agenda 等上下文来源；每个 step 必须带 phase、taskId、group 和 role，`.agent/run.latest.json.taskGraph` 必须记录 goal、readyTaskIds、activeTaskIds、carriedTaskIds、repairFocus、repairRoutes 和当前任务状态；`.agent/run.latest.json.passPlans` 必须按轮记录 mode、summary、activeTaskIds、carriedTaskIds、dependencyWaves、repairFocus 和 repairRoutes，作为 `/trace` 与开发窗口判断编排 loop 是否真实发生的直接证据；`run.latest.json` 是当前指针，`.agent/runs/` 保留历史 run trace；开发窗口读取 latest 展示编排过程，并复用 `file.list/read` 列出和载入历史 run，普通用户窗口不展示。
+- loop 每次运行会写 `.agent/run.latest.json` 和 `.agent/runs/<runId>.json`，记录 `Planner` / `Orchestrator` agenda / 16 个组内角色 brief 或 carry-over / 6 个 `GroupCoordinator` 汇总 / `Generator` / 6 个专业组交接 / `Evaluator` 质量评审 / `ArtifactWriter` / `Playtest` step、每步 `toolCalls`、输入文件、输出文件、状态、轮次、maxPasses、toolCallCount、maxToolCalls、stopReason、nextStep 和错误摘要；Planner、角色 agent 和 Generator 的 `inputPaths` 必须包含对应记忆文件、`.agent/manifest.json` 和 agenda 等上下文来源；每个 step 必须带 phase、taskId、group 和 role，`.agent/run.latest.json.taskGraph` 必须记录 goal、readyTaskIds、activeTaskIds、carriedTaskIds、repairFocus、repairRoutes 和当前任务状态；`.agent/run.latest.json.passPlans` 必须按轮记录 mode、summary、activeTaskIds、carriedTaskIds、dependencyWaves、repairFocus 和 repairRoutes，作为 `/trace` 与开发窗口判断编排 loop 是否真实发生的直接证据；`run.latest.json` 是当前指针，`.agent/runs/` 保留历史 run trace；开发窗口读取 latest 展示编排过程，并复用 `file.list/read` 列出和载入历史 run，普通用户窗口不展示。
 - `.agent/run.latest.json` 的 `artifacts` 使用结构化记录，包含相对路径、字节数和 `fnv1a64:` checksum；除最终本地产物外，也会收集 `.agent/passes/pass-N/` 快照，便于确认返工前后的产物差异。
 - 通过 Evaluator 和 `game.static_smoke` 后，Agent loop 会把本次 runId、状态、轮次、下一步、active / carry-over 任务和最终本地产物摘要追加到 `memory/session.md` 与 `memory/project.md`；下一次 Planner、组内角色和 Generator 会通过记忆输入自然读取上一轮稳定原型状态，而不只依赖开发窗口 trace。
 - `.agent/agent.db` 当前作为最小本地索引文件使用 JSONL：初始化写入 `project.init`，每次 `game.generate_draft` 追加目标、标题和本地产物路径，上传 / 登记 / 画板导入资产时追加 `asset.register` 或 `asset.update`；v1 不引入 SQLite 依赖。
@@ -118,7 +118,7 @@ game-project/
 - `game.generate_draft` 写入最终产物后会复用白名单受限命令 `game.static_smoke` 做一次生成后自检，至少检查 `game/index.html` 包含 canvas、canvas 渲染上下文、绘制调用、主循环、非空输入监听、明确目标、失败或胜利状态和重开路径，且不使用远程资源、`eval`、`new Function`、`localStorage`、`fetch`、`WebSocket` 或 `ServiceWorker`，也不得包含固定星核传送门模板词、纯按钮计分模板或 `TODO` / `待实现` / `这里省略` 等未完成实现；画板资源占位引用允许出现在 asset id 或说明中，并把该工具调用写入 `.agent/run.latest.json` 与 `.agent/logs/command.log`；自检失败则本次命令失败，不继续启动预览。
 - `ArtifactWriter` step 使用 `file.write.local_artifacts` 工具调用记录最终写入的 `memory/`、`game/`、`assets/`、`exports/` 和 `.agent/manifest.json` 路径；写入完成后 `nextStep` 指向 `game.static_smoke`。
 - `preview.start` / `preview.stop` 会追加 `.agent/logs/preview.log`，并在 `.agent/run.latest.json` 已存在时追加 `Preview` step 和 `preview.*` toolCall，记录本地 HTTP 预览 URL 与停止事件；单全局本地预览被新项目替换时，会 best-effort 把旧项目 manifest、preview log 和 trace 记录为 stopped，避免旧项目残留 running；本地 HTTP server 的 `/` 映射到 `game/index.html`，只允许读取 canonical 后仍位于项目真实 `game/` 或真实 `assets/` 下的文件，拒绝 `memory/`、`.agent/`、`exports/`、`..`、一级 `game` / `assets` 符号链接目录和内部符号链接越界，并为常见图片、音频、视频和 Web 资源返回对应 MIME；静态 `HEAD` 返回真实 `Content-Length` 但不返回 body，确保浏览器和媒体资源探测可用；上传和画板回流资产可被生成游戏引用但不会暴露记忆或 trace；没有 run trace 的手动预览启动不阻断。
-- 聊天输入会生成待确认的 `game.generate_draft` 内置命令；用户确认后才把 LLM 返回的结构化草案写入短期记忆 `memory/session.md`、长期记忆 `memory/project.md`、设计草案 `game/game_design.md`、数值配置 `game/balance.json`、美术清单 `assets/manifest.art.json`、音乐音效清单 `assets/manifest.audio.json`、发布包装草案 `exports/README.md` 和可运行 `game/index.html`。生成完成后，普通聊天消息会自动展示最近一次 Agent loop 的轮次、工具调用、active / carry-over 任务、返工焦点、画板同步建议命令和本地产物摘要；完整证据仍由 `/trace` 读取同一份 `.agent/run.latest.json`。
+- 聊天输入会生成待确认的 `game.generate_draft` 内置命令；用户确认后，正式用户聊天会实时展示 Planner LLM、Orchestrator、6 组角色 brief、Generator LLM、Evaluator 质量评审、ArtifactWriter 和 `game.static_smoke` 的进度，再把 LLM 返回的结构化草案写入短期记忆 `memory/session.md`、长期记忆 `memory/project.md`、设计草案 `game/game_design.md`、数值配置 `game/balance.json`、美术清单 `assets/manifest.art.json`、音乐音效清单 `assets/manifest.audio.json`、发布包装草案 `exports/README.md` 和可运行 `game/index.html`。生成完成后，普通聊天消息会自动展示最近一次 Agent loop 的 Run、LLM 对话、轮次、工具调用、active / carry-over 任务、返工焦点、编排轮次、最近步骤、画板同步建议命令和本地产物快照；完整证据仍由 `/trace` 读取同一份 `.agent/run.latest.json`。
 - `game.generate_draft` 的 `game/index.html` 必须是可试玩原型，至少包含输入、主循环、目标、失败或胜利状态和重开路径；不能只输出按钮计分或纯展示页。
 - `game.generate_draft` 会校验 LLM 输出：`balance`、美术清单和音乐清单必须是 JSON object，`gameHtml` 必须是自包含 HTML、包含 `canvas` 与 `requestAnimationFrame`，不得加载远程脚本或资源，不得使用 `eval` / `new Function` / `localStorage` / `fetch` / `WebSocket` / `ServiceWorker`，不得把包含 `<` / `>` 的用户输入原样写入 HTML。
 - 同一项目内多次 `game.generate_draft` 不覆盖记忆文件，而是继续追加短期对话记录和长期创作目标记录，保留用户迭代历史。

packages/shared/src/contracts/gameCreationApp.ts

@@ -87,6 +87,7 @@ export const GAME_CREATION_AGENT_CAPABILITIES = [
     area: 'agent-runtime',
     title: '组内角色协作',
   },
+  { id: 'quality-review', area: 'agent-runtime', title: '质量评审' },
   {
     id: 'repair-loop-carryover',
     area: 'agent-runtime',
@@ -257,13 +258,23 @@ export const GAME_CREATION_APP_SEED_TASKS = [
     artifacts: ['game/'],
     acceptanceCriteria: ['本地 Web 游戏项目可以通过 HTTP server 打开'],
   },
+  {
+    id: 'quality-review',
+    title: '执行质量评审',
+    group: 'code',
+    role: 'Review',
+    status: 'pending',
+    dependencies: ['code-prototype'],
+    artifacts: ['.agent/findings.md', '.agent/run.latest.json'],
+    acceptanceCriteria: ['玩法、资产、数值、程序和发布包装通过跨专业组质量评审'],
+  },
   {
     id: 'preview-readiness',
     title: '执行静态自检',
     group: 'code',
     role: 'Preview',
     status: 'pending',
-    dependencies: ['code-prototype'],
+    dependencies: ['quality-review'],
     artifacts: ['.agent/logs/command.log', '.agent/run.latest.json'],
     acceptanceCriteria: ['HTML 自包含且通过本地静态 smoke'],
   },

server-rs/crates/platform-agent/src/game_creation.rs

@@ -170,12 +170,21 @@ pub fn build_game_creation_seed_task_graph(
                 ["game/"],
                 ["本地 Web 游戏项目可以通过 HTTP server 打开"],
             ),
+            task(
+                "quality-review",
+                "执行质量评审",
+                GameCreationAgentGroup::Code,
+                "Review",
+                ["code-prototype"],
+                [".agent/findings.md", ".agent/run.latest.json"],
+                ["玩法、资产、数值、程序和发布包装通过跨专业组质量评审"],
+            ),
             task(
                 "preview-readiness",
                 "执行静态自检",
                 GameCreationAgentGroup::Code,
                 "Preview",
-                ["code-prototype"],
+                ["quality-review"],
                 [".agent/logs/command.log", ".agent/run.latest.json"],
                 ["HTML 自包含且通过本地静态 smoke"],
             ),
@@ -858,7 +867,7 @@ mod tests {
         );
 
         assert_eq!(plan.mode, "initial");
-        assert_eq!(plan.active_task_ids.len(), 15);
+        assert_eq!(plan.active_task_ids.len(), 16);
         assert!(plan.carried_task_ids.is_empty());
         assert_eq!(plan.repair_routes, Vec::new());
         assert_eq!(plan.dependency_waves[0], vec!["design-director"]);
@@ -879,6 +888,7 @@ mod tests {
             vec![
                 "code-director",
                 "code-prototype",
+                "quality-review",
                 "preview-readiness",
                 "preview-playtest",
                 "publish-strategy",
@@ -926,6 +936,7 @@ mod tests {
             plan.active_task_ids,
             vec![
                 "code-prototype",
+                "quality-review",
                 "preview-readiness",
                 "preview-playtest",
                 "publish-strategy",
@@ -936,6 +947,7 @@ mod tests {
             plan.repair_routes[0].task_ids,
             vec![
                 "code-prototype",
+                "quality-review",
                 "preview-readiness",
                 "preview-playtest",
                 "publish-strategy",
@@ -981,6 +993,7 @@ mod tests {
                 "art-polish",
                 "code-director",
                 "code-prototype",
+                "quality-review",
                 "preview-readiness",
                 "preview-playtest",
                 "publish-strategy",
@@ -1014,7 +1027,7 @@ mod tests {
         );
 
         assert_eq!(plan.mode, "repair");
-        assert_eq!(plan.active_task_ids.len(), 15);
+        assert_eq!(plan.active_task_ids.len(), 16);
         assert!(plan.carried_task_ids.is_empty());
         assert_eq!(plan.repair_routes[0].reason, "cross-group-handoff");
     }

server-rs/crates/shared-contracts/src/game_creation_app.rs

@@ -68,7 +68,7 @@ pub struct GameCreationAgentCapabilityDescriptor {
     pub title: &'static str,
 }
 
-pub const GAME_CREATION_AGENT_CAPABILITIES: [GameCreationAgentCapabilityDescriptor; 21] = [
+pub const GAME_CREATION_AGENT_CAPABILITIES: [GameCreationAgentCapabilityDescriptor; 22] = [
     capability("chat", "user", "聊天入口"),
     capability("file-upload", "user", "上传文件"),
     capability("built-in-commands", "agent-runtime", "内置命令调用"),
@@ -89,6 +89,7 @@ pub const GAME_CREATION_AGENT_CAPABILITIES: [GameCreationAgentCapabilityDescript
     capability("tool-call-budget", "agent-runtime", "工具调用预算"),
     capability("multi-agent-collaboration", "agent-runtime", "多智能体协作"),
     capability("role-level-collaboration", "agent-runtime", "组内角色协作"),
+    capability("quality-review", "agent-runtime", "质量评审"),
     capability(
         "repair-loop-carryover",
         "agent-runtime",
@@ -269,12 +270,21 @@ pub fn new_game_creation_app_seed_tasks() -> Vec<GameCreationAppTaskState> {
             ["game/"],
             ["本地 Web 游戏项目可以通过 HTTP server 打开"],
         ),
+        task(
+            "quality-review",
+            "执行质量评审",
+            GameCreationAppAgentGroup::Code,
+            "Review",
+            ["code-prototype"],
+            [".agent/findings.md", ".agent/run.latest.json"],
+            ["玩法、资产、数值、程序和发布包装通过跨专业组质量评审"],
+        ),
         task(
             "preview-readiness",
             "执行静态自检",
             GameCreationAppAgentGroup::Code,
             "Preview",
-            ["code-prototype"],
+            ["quality-review"],
             [".agent/logs/command.log", ".agent/run.latest.json"],
             ["HTML 自包含且通过本地静态 smoke"],
         ),
@@ -869,6 +879,7 @@ mod tests {
                 "audio-asset-plan",
                 "code-director",
                 "code-prototype",
+                "quality-review",
                 "preview-readiness",
                 "preview-playtest",
                 "publish-strategy",