Merge remote-tracking branch 'origin/master' into opt/design_agent

按项目与回合绑定 reasoning 事件

保留回合结束后的思考过程并在新回合清理

补充默认折叠的展开收起交互与定向测试

更新推理与正文分离里程碑进度

.gitea/workflows/project-ci.yml

@@ -27,18 +27,9 @@ env:
   RUSTC_WRAPPER: ''
   CARGO_BUILD_RUSTC_WRAPPER: ''
 
-# job 声明顺序就是 runner 领取顺序，因此把最长尾的客户端 Rust 门禁排在前面，
-# 让它在最少的等待下占用并发槽位；其余 job 按时长递减排列。
-#
-# 客户端（微信壳 / Expo 移动壳 / Tauri 桌面壳 / AI 游戏创作壳）门禁原先全部串在
-# `Native shell tests` 一个 job 里，实测 18 分 37 秒，其中 AI 游戏创作壳的串行
-# Rust 套件（2451 个用例，`--test-threads=1`）单独占 533 秒。现在按门禁组拆成
-# `Native shell tests`、`AI game creator shell web tests` 与
-# `AI game creator shell Rust tests` 三个 job，各自的命令与拆分前逐一对应。
 jobs:
-  # 该 job 最长：AI 游戏创作壳的共享 / 平台 crate 测试加串行壳测试。
+  repository-checks:
-  ai-game-creator-shell-rust-tests:
+    name: Repository checks
-    name: AI game creator shell Rust tests
     runs-on: genarrative-ci
     steps:
       - name: Checkout full history from Gitea
@@ -50,63 +41,76 @@ jobs:
       - name: Validate preinstalled CI job image and sandbox
         run: GENARRATIVE_GITEA_CI_CHECK_RUNTIME=1 bash scripts/check-gitea-ci-job-image.sh
 
+      - name: Resolve comparison base
+        shell: bash
+        run: |
+          set -euo pipefail
+          base_ref="$(node -e '
+            const fs = require("node:fs");
+            const event = JSON.parse(fs.readFileSync(process.env.GITHUB_EVENT_PATH, "utf8"));
+            process.stdout.write(event.pull_request?.base?.sha ?? event.before ?? "");
+          ')"
+          if [[ -n "${base_ref}" && ! "${base_ref}" =~ ^0+$ ]]; then
+            git cat-file -e "${base_ref}^{commit}" 2>/dev/null || {
+              echo "comparison base commit is unavailable: ${base_ref}" >&2
+              exit 1
+            }
+          else
+            base_ref="$(git merge-base HEAD origin/master 2>/dev/null || git rev-parse HEAD)"
+          fi
+          resolved_base_ref="$(git rev-parse --verify "${base_ref}^{commit}" 2>/dev/null || true)"
+          head_ref="$(git rev-parse HEAD)"
+          if [[ "${resolved_base_ref}" == "${head_ref}" ]]; then
+            resolved_base_ref="$(git rev-parse --verify HEAD^ 2>/dev/null || true)"
+          fi
+          if [[ -z "${resolved_base_ref}" ]]; then
+            echo 'comparison base must resolve to a commit distinct from HEAD.' >&2
+            exit 1
+          fi
+          base_ref="${resolved_base_ref}"
+          if [[ "${GITHUB_EVENT_NAME:-}" == 'pull_request' ]] \
+            && ! git merge-base --is-ancestor "${base_ref}" HEAD; then
+            echo 'pull request head does not contain the latest base commit; update the branch and rerun CI.' >&2
+            exit 1
+          fi
+          echo "SPACETIME_SCHEMA_BASE_REF=${base_ref}" >> "${GITHUB_ENV}"
+
       - name: Install npm dependencies
         run: bash scripts/ci-npm-ci-with-retry.sh
 
-      - name: Prepare AI game creator shell Rust dependencies
+      - name: Run repository checks
-        shell: bash
+        run: npm run check:repository-ci
-        run: |
-          set -euo pipefail
-          for manifest_path in \
-            server-rs/Cargo.toml \
-            apps/ai-game-creator-shell/src-tauri/Cargo.toml; do
-            for attempt in $(seq 1 5); do
-              if cargo fetch --locked \
-                --target x86_64-unknown-linux-gnu \
-                --manifest-path "${manifest_path}"; then
-                break
-              fi
-              if [[ "${attempt}" -eq 5 ]]; then
-                echo "Cargo dependency fetch failed after 5 attempts: ${manifest_path}" >&2
-                exit 1
-              fi
-              sleep $((attempt * 2))
-            done
-          done
 
-      - name: Prepare standalone Rust crate dependencies
+  frontend-tests:
-        shell: bash
+    name: Frontend tests
-        run: |
+    runs-on: genarrative-ci
-          set -euo pipefail
+    steps:
-          # agent-runtime-core / agent-runtime-orchestration 被 server-rs/Cargo.toml 的
+      - name: Checkout source from Gitea
-          # exclude 排除，不参与上面的 workspace 锁文件，因此上面那次锁定 fetch 覆盖不到它们；
+        env:
-          # 而 `npm run ai-game-creator-shell:check:rust` 会用
+          GENARRATIVE_GITEA_FETCH_DEPTH: '1'
-          # `cargo test --manifest-path` 单独跑这两个 crate。不在这里预热的话，这两条测试
+          GENARRATIVE_GITEA_TOKEN: ${{ github.token }}
-          # 会在测试阶段自己 `Updating crates.io index`，crates.io 一抖动整条 job 就红
+        run: genarrative-gitea-checkout
-          # （见 #327 / PR #316 run 1950）。
-          # 两个 crate 都没有提交 Cargo.lock，所以这里只能做不带锁标志的 fetch：
-          # 加锁标志会因为缺少锁文件直接失败。生成的 Cargo.lock 落在两个 crate 目录内，
-          # 已被各自的 .gitignore 忽略，只留在容器里；随后的测试阶段因此能用锁定版本
-          # 解析，不再触碰 registry index。
-          for manifest_path in \
-            server-rs/crates/agent-runtime-core/Cargo.toml \
-            server-rs/crates/agent-runtime-orchestration/Cargo.toml; do
-            for attempt in $(seq 1 5); do
-              if cargo fetch \
-                --target x86_64-unknown-linux-gnu \
-                --manifest-path "${manifest_path}"; then
-                break
-              fi
-              if [[ "${attempt}" -eq 5 ]]; then
-                echo "standalone crate dependency fetch failed after 5 attempts: ${manifest_path}" >&2
-                exit 1
-              fi
-              sleep $((attempt * 2))
-            done
-          done
 
-      - name: Run AI game creator shell Rust gates
+      - name: Validate preinstalled CI job image and sandbox
-        run: npm run check:native-shells:agc-rust
+        run: GENARRATIVE_GITEA_CI_CHECK_RUNTIME=1 bash scripts/check-gitea-ci-job-image.sh
+
+      - name: Install npm dependencies
+        run: bash scripts/ci-npm-ci-with-retry.sh
+
+      - name: Run frontend and script tests
+        run: npm run test
+
+      - name: Run BgFilter worker smoke harness tests
+        run: npm run bgfilter-worker:smoke-test
+
+      - name: Validate production health patrol behavior
+        run: npm run check:production-health-patrol
+
+      - name: Validate production API release behavior
+        run: npm run check:production-api-release
+
+      - name: Validate production API deploy behavior
+        run: npm run check:production-api-deploy
 
   backend-tests:
     name: Backend tests
@@ -190,8 +194,6 @@ jobs:
       - name: Check SpacetimeDB module
         run: cargo check --locked -p spacetime-module --manifest-path server-rs/Cargo.toml
 
-  # 客户端的壳级与契约门禁：静态契约断言、H5 / 微信 / 移动 / 桌面壳运行时门禁，
-  # 以及依赖发布产物的构建 smoke。
   native-shell-tests:
     name: Native shell tests
     runs-on: genarrative-ci
@@ -213,6 +215,7 @@ jobs:
         run: |
           set -euo pipefail
           for manifest_path in \
+            server-rs/Cargo.toml \
             apps/desktop-shell/src-tauri/Cargo.toml \
             apps/ai-game-creator-shell/src-tauri/Cargo.toml; do
             for attempt in $(seq 1 5); do
@@ -229,118 +232,39 @@ jobs:
             done
           done
 
-      - name: Run native shell contract gates
+      - name: Prepare standalone Rust crate dependencies
-        run: npm run check:native-shells:contract
-
-      - name: Run native shell gates
-        run: npm run check:native-shells:shells
-
-      - name: Run native shell release build smoke
-        run: npm run check:native-shells:release
-
-      - name: Ensure native lockfiles are unchanged
-        run: git diff --exit-code -- apps/desktop-shell/src-tauri/Cargo.lock apps/ai-game-creator-shell/src-tauri/Cargo.lock
-
-  frontend-tests:
-    name: Frontend tests
-    runs-on: genarrative-ci
-    steps:
-      - name: Checkout source from Gitea
-        env:
-          GENARRATIVE_GITEA_FETCH_DEPTH: '1'
-          GENARRATIVE_GITEA_TOKEN: ${{ github.token }}
-        run: genarrative-gitea-checkout
-
-      - name: Validate preinstalled CI job image and sandbox
-        run: GENARRATIVE_GITEA_CI_CHECK_RUNTIME=1 bash scripts/check-gitea-ci-job-image.sh
-
-      - name: Install npm dependencies
-        run: bash scripts/ci-npm-ci-with-retry.sh
-
-      - name: Run frontend and script tests
-        run: npm run test
-
-      - name: Run BgFilter worker smoke harness tests
-        run: npm run bgfilter-worker:smoke-test
-
-      - name: Validate production health patrol behavior
-        run: npm run check:production-health-patrol
-
-      - name: Validate production API release behavior
-        run: npm run check:production-api-release
-
-      - name: Validate production API deploy behavior
-        run: npm run check:production-api-deploy
-
-  repository-checks:
-    name: Repository checks
-    runs-on: genarrative-ci
-    steps:
-      - name: Checkout full history from Gitea
-        env:
-          GENARRATIVE_GITEA_FETCH_DEPTH: '0'
-          GENARRATIVE_GITEA_TOKEN: ${{ github.token }}
-        run: genarrative-gitea-checkout
-
-      - name: Validate preinstalled CI job image and sandbox
-        run: GENARRATIVE_GITEA_CI_CHECK_RUNTIME=1 bash scripts/check-gitea-ci-job-image.sh
-
-      - name: Resolve comparison base
         shell: bash
         run: |
           set -euo pipefail
-          base_ref="$(node -e '
+          # agent-runtime-core / agent-runtime-orchestration 被 server-rs/Cargo.toml 的
-            const fs = require("node:fs");
+          # exclude 排除，不参与上面的 workspace 锁文件，因此上面那次锁定 fetch 覆盖不到它们；
-            const event = JSON.parse(fs.readFileSync(process.env.GITHUB_EVENT_PATH, "utf8"));
+          # 而 check:native-shells 会经 agent-runtime-*:check 用 `cargo test --manifest-path`
-            process.stdout.write(event.pull_request?.base?.sha ?? event.before ?? "");
+          # 单独跑这两个 crate。不在这里预热的话，这两条测试会在测试阶段自己
-          ')"
+          # `Updating crates.io index`，crates.io 一抖动整条 native shell 作业就红
-          if [[ -n "${base_ref}" && ! "${base_ref}" =~ ^0+$ ]]; then
+          # （见 #327 / PR #316 run 1950）。
-            git cat-file -e "${base_ref}^{commit}" 2>/dev/null || {
+          # 两个 crate 都没有提交 Cargo.lock，所以这里只能做不带锁标志的 fetch：
-              echo "comparison base commit is unavailable: ${base_ref}" >&2
+          # 加锁标志会因为缺少锁文件直接失败。生成的 Cargo.lock 落在两个 crate 目录内，
-              exit 1
+          # 已被各自的 .gitignore 忽略，只留在容器里；随后的测试阶段因此能用锁定版本
-            }
+          # 解析，不再触碰 registry index。
-          else
+          for manifest_path in \
-            base_ref="$(git merge-base HEAD origin/master 2>/dev/null || git rev-parse HEAD)"
+            server-rs/crates/agent-runtime-core/Cargo.toml \
+            server-rs/crates/agent-runtime-orchestration/Cargo.toml; do
+            for attempt in $(seq 1 5); do
+              if cargo fetch \
+                --target x86_64-unknown-linux-gnu \
+                --manifest-path "${manifest_path}"; then
+                break
               fi
-          resolved_base_ref="$(git rev-parse --verify "${base_ref}^{commit}" 2>/dev/null || true)"
+              if [[ "${attempt}" -eq 5 ]]; then
-          head_ref="$(git rev-parse HEAD)"
+                echo "standalone crate dependency fetch failed after 5 attempts: ${manifest_path}" >&2
-          if [[ "${resolved_base_ref}" == "${head_ref}" ]]; then
-            resolved_base_ref="$(git rev-parse --verify HEAD^ 2>/dev/null || true)"
-          fi
-          if [[ -z "${resolved_base_ref}" ]]; then
-            echo 'comparison base must resolve to a commit distinct from HEAD.' >&2
                 exit 1
               fi
-          base_ref="${resolved_base_ref}"
+              sleep $((attempt * 2))
-          if [[ "${GITHUB_EVENT_NAME:-}" == 'pull_request' ]] \
+            done
-            && ! git merge-base --is-ancestor "${base_ref}" HEAD; then
+          done
-            echo 'pull request head does not contain the latest base commit; update the branch and rerun CI.' >&2
-            exit 1
-          fi
-          echo "SPACETIME_SCHEMA_BASE_REF=${base_ref}" >> "${GITHUB_ENV}"
 
-      - name: Install npm dependencies
+      - name: Run native shell gates
-        run: bash scripts/ci-npm-ci-with-retry.sh
+        run: npm run check:native-shells
 
-      - name: Run repository checks
+      - name: Ensure native lockfiles are unchanged
-        run: npm run check:repository-ci
+        run: git diff --exit-code -- apps/desktop-shell/src-tauri/Cargo.lock apps/ai-game-creator-shell/src-tauri/Cargo.lock
-
-  # 客户端的 AI 游戏创作壳前端门禁：typecheck、壳内测试与本地 provider agent-run smoke。
-  ai-game-creator-shell-web-tests:
-    name: AI game creator shell web tests
-    runs-on: genarrative-ci
-    steps:
-      - name: Checkout full history from Gitea
-        env:
-          GENARRATIVE_GITEA_FETCH_DEPTH: '0'
-          GENARRATIVE_GITEA_TOKEN: ${{ github.token }}
-        run: genarrative-gitea-checkout
-
-      - name: Validate preinstalled CI job image and sandbox
-        run: GENARRATIVE_GITEA_CI_CHECK_RUNTIME=1 bash scripts/check-gitea-ci-job-image.sh
-
-      - name: Install npm dependencies
-        run: bash scripts/ci-npm-ci-with-retry.sh
-
-      - name: Run AI game creator shell web gates
-        run: npm run check:native-shells:agc-web

apps/ai-game-creator-shell/src-tauri/src/agent/codex_app_server.rs

@@ -2829,6 +2829,8 @@ impl CodexAppServerConnection {
                             callback(&platform_llm::LlmStreamDelta {
                                 accumulated_text: streamed_text.clone(),
                                 delta_text: delta,
+                                accumulated_reasoning: String::new(),
+                                reasoning_delta: String::new(),
                                 finish_reason: None,
                             });
                         }
@@ -3133,6 +3135,7 @@ fn parse_game_creator_codex_app_server_text(
         } else {
             String::new()
         },
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id: Some(thread_id.to_string()),
         usage: None,

apps/ai-game-creator-shell/src-tauri/src/agent/codex_cli.rs

@@ -589,6 +589,7 @@ fn parse_game_creator_codex_cli_response(
         } else {
             String::new()
         },
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id,
         usage,

apps/ai-game-creator-shell/src-tauri/src/agent/design_runtime.rs

@@ -104,6 +104,17 @@ fn design_event(
     }
 }
 
+fn design_reasoning_event(
+    root: &Path,
+    turn_id: &str,
+    id: Option<&str>,
+    reasoning: String,
+) -> DesignEvent {
+    let mut event = design_event(root, turn_id, "reasoning", id, None, None);
+    event.reasoning_text = Some(reasoning);
+    event
+}
+
 fn design_project_id(root: &Path) -> Result<String, String> {
     validate_project_root(root)?;
     Ok(read_existing_manifest_for_project(root)?.project_id)
@@ -462,6 +473,7 @@ fn build_design_request(
         .with_tool_choice(platform_llm::LlmToolChoice::Auto)
         .with_web_search(false);
     apply_game_creator_llm_reasoning_effort(request, llm)
+        .map(|request| request.with_reasoning_capture(true))
 }
 
 // 调试队列只接收副本，写盘慢或失败时丢弃，不参与会话恢复。
@@ -548,6 +560,12 @@ async fn request_design_provider(
             Some(String::new()),
             None,
         ));
+        emit(design_reasoning_event(
+            root,
+            &turn_id,
+            Some(&message_id),
+            String::new(),
+        ));
         let result = if llm.stream {
             let mut stream_sequence = 0_u64;
             client
@@ -565,10 +583,13 @@ async fn request_design_provider(
                             "model": llm.model,
                             "deltaChars": delta.delta_text.chars().count(),
                             "accumulatedChars": delta.accumulated_text.chars().count(),
+                            "reasoningDeltaChars": delta.reasoning_delta.chars().count(),
+                            "reasoningAccumulatedChars": delta.accumulated_reasoning.chars().count(),
                             "deltaText": delta.delta_text,
                             "finishReason": delta.finish_reason,
                         }),
                     );
+                    if !delta.delta_text.is_empty() || delta.finish_reason.is_some() {
                         emit(design_event(
                             root,
                             &turn_id,
@@ -577,6 +598,15 @@ async fn request_design_provider(
                             Some(delta.accumulated_text.clone()),
                             None,
                         ));
+                    }
+                    if !delta.reasoning_delta.is_empty() {
+                        emit(design_reasoning_event(
+                            root,
+                            &turn_id,
+                            Some(&message_id),
+                            delta.accumulated_reasoning.clone(),
+                        ));
+                    }
                 })
                 .await
         } else {
@@ -584,6 +614,14 @@ async fn request_design_provider(
         };
         match result {
             Ok(response) => {
+                if !response.reasoning.is_empty() {
+                    emit(design_reasoning_event(
+                        root,
+                        &turn_id,
+                        Some(&message_id),
+                        response.reasoning.clone(),
+                    ));
+                }
                 design_debug(
                     root,
                     "response",
@@ -606,6 +644,12 @@ async fn request_design_provider(
                     || game_creator_agent_runtime_transient_provider_error_kind(&error, false)
                         .is_none()
                 {
+                    emit(design_reasoning_event(
+                        root,
+                        &turn_id,
+                        Some(&message_id),
+                        String::new(),
+                    ));
                     return Err(detail);
                 }
                 tokio::time::sleep(Duration::from_millis(
@@ -654,8 +698,24 @@ async fn request_scripted_design_provider(
             Some(String::new()),
             None,
         ));
+        emit(design_reasoning_event(
+            root,
+            &turn_id,
+            Some(&message_id),
+            String::new(),
+        ));
         match fake_provider::take() {
-            Some(Ok(response)) => return Ok(response),
+            Some(Ok(response)) => {
+                if !response.reasoning.is_empty() {
+                    emit(design_reasoning_event(
+                        root,
+                        &turn_id,
+                        Some(&message_id),
+                        response.reasoning.clone(),
+                    ));
+                }
+                return Ok(response);
+            }
             Some(Err(error)) => {
                 let detail = redact_agent_runtime_error(
                     root,
@@ -666,10 +726,24 @@ async fn request_scripted_design_provider(
                     || game_creator_agent_runtime_transient_provider_error_kind(&error, false)
                         .is_none()
                 {
+                    emit(design_reasoning_event(
+                        root,
+                        &turn_id,
+                        Some(&message_id),
+                        String::new(),
+                    ));
                     return Err(detail);
                 }
             }
-            None => return Err("假 Provider 脚本耗尽".into()),
+            None => {
+                emit(design_reasoning_event(
+                    root,
+                    &turn_id,
+                    Some(&message_id),
+                    String::new(),
+                ));
+                return Err("假 Provider 脚本耗尽".into());
+            }
         }
     }
     unreachable!()
@@ -1285,6 +1359,7 @@ mod tests {
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: "fake-design".into(),
             text: text.into(),
+            reasoning: String::new(),
             finish_reason: Some(if calls.is_empty() {
                 "stop".into()
             } else {
@@ -1345,6 +1420,69 @@ mod tests {
             .clone()
     }
 
+    #[test]
+    fn design_request_enables_reasoning_capture_only_for_design_runtime() {
+        let session = new_design_session("project", "quality");
+        let request = build_design_request(&session, &pack(), &GameCreatorLlmConfig::default())
+            .expect("design request");
+        assert!(request.capture_reasoning);
+    }
+
+    #[tokio::test(flavor = "current_thread")]
+    async fn scripted_design_provider_emits_reasoning_without_persisting_it() {
+        let (_temp, root, _resources) = init_design_project();
+        let mut session = new_design_session("design-fake", "quality");
+        begin_design_turn(&mut session, "turn-reasoning");
+        let mut response = fake_response("reasoning", "正文", Vec::new());
+        response.reasoning = "先分析需求，再组织方案。".into();
+        let _fake = fake_provider::install(vec![Ok(response)], 0);
+        let mut events = Vec::new();
+        let response =
+            request_scripted_design_provider(&root, &mut session, &mut |event| events.push(event))
+                .await
+                .expect("scripted provider");
+
+        let reasoning_events = events
+            .iter()
+            .filter_map(|event| event.reasoning_text.as_deref())
+            .collect::<Vec<_>>();
+        assert_eq!(reasoning_events, vec!["", "先分析需求，再组织方案。"]);
+        assert_eq!(response.text, "正文");
+        assert!(session.history.is_empty());
+    }
+
+    #[tokio::test(flavor = "current_thread")]
+    async fn scripted_design_provider_retry_clears_previous_reasoning_attempt() {
+        let (_temp, root, _resources) = init_design_project();
+        let mut session = new_design_session("design-fake", "quality");
+        begin_design_turn(&mut session, "turn-reasoning-retry");
+        let mut response = fake_response("reasoning-retry", "重试后的正文", Vec::new());
+        response.reasoning = "重试后的推理".into();
+        let _fake = fake_provider::install(
+            vec![
+                Err(platform_llm::LlmError::Upstream {
+                    status_code: 503,
+                    message: "busy".into(),
+                }),
+                Ok(response),
+            ],
+            1,
+        );
+        let mut events = Vec::new();
+        let response =
+            request_scripted_design_provider(&root, &mut session, &mut |event| events.push(event))
+                .await
+                .expect("scripted retry provider");
+
+        let reasoning_events = events
+            .iter()
+            .filter_map(|event| event.reasoning_text.as_deref())
+            .collect::<Vec<_>>();
+        assert_eq!(reasoning_events, vec!["", "", "重试后的推理"]);
+        assert_eq!(response.text, "重试后的正文");
+        assert!(session.history.is_empty());
+    }
+
     #[tokio::test(flavor = "current_thread")]
     async fn fake_provider_walks_five_phases_and_enters_consultant() {
         let (_temp, root, resources) = init_design_project();

apps/ai-game-creator-shell/src-tauri/src/agent/interaction.rs

@@ -409,6 +409,8 @@ where
             (self.on_delta)(&platform_llm::LlmStreamDelta {
                 accumulated_text,
                 delta_text,
+                accumulated_reasoning: String::new(),
+                reasoning_delta: String::new(),
                 finish_reason,
             });
         }
@@ -499,6 +501,7 @@ mod tests {
             provider: LlmProvider::OpenAiCompatible,
             model: "interaction-test".to_string(),
             text: text.to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: Some("interaction-response".to_string()),
             usage: None,

apps/ai-game-creator-shell/src-tauri/src/agent/runtime_actions/response_stream_tests.rs

@@ -115,6 +115,7 @@ fn persist_tool_plan_handoff_repair_chain(
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: llm.model.clone(),
         text: text.to_string(),
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id: None,
         usage: None,
@@ -146,6 +147,8 @@ fn stream_delta(delta_text: &str, accumulated_text: &str) -> platform_llm::LlmSt
     platform_llm::LlmStreamDelta {
         accumulated_text: accumulated_text.to_string(),
         delta_text: delta_text.to_string(),
+        accumulated_reasoning: String::new(),
+        reasoning_delta: String::new(),
         finish_reason: None,
     }
 }
@@ -1003,6 +1006,7 @@ async fn provider_handoff_identity_drift_closes_lifecycle_without_leaking_respon
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: old_llm.model.clone(),
             text: private_response.to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: None,
             usage: None,
@@ -1116,6 +1120,7 @@ async fn tool_plan_handoff_identity_drift_closes_entire_repair_chain_before_remo
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: old_llm.model.clone(),
         text: text.to_string(),
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id: None,
         usage: None,
@@ -1302,6 +1307,7 @@ async fn tool_plan_capacity_gate_runs_before_provider_lifecycle_and_network() {
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: llm.model.clone(),
             text: format!("capacity response {loop_iteration}"),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: None,
             usage: None,
@@ -1436,6 +1442,7 @@ async fn tool_plan_handoff_durable_control_closes_entire_repair_chain_before_rem
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: llm.model.clone(),
         text: text.to_string(),
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id: None,
         usage: None,
@@ -1552,6 +1559,7 @@ fn provider_recovery_cleanup_closes_tool_plan_lifecycle_before_removing_handoff(
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: llm.model.clone(),
             text: "cleanup handoff".to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: None,
             usage: None,
@@ -1623,6 +1631,7 @@ fn runtime_resume_scans_and_cleans_terminal_tool_plan_handoff() {
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: llm.model.clone(),
             text: "terminal handoff".to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: None,
             usage: None,
@@ -1718,6 +1727,7 @@ async fn provider_handoff_retry_conflict_preserves_both_sidecars_for_reconciliat
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: llm.model.clone(),
             text: "已成功但尚未消费的回复".to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: None,
             usage: None,

apps/ai-game-creator-shell/src-tauri/src/agent/runtime_protocol/provider_control.rs

@@ -798,6 +798,7 @@ mod provider_reconciliation_diagnostic_tests {
         let response = platform_llm::LlmRunResponse {
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: "test-model".to_string(),
+            reasoning: String::new(),
             text: "C:\\private\\response".to_string(),
             finish_reason: Some("completed".to_string()),
             response_id: Some("response-1".to_string()),

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

@@ -905,6 +905,7 @@ mod tests {
             provider: platform_llm::LlmProvider::OpenAiCompatible,
             model: "context-compaction-test".to_string(),
             text: summary.into(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: Some("context-compaction-response".to_string()),
             usage: Some(platform_llm::LlmTokenUsage {

apps/ai-game-creator-shell/src-tauri/src/project/write_lock.rs

@@ -14,59 +14,16 @@ const PROJECT_WRITE_LOCK_UNWRITTEN_GRACE_SECONDS: u64 = 30;
 const PROJECT_WRITE_LOCK_PID_REUSE_TOLERANCE_SECONDS: u64 = 5;
 const PROJECT_WRITE_LOCK_MAX_BYTES: u64 = 4 * 1024;
 
-/// 本进程内真正落盘持有项目写锁的线程登记表。
-///
-/// `.agent/project.lock` 的 `pid` 只能证明“锁由本进程的某条写通道持有”，它分不清
-/// 两种完全不同的局面：
-/// - **同一条调用链再次取锁**：持锁方就是自己，必须放行，否则每次嵌套项目写入都要
-///   白等一个等待预算再报“项目正在被其他写操作占用”；
-/// - **本进程另一条写通道正在写**：项目 revision 侧车、steer 序号、一致快照读、
-///   pending sidecar 复核和恢复安装都靠这把锁串行化，必须照旧等待。
-///
-/// 复用判据因此不能停在 `pid`：只有**当前线程**就是真实持锁线程时才返回 advisory
-/// guard，本进程其余争用继续走有界等待与终态占用。登记按路径进行、按路径注销：
-/// guard 可能被移到别的线程再 Drop（例如写入路径把锁交给阻塞线程池的持有者），
-/// 按线程注销会漏项，让后续的重入判断失真。
-static PROJECT_WRITE_LOCK_THREAD_OWNERS: std::sync::Mutex<Vec<(PathBuf, std::thread::ThreadId)>> =
-    std::sync::Mutex::new(Vec::new());
-
-fn project_write_lock_thread_owners(
-) -> std::sync::MutexGuard<'static, Vec<(PathBuf, std::thread::ThreadId)>> {
-    // 登记表只是复用判据的加速器：中毒时继续用内部值，不能让一次取锁失败升级成
-    // 整个进程再也写不了项目。
-    PROJECT_WRITE_LOCK_THREAD_OWNERS
-        .lock()
-        .unwrap_or_else(|poisoned| poisoned.into_inner())
-}
-
-fn register_project_write_lock_thread_owner(path: &Path) {
-    let mut owners = project_write_lock_thread_owners();
-    if owners.iter().any(|(owner, _)| owner == path) {
-        return;
-    }
-    owners.push((path.to_path_buf(), std::thread::current().id()));
-}
-
-fn unregister_project_write_lock_thread_owner(path: &Path) {
-    project_write_lock_thread_owners().retain(|(owner, _)| owner != path);
-}
-
-/// 当前线程是否就是这条锁路径上真实落盘的持有者（同线程重入）。
-fn project_write_lock_reentered_by_current_thread(path: &Path) -> bool {
-    let thread = std::thread::current().id();
-    project_write_lock_thread_owners()
-        .iter()
-        .any(|(owner, owner_thread)| owner == path && *owner_thread == thread)
-}
-
 #[derive(Debug)]
 pub(crate) struct ProjectWriteLock {
     path: PathBuf,
     content: String,
-    /// 两种“本进程持锁但不必自等”的争用会拿到 advisory guard：同一线程重入（同一条
+    /// In the free-form autonomous lane a single Runtime process may have
-    /// 调用链再次取锁）和自主游戏构建流水线（它有意让并行专家动作同时在飞）。这两种
+    /// several specialist actions in flight at once.  A file lock is still
-    /// 情况下争用是进程内重叠而不是另一个客户端在改项目，返回的 guard 不拥有
+    /// useful across processes, but making same-process contenders fail turns
-    /// `.agent/project.lock`，Drop 时也不得删除真实持有者的锁。
+    /// ordinary parallel work into a dead run (and can deadlock nested tool
+    /// calls).  Such a contender receives an in-process/advisory guard instead
+    /// of deleting the real holder's lock on drop.
     bypassed_same_process: bool,
 }
 
@@ -90,7 +47,6 @@ impl Drop for ProjectWriteLock {
         if self.bypassed_same_process {
             return;
         }
-        unregister_project_write_lock_thread_owner(&self.path);
         if fs::read_to_string(&self.path).is_ok_and(|content| content == self.content) {
             let _ = fs::remove_file(&self.path);
         }
@@ -859,7 +815,6 @@ pub(crate) fn acquire_project_write_lock_failure(
                         path.display()
                     )));
                 }
-                register_project_write_lock_thread_owner(&path);
                 return Ok(ProjectWriteLock {
                     path,
                     content: content.clone(),
@@ -905,15 +860,11 @@ pub(crate) fn acquire_project_write_lock_failure(
                         }
                     }
                 }
-                if project_write_lock_is_owned_by_current_process(&path)
+                if project_write_lock_is_owned_by_current_process(&path) {
-                    && (crate::agent::autonomous_game_build_root_run_active_at(root)
+                    // A project lock is the client-use lock.  Nested calls in
-                        || project_write_lock_reentered_by_current_thread(&path))
+                    // the same client process must reuse that ownership instead
-                {
+                    // of waiting on their own durable marker.  Cross-process
-                    // 持锁方就是本进程自己时必须区分重入与并发：同一条调用链（同一
+                    // contenders still take the normal retryable path.
-                    // 线程）再次取锁，以及自主流水线有意并行专家动作，返回 advisory
-                    // guard、不自等、不动真实锁；本进程**其它线程**正在写则继续走
-                    // 有界等待，保住 revision 侧车、steer 序号、一致快照读与恢复安装
-                    // 的串行化。
                     return Ok(ProjectWriteLock {
                         path,
                         content: String::new(),

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

@@ -51,6 +51,7 @@ impl AgentRuntimeProviderHandoffRecord {
             provider: self.response.provider,
             model: self.response.model.clone(),
             text: self.response.text.clone(),
+            reasoning: String::new(),
             finish_reason: self.response.finish_reason.clone(),
             response_id: self.response.response_id.clone(),
             usage: self.response.usage.clone(),
@@ -340,6 +341,7 @@ mod tests {
             provider: LlmProvider::OpenAiCompatible,
             model: "handoff-model".to_string(),
             text: text.to_string(),
+            reasoning: String::new(),
             finish_reason: Some("stop".to_string()),
             response_id: Some("response-handoff".to_string()),
             usage: Some(LlmTokenUsage {

apps/ai-game-creator-shell/src-tauri/src/runner/tests.rs

@@ -2775,6 +2775,7 @@ fn durable_provider_handoff_prevents_shutdown_even_when_corrupt() {
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: "provider-handoff-runner-test".to_string(),
         text: "durable final reply".to_string(),
+        reasoning: String::new(),
         finish_reason: Some("stop".to_string()),
         response_id: Some("provider-handoff-response".to_string()),
         usage: None,

apps/ai-game-creator-shell/src-tauri/src/tests/mod.rs

@@ -4475,6 +4475,7 @@ fn real_e2e_tool_plan_checkpoint_response() -> platform_llm::LlmRunResponse {
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: "real-e2e-checkpoint-model".to_string(),
         text: REAL_E2E_TOOL_PLAN_CHECKPOINT_PRIVATE_RESPONSE.to_string(),
+        reasoning: String::new(),
         finish_reason: Some("tool_calls".to_string()),
         response_id: Some("real-e2e-checkpoint-private-response-id".to_string()),
         usage: None,
@@ -4720,6 +4721,7 @@ fn agent_tool_plan_llm_response(
         provider: platform_llm::LlmProvider::OpenAiCompatible,
         model: "mock-game-model".to_string(),
         text: text.into(),
+        reasoning: String::new(),
         finish_reason: Some("tool_calls".to_string()),
         response_id: Some("response-tool-plan-test".to_string()),
         usage: None,

apps/ai-game-creator-shell/src-tauri/src/tests/project_tools.rs

@@ -5818,27 +5818,8 @@ async fn agent_runtime_file_write_lock_failure_redacts_project_path() {
         },
     )
     .expect("allow direct file write");
-    // 持锁方必须是**另一条线程**：本用例验证的是“别的写通道正在写时 file.write 必须
+    let lock = acquire_project_write_lock(&root, "persistent-writer")
-    // 走满等待预算并失败关闭”，同一条调用链自持锁属于重入复用，不会失败。
-    let holder_root = root.clone();
-    let (release_sender, release_receiver) = mpsc::channel::<()>();
-    let holder = std::thread::spawn(move || {
-        let lock = acquire_project_write_lock(&holder_root, "persistent-writer")
         .expect("acquire persistent project writer");
-        let _ = release_receiver.recv();
-        drop(lock);
-    });
-    let lock_path = root.join(PROJECT_WRITE_LOCK_PATH);
-    for _ in 0..400 {
-        if lock_path.is_file() {
-            break;
-        }
-        std::thread::sleep(Duration::from_millis(5));
-    }
-    assert!(
-        lock_path.is_file(),
-        "persistent writer must hold the project write lock"
-    );
 
     let observation = execute_game_creator_agent_runtime_tool_action(
         &root,
@@ -5856,8 +5837,7 @@ async fn agent_runtime_file_write_lock_failure_redacts_project_path() {
     )
     .await;
 
-    let _ = release_sender.send(());
+    drop(lock);
-    holder.join().expect("join persistent project writer");
     assert_eq!(observation.status, "failed");
     assert!(!observation
         .summary

apps/ai-game-creator-shell/src-tauri/src/tool_plan_handoff/model.rs

@@ -127,6 +127,7 @@ impl AgentRuntimeToolPlanHandoffEntry {
             provider: self.response.provider,
             model: self.response.model.clone(),
             text,
+            reasoning: String::new(),
             finish_reason: self.response.finish_reason.clone(),
             response_id: self.response.response_id.clone(),
             usage: self.response.usage.as_ref().map(LlmTokenUsage::from),

apps/ai-game-creator-shell/src-tauri/src/tool_plan_handoff/tests.rs

@@ -84,6 +84,7 @@ fn response(text: &str, tool_calls: Vec<LlmToolCall>) -> LlmRunResponse {
         provider: LlmProvider::OpenAiCompatible,
         model: "tool-plan-handoff-model".to_string(),
         text: text.to_string(),
+        reasoning: String::new(),
         finish_reason: Some("tool_calls".to_string()),
         response_id: Some("tool-plan-handoff-response".to_string()),
         usage: Some(LlmTokenUsage {

apps/ai-game-creator-shell/src-tauri/src/ui_editor/persistence.rs

@@ -1214,28 +1214,8 @@ mod tests {
             .expect("resolve primary");
         fs::write(&primary, b"{broken").expect("corrupt primary");
 
-        // 持锁方必须是**另一条线程**：本用例验证的是“另一个写者持锁时恢复安装必须失败
+        let project_lock = acquire_project_write_lock(directory.path(), "test.concurrent-save")
-        // 关闭”，同一条调用链自持锁属于重入复用，不再产生占用失败。
-        let holder_root = directory.path().to_path_buf();
-        let (release_sender, release_receiver) = std::sync::mpsc::channel::<()>();
-        let holder = std::thread::spawn(move || {
-            let lock = acquire_project_write_lock(&holder_root, "test.concurrent-save")
             .expect("hold project write lock");
-            let _ = release_receiver.recv();
-            drop(lock);
-        });
-        let lock_path = resolve_local_project_path(directory.path(), PROJECT_WRITE_LOCK_PATH)
-            .expect("resolve project write lock path");
-        for _ in 0..400 {
-            if lock_path.is_file() {
-                break;
-            }
-            std::thread::sleep(std::time::Duration::from_millis(5));
-        }
-        assert!(
-            lock_path.is_file(),
-            "concurrent writer must hold the project write lock"
-        );
         let error = load_ui_design_state_at(LoadUiDesignStateInput {
             project_path: directory.path().to_string_lossy().into_owned(),
             expected_project_id: PROJECT_ID.to_string(),
@@ -1244,8 +1224,7 @@ mod tests {
         .expect_err("recovery must not install while another writer holds the lock");
         assert!(error.contains("项目正在被其他写操作占用"));
         assert!(read_ui_design_document_path(&primary).is_err());
-        let _ = release_sender.send(());
+        drop(project_lock);
-        holder.join().expect("join concurrent writer");
 
         let recovered = load_ui_design_state_at(LoadUiDesignStateInput {
             project_path: directory.path().to_string_lossy().into_owned(),

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

@@ -596,6 +596,12 @@ export function App({
     projectPath: string;
     clientTurnId: string;
   } | null>(null);
+  const designAgentEventSubscriptionReadyRef = useRef<Promise<void> | null>(
+    null,
+  );
+  const designAgentEventSubscriptionResolveRef = useRef<(() => void) | null>(
+    null,
+  );
   // 做方案入口独立成链：立项策划需要委派、澄清 pending 与 GDD 审批，这些只存在于
   // Supervisor Runtime；direct-codex 是单回合「生成→试玩→修」循环，没有对应机制。
   // 因此策划入口不走产品默认的 direct-codex，做游戏与做素材保持 master 的新默认。
@@ -821,7 +827,17 @@ export function App({
     useState('');
   const planningV2TransientReplyTargetRef = useRef('');
   const planningV2VisibleReplyRef = useRef('');
+  const designAgentPendingViewRef = useRef<{
+    clientTurnId: string;
+    projectPath: string;
+    view: DesignView;
+  } | null>(null);
   const [planningV2Reasoning, setPlanningV2Reasoning] = useState('');
+  const designAgentReasoningTurnRef = useRef<{
+    projectPath: string;
+    clientTurnId: string;
+    text: string;
+  } | null>(null);
   const planningV2TurnRef = useRef<{
     projectPath: string;
     clientTurnId: string;
@@ -849,6 +865,22 @@ export function App({
     }
   }
 
+  function designAgentEventSubscriptionReady() {
+    if (!designAgentEventSubscriptionReadyRef.current) {
+      designAgentEventSubscriptionReadyRef.current = new Promise<void>(
+        (resolve) => {
+          designAgentEventSubscriptionResolveRef.current = resolve;
+        },
+      );
+    }
+    return designAgentEventSubscriptionReadyRef.current;
+  }
+
+  function resolveDesignAgentEventSubscriptionReady() {
+    designAgentEventSubscriptionResolveRef.current?.();
+    designAgentEventSubscriptionResolveRef.current = null;
+  }
+
   useEffect(() => {
     const timer = window.setInterval(() => {
       const target = planningV2TransientReplyTargetRef.current;
@@ -1005,6 +1037,66 @@ export function App({
     latestMessagesRef.current = conversation;
   }
 
+  function commitDesignAgentView(view: DesignView, projectPath: string) {
+    const pendingTurnId = designAgentPendingViewRef.current?.clientTurnId;
+    designAgentPendingViewRef.current = null;
+    applyDesignView(view, projectPath);
+    setPlanningV2TransientReplyTarget('');
+    if (designAgentTurnRef.current?.clientTurnId === pendingTurnId) {
+      designAgentTurnRef.current = null;
+    }
+  }
+
+  function applyDesignAgentViewAfterTransient(
+    view: DesignView,
+    projectPath: string,
+    clientTurnId: string,
+  ) {
+    let target = planningV2TransientReplyTargetRef.current;
+    const tracked = designAgentTurnRef.current;
+    if (!target.trim() && !view.running) {
+      const latestAssistantText = [...view.messages]
+        .reverse()
+        .find((message) => message.role !== 'user' && message.text.trim())
+        ?.text.trim();
+      if (latestAssistantText) {
+        setPlanningV2TransientReplyTarget(latestAssistantText);
+        target = latestAssistantText;
+      }
+    }
+    if (
+      !view.running &&
+      tracked?.clientTurnId === clientTurnId &&
+      target.trim() &&
+      planningV2VisibleReplyRef.current !== target
+    ) {
+      designAgentPendingViewRef.current = {
+        clientTurnId,
+        projectPath,
+        view,
+      };
+      return;
+    }
+    commitDesignAgentView(view, projectPath);
+  }
+
+  useEffect(() => {
+    const timer = window.setInterval(() => {
+      const pending = designAgentPendingViewRef.current;
+      if (!pending) {
+        return;
+      }
+      const target = planningV2TransientReplyTargetRef.current;
+      if (target && planningV2VisibleReplyRef.current !== target) {
+        return;
+      }
+      commitDesignAgentView(pending.view, pending.projectPath);
+    }, 50);
+    return () => window.clearInterval(timer);
+    // 收尾定时器只需注册一次；它读取 refs，避免随每次渲染重建。
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []);
+
   async function hydrateDesignAgentSession(nextProjectPath: string) {
     const invoke = resolveTauriInvoke();
     if (!invoke || !nextProjectPath.trim()) {
@@ -1038,9 +1130,17 @@ export function App({
       projectPath: nextProjectPath,
       clientTurnId,
     };
+    designAgentReasoningTurnRef.current = {
+      projectPath: nextProjectPath,
+      clientTurnId,
+      text: '',
+    };
+    designAgentPendingViewRef.current = null;
+    await designAgentEventSubscriptionReady();
     setChatAgentBusy(true);
     setProjectSupervisorRuntimeError('');
     setPlanningV2TransientReplyTarget('');
+    setPlanningV2Reasoning('');
     try {
       const view = await invoke<DesignView>('continue_design_agent_session', {
         projectPath: nextProjectPath,
@@ -1050,7 +1150,7 @@ export function App({
       if (localProjectPathRef.current !== nextProjectPath) {
         return;
       }
-      applyDesignView(view, nextProjectPath);
+      applyDesignAgentViewAfterTransient(view, nextProjectPath, clientTurnId);
     } catch (error) {
       if (localProjectPathRef.current !== nextProjectPath) {
         return;
@@ -1062,8 +1162,10 @@ export function App({
       setProjectSupervisorRuntimeError(message);
       setPlanGddError(message);
     } finally {
+      if (!designAgentPendingViewRef.current) {
         designAgentTurnRef.current = null;
         setPlanningV2TransientReplyTarget('');
+      }
       setChatAgentBusy(false);
     }
   }
@@ -1555,6 +1657,9 @@ export function App({
     setProjectSupervisorRuntimeError('');
     setPlanningV2Session(null);
     setPlanningV2TransientReplyTarget('');
+    designAgentPendingViewRef.current = null;
+    designAgentReasoningTurnRef.current = null;
+    setPlanningV2Reasoning('');
     setPlanningV2Active(planningStartMode);
     planningV2ActiveRef.current = planningStartMode;
     designAgentLaneRef.current = planningStartMode;
@@ -1990,8 +2095,14 @@ export function App({
   }, [planningV2Active]);
 
   useEffect(() => {
+    const ready = designAgentEventSubscriptionReady();
     if (!canSubscribeTauriEvents() || !planningV2Active) {
-      return;
+      resolveDesignAgentEventSubscriptionReady();
+      return () => {
+        if (designAgentEventSubscriptionReadyRef.current === ready) {
+          designAgentEventSubscriptionReadyRef.current = null;
+        }
+      };
     }
     let cleanup: (() => void) | null = null;
     let disposed = false;
@@ -2008,26 +2119,45 @@ export function App({
         setPlanningV2TransientReplyTarget(payload.text);
       }
       if (payload.reasoningText != null) {
+        const reasoningTurn = designAgentReasoningTurnRef.current;
+        if (
+          reasoningTurn &&
+          reasoningTurn.projectPath === payload.projectPath &&
+          reasoningTurn.clientTurnId === payload.clientTurnId
+        ) {
+          reasoningTurn.text = payload.reasoningText;
           setPlanningV2Reasoning(payload.reasoningText);
         }
+      }
       if (payload.kind === 'tool' && payload.text) {
         setPlanningV2TransientReplyTarget(payload.text);
       }
       if (payload.view) {
-        applyDesignView(payload.view, payload.projectPath);
+        applyDesignAgentViewAfterTransient(
+          payload.view,
+          payload.projectPath,
+          payload.clientTurnId,
+        );
       }
     })
       .then((unlisten) => {
+        resolveDesignAgentEventSubscriptionReady();
         if (disposed) {
           unlisten();
           return;
         }
         cleanup = unlisten;
       })
-      .catch(() => undefined);
+      .catch(() => {
+        resolveDesignAgentEventSubscriptionReady();
+      });
     return () => {
       disposed = true;
       cleanup?.();
+      resolveDesignAgentEventSubscriptionReady();
+      if (designAgentEventSubscriptionReadyRef.current === ready) {
+        designAgentEventSubscriptionReadyRef.current = null;
+      }
     };
     // applyDesignView 读的是 refs 和当前项目路径，
     // 把它写进依赖会在每轮回复时重订事件。
@@ -6072,6 +6202,7 @@ export function App({
     setChatAgentBusy(true);
     setProjectSupervisorRuntimeError('');
     setPlanningV2TransientReplyTarget('');
+    setPlanningV2Reasoning('');
     try {
       const result = currentSessionId
         ? await invoke<PlanningSessionCommandResultV2>(
@@ -11824,15 +11955,25 @@ export function App({
                   projectPath: nextProjectPath,
                   clientTurnId,
                 };
+                designAgentReasoningTurnRef.current = {
+                  projectPath: nextProjectPath,
+                  clientTurnId,
+                  text: '',
+                };
+                designAgentPendingViewRef.current = null;
                 setPlanningV2TransientReplyTarget('');
+                setPlanningV2Reasoning('');
                 setChatAgentBusy(true);
                 setPlanGddDecisionBusy(true);
-                void invoke<DesignView>('decide_design_phase', {
+                void designAgentEventSubscriptionReady()
+                  .then(() =>
+                    invoke<DesignView>('decide_design_phase', {
                       projectPath: nextProjectPath,
                       clientTurnId,
                       requestId,
                       approved,
-                })
+                    }),
+                  )
                   .then((view) => {
                     if (
                       localProjectPathRef.current !== nextProjectPath ||
@@ -11842,7 +11983,11 @@ export function App({
                     ) {
                       return;
                     }
-                    applyDesignView(view, nextProjectPath);
+                    applyDesignAgentViewAfterTransient(
+                      view,
+                      nextProjectPath,
+                      clientTurnId,
+                    );
                   })
                   .catch((error) => {
                     if (
@@ -11864,8 +12009,10 @@ export function App({
                     ) {
                       return;
                     }
+                    if (!designAgentPendingViewRef.current) {
                       designAgentTurnRef.current = null;
                       setPlanningV2TransientReplyTarget('');
+                    }
                     setChatAgentBusy(false);
                     setPlanGddDecisionBusy(false);
                   });

apps/ai-game-creator-shell/src/features/project-workspace/ProjectSupervisorView.tsx

@@ -263,8 +263,11 @@ export function ProjectSupervisorView({
             </div>
           ))}
           {designReasoning ? (
-            <details className="design-agent-reasoning">
+            <details
-              <summary>显示思考过程</summary>
+              className="design-agent-reasoning"
+              aria-label="策划 Agent 思考过程"
+            >
+              <summary>思考过程（点击展开）</summary>
               <pre>{designReasoning}</pre>
             </details>
           ) : null}

apps/ai-game-creator-shell/tests/appSurface/design-agent.suite.ts

@@ -7,6 +7,7 @@ import {
   React,
   render,
   screen,
+  setComposerText,
   waitFor,
 } from './harness';
 
@@ -55,6 +56,24 @@ function designClarificationView() {
   };
 }
 
+function designConversationView() {
+  return {
+    session: {
+      sessionId: 'design-session-reasoning',
+      projectId: 'local-project-draft',
+      currentPhase: 'concept',
+      approvedPhases: [],
+      pendingApproval: null,
+      pendingClarification: null,
+      turnIndex: 1,
+      lastError: null,
+    },
+    messages: [],
+    running: false,
+    canRetry: false,
+  };
+}
+
 export function registerDesignAgentSurfaceTests() {
   it('hydrates an existing design session and decides approval through design commands', async () => {
     const harness = createProjectSupervisorRuntimeHarness({
@@ -141,4 +160,54 @@ export function registerDesignAgentSurfaceTests() {
       ).toBe(true);
     });
   });
+
+  it('keeps the current turn reasoning after completion and supports collapse/expand', async () => {
+    const harness = createProjectSupervisorRuntimeHarness({
+      designAgentView: designConversationView(),
+      designAgentContinueView: designConversationView(),
+    });
+    window.__TAURI__ = {
+      core: { invoke: harness.invoke },
+      event: { listen: harness.listen },
+    };
+    window.history.pushState({}, '', '/');
+    render(
+      React.createElement(App, {
+        initialProjectPath: harness.projectPath,
+        orchestrationMode: 'single-supervisor',
+        planningStartMode: true,
+        projectSupervisorOnly: true,
+      }),
+    );
+
+    const input = await screen.findByLabelText('项目需求');
+    await setComposerText(input, '请给出核心玩法方案');
+    fireEvent.submit(input.closest('form') as HTMLFormElement);
+    await waitFor(() => {
+      expect(
+        harness.invoke.mock.calls.some(
+          ([command]) => command === 'continue_design_agent_session',
+        ),
+      ).toBe(true);
+    });
+    const continueCall = [...harness.invoke.mock.calls]
+      .reverse()
+      .find(([command]) => command === 'continue_design_agent_session');
+    const clientTurnId = String(
+      (continueCall?.[1] as { clientTurnId?: string }).clientTurnId,
+    );
+    harness.emitDesignAgentEvent({
+      projectPath: harness.projectPath,
+      clientTurnId,
+      kind: 'reasoning',
+      reasoningText: '先分析需求，再组织方案。',
+    });
+
+    const summary = await screen.findByText('思考过程（点击展开）');
+    const details = summary.closest('details') as HTMLDetailsElement;
+    expect(details.open).toBe(false);
+    fireEvent.click(summary);
+    expect(details.open).toBe(true);
+    expect(screen.getByText('先分析需求，再组织方案。')).not.toBeNull();
+  });
 }

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

@@ -741,6 +741,9 @@ function createProjectSupervisorRuntimeHarness({
         };
       }) => void)
     | null = null;
+  let designAgentUpdateHandler:
+    | ((event: { payload: Record<string, unknown> }) => void)
+    | null = null;
 
   const conversationRecord = (
     role: 'user' | 'assistant',
@@ -1113,6 +1116,10 @@ function createProjectSupervisorRuntimeHarness({
       if (eventName === 'game-creator-agent-progress') {
         progressHandler = handler as unknown as typeof progressHandler;
       }
+      if (eventName === 'design-agent-update') {
+        designAgentUpdateHandler =
+          handler as unknown as typeof designAgentUpdateHandler;
+      }
       return () => {
         if (runtimeUpdateHandler === handler) {
           runtimeUpdateHandler = null;
@@ -1123,6 +1130,9 @@ function createProjectSupervisorRuntimeHarness({
         if (progressHandler === handler) {
           progressHandler = null;
         }
+        if (designAgentUpdateHandler === handler) {
+          designAgentUpdateHandler = null;
+        }
       };
     },
   );
@@ -1149,6 +1159,9 @@ function createProjectSupervisorRuntimeHarness({
     setPlanningV2Result(state: Record<string, unknown> | null) {
       currentPlanningV2Result = state;
     },
+    emitDesignAgentEvent(payload: Record<string, unknown>) {
+      designAgentUpdateHandler?.({ payload });
+    },
     setPlanningV2StartResult(state: Record<string, unknown> | null) {
       currentPlanningV2StartResult = state;
     },

docs/project-memory/plans/【实施计划】项目客户端占用锁收敛-2026-09-14.md

@@ -15,16 +15,14 @@ Milestone: `【里程碑】项目客户端占用锁收敛-2026-09-14.md`
 ## 修改顺序
 
 1. 统一同进程嵌套调用的项目锁语义，禁止自等待。
-2. 收窄复用判据：按 `pid` 放行会放过本进程其它线程的并行写，改为按“当前线程就是真实持锁线程”判定重入，并保住同进程跨线程的等待与终态占用。
+2. 盘点并迁移 Runner 的项目级 owner 文件到统一锁，保留诊断投影与跨 boot 恢复。
-3. 盘点并迁移 Runner 的项目级 owner 文件到统一锁，保留诊断投影与跨 boot 恢复。
+3. 删除重复项目级锁路径及其专属调用，保留底层原子写和 Git 锁。
-4. 删除重复项目级锁路径及其专属调用，保留底层原子写和 Git 锁。
+4. 补齐同进程重入、跨进程占用、崩溃恢复和锁释放测试。
-5. 补齐同进程重入、同进程跨线程争用、跨进程占用、崩溃恢复和锁释放测试。
 
 ## 验证命令
 
 - `cargo fmt --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml -- --check`
-- `cargo test --locked --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml -- --test-threads=1`
+- `cargo test --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml project_write_lock_reuses_same_process_owner_and_releases_on_drop --no-default-features`
-- `cargo test --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml project_write_lock --no-default-features`
 - Runner owner 与 response stream 相关定向测试
 - `npm run check:encoding`
 - `git diff --check`
@@ -33,5 +31,4 @@ Milestone: `【里程碑】项目客户端占用锁收敛-2026-09-14.md`
 
 - Runner 与 GUI 可能是不同进程；统一锁前必须验证同一客户端不会互相阻塞。
 - 旧 `.agent/runtime/execution-owner.lock` 残留需要按 PID/启动身份安全回收，不能直接删除。
-- 复用判据按线程判定：出现同进程跨线程重入的现场时先按 `*_locked` 入口处置，不要把判据退回按 `pid` 一律放行（那会放过并行写，见里程碑「边界」末条）。
 - 若跨 boot 恢复或 GUI/Runner 联动回归，回滚统一路径迁移，保留已验证的同进程重入修复。

docs/project-memory/plans/【里程碑】Provider推理与正文分离及策划Agent展示-2026-09-14.md

@@ -0,0 +1,200 @@
+# 【里程碑】Provider 推理与正文分离及策划 Agent 展示
+
+| 字段 | 值 |
+| --- | --- |
+| Version | 1.0 |
+| Status | in-progress |
+| Date | 2026-09-14 |
+| Parent Spec | `docs/technical/【技术方案】策划Agent生产迁移与工作区浏览-2026-09-10.md` |
+| Related Issue | `GenarrativeAI/Genarrative#331` |
+
+## 一句话交付结果
+
+让策划 Agent 能在流式回合中单独收到 Provider reasoning，并在 UI 中以默认折叠的思考过程展示；用户可见正文、工具调用和 GameAgent 现有行为保持不变。
+
+## 当前实现进度（2026-09-14）
+
+- 已完成共享 reasoning 字段、Provider 解析、策划事件映射以及正文流式收尾的前两轮提交。
+- 当前第三轮聚焦策划 Agent 前端 reasoning 生命周期：按 `projectPath + clientTurnId` 绑定事件，回合结束后保留本轮 reasoning，下一轮或项目切换时清理。
+- 现有 `<details>` 展示保持默认收起，并提供明确的展开/收起入口；不新增持久化字段，也不改动 GameAgent、Direct/Codex、supervisor 或已退役链路。
+
+## 背景与现状
+
+- `platform-llm` 当前只向上层提供正文累计值、正文增量和结束状态。
+- Chat 兼容响应中的 `reasoning`、`reasoning_content` 以及 reasoning content part 会被正文提取器过滤。
+- Responses 响应中的 reasoning 类型 output item 也不会进入独立的上层字段。
+- 策划 Agent 已经预留 `DesignEvent.reasoningText`、`planningV2Reasoning` 和默认折叠 UI，但 Provider 解析链没有产出数据，因此折叠区通常不出现。
+- GameAgent 当前只消费 `delta_text`、`accumulated_text` 和 `finish_reason`，没有消费策划 Agent 的 `reasoningText`。
+
+## 目标
+
+1. 为 Provider 流式响应增加独立 reasoning 增量和累计通道。
+2. 为非流式终态响应提供独立 reasoning 字段。
+3. 支持 Responses 和 Chat 兼容协议的 reasoning 解析。
+4. 仅由策划 Agent 显式启用 reasoning 捕获和 UI 转发。
+5. 保证 reasoning 不进入用户可见正文、工具调用参数或 GameAgent 消息流。
+6. 在无 reasoning、reasoning 解析异常、重试和工具调用共存场景下保持可恢复行为。
+
+## 非目标
+
+- 不改变 GameAgent 的正文展示、工具调用、`<think>` 过滤和运行时状态语义。
+- 不把 reasoning 自动拼接到 `delta_text`、`accumulated_text` 或正式 assistant 消息。
+- 不把 reasoning 作为新的业务消息类型写入策划会话历史。
+- 不新增通用 reasoning UI，不改造 Direct/Codex 的过程卡展示。
+- 不修改 Provider 请求模型、推理档位或 token 预算。
+- 不为 reasoning 增加新的 SpacetimeDB 表、公开 API 或持久化 schema。
+
+## 受影响模块与边界
+
+### Provider 共享层
+
+`server-rs/crates/platform-llm` 负责协议解析和流式累计：
+
+- `LlmStreamDelta` 增加 `reasoning_delta` 与 `accumulated_reasoning`。
+- `LlmRunResponse` 增加终态 reasoning 字段。
+- `LlmRunRequest` 增加默认关闭的 reasoning 捕获开关。
+- 正文提取继续排除隐藏 reasoning part；reasoning 进入旁路字段。
+- reasoning 解析失败只丢弃 reasoning，不影响正文和工具调用。
+
+### 策划 Runtime
+
+`apps/ai-game-creator-shell/src-tauri/src/agent/design_runtime.rs` 仅在策划专用请求中打开 reasoning 捕获：
+
+- 流式 reasoning 更新映射到已有 `DesignEvent.reasoningText`。
+- 正文继续使用已有 `text` 事件。
+- 新回合、重试、项目切换和请求失败时清理旧 reasoning。
+- debug 记录与正文记录分开，内容受现有 debug 开关和长度限制约束。
+
+### 其它调用方
+
+GameAgent、Agent Interaction、Direct/Codex 适配层和通用 runtime 继续只读取正文字段。新增 reasoning 字段默认为空，不改变这些调用方的业务判断。
+
+### 前端
+
+复用现有 `ProjectSupervisorView` 的 `designReasoning` 和默认折叠 `<details>` 展示。只补事件生命周期和状态清理，不新建平行组件或平行状态协议。
+
+## 分步实施方案
+
+### 第一步：冻结共享契约与兼容开关
+
+明确字段语义、空值语义和捕获开关：
+
+- reasoning 字段只表示 Provider 返回的内部推理内容，不代表用户正文。
+- 捕获开关默认关闭；未启用时新增字段为空。
+- 正文、工具调用、finish reason 和 Responses 原生 output 的现有语义保持不变。
+- 该步只更新规范、类型定义和构造点，不接入策划 UI。
+
+验收重点：所有现有 Rust 构造点可编译，GameAgent 现有调用仍只依赖正文字段。
+
+### 第二步：实现 `platform-llm` 协议解析
+
+分别补齐：
+
+- Responses reasoning 增量事件；
+- Responses 终态 reasoning output item / summary；
+- Chat `reasoning`、`reasoning_content` 和 reasoning content part；
+- 正文与 reasoning 的独立累计；
+- reasoning 与正文、工具调用同时出现时的顺序和去重；
+- reasoning 解析失败时的降级行为。
+
+Responses 的原生 output 仍按当前方式保留，用于后续 Responses 会话回放；新增 reasoning 字段只用于上层展示和调试消费。
+
+验收重点：正文永远不含 reasoning；无 reasoning 的响应与当前行为一致。
+
+### 第三步：补齐共享适配层并锁定 GameAgent 不变
+
+更新 `LlmStreamDelta` 构造点、适配器和测试辅助函数，使它们为新增字段提供空值。检查并锁定：
+
+- GameAgent 正文流不读取 reasoning；
+- 工具调用判断不读取 reasoning；
+- Direct/Codex 过程卡不显示 reasoning；
+- 通用 response stream 过滤逻辑不因新增字段改变。
+
+验收重点：现有工具调用、正文流式、Direct 和 Agent Interaction 测试无行为回归。
+
+### 第四步：接通策划 Runtime 与现有 UI
+
+仅在策划 Agent Provider 请求中启用捕获开关：
+
+- 收到 reasoning 增量时发出独立 `reasoningText`；
+- 收到正文增量时继续发出原有 `text`；
+- 重试时替换同一回合的临时 reasoning，不残留上一 attempt；
+- 正式回合结束后保留本回合展示，下一回合开始时清理；
+- UI 默认折叠，展开后显示累计 reasoning，不影响正文滚动和输入。
+
+验收重点：策划 Agent 能看到独立 reasoning，正文气泡不重复、不混入推理文本。
+
+### 第五步：完成回归、文档和验收证据
+
+形成逐条证据矩阵，至少覆盖：
+
+- Responses reasoning 增量和终态；
+- Chat reasoning 字段和 content part；
+- 正文与 reasoning 分离；
+- reasoning 与工具调用并存；
+- reasoning 解析失败降级；
+- 无 reasoning 兼容行为；
+- 策划 Runtime 事件映射和 UI 生命周期；
+- GameAgent 正文与工具调用回归。
+
+## 第五轮验收证据
+
+| 验收面 | 证据 | 结果 |
+| --- | --- | --- |
+| Chat / Responses reasoning 解析 | `cargo test --manifest-path server-rs/Cargo.toml -p platform-llm` | PASS，152 个单元测试；含字段、content part、SSE 增量、终态快照和正文隔离 |
+| reasoning 与工具调用共存 | `responses_response_captures_reasoning_alongside_tool_call`、既有 Chat/Responses 流式工具测试 | PASS |
+| 默认关闭与请求兼容 | `run_request_defaults_to_openai_responses_api_kind`、`reasoning_capture_switch_does_not_change_provider_request_body` | PASS |
+| 策划 Runtime 生命周期 | `cargo test --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml --bin genarrative-ai-game-creator-shell design_runtime` | PASS，10 个测试；含事件映射、history 隔离、重试清理和失败清理 |
+| GameAgent / Direct/Codex 正文回归 | `cargo check --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml --tests` 与现有 response stream / direct tests 编译 | PASS；新增字段未进入正文消费路径 |
+| 前端与文档门禁 | `npx tsc -p apps/ai-game-creator-shell/tsconfig.json --noEmit`、`npm run check:encoding`、`npm run check:doc-index`、`git diff --check` | PASS |
+| 格式门禁 | `cargo fmt --all --manifest-path server-rs/Cargo.toml -- --check`、AGC Tauri 同命令 | PASS |
+
+真实 Provider、浏览器运行时 smoke 和 `check-config.mjs` 的 Windows 私有 DACL 路径本轮未验证；前者需要凭据和运行环境，后者受当前沙箱权限限制，不能据此扩大验收结论。
+
+## 契约与持久化策略
+
+- 不修改 HTTP API、OpenAPI、SpacetimeDB schema 或生成绑定。
+- 不新增正式持久化字段；策划会话仍保存既有对话和 Responses 原生 output。
+- reasoning 捕获开关属于 Provider 请求的内部调用语义，默认关闭，不改变已有请求的默认指纹和展示行为。
+- reasoning 不作为下一轮普通用户可见正文回灌；Responses 原生 output 的恢复语义保持现状。
+
+## 失败、重试与恢复
+
+- reasoning 解析失败：保留正文和工具调用，reasoning 字段置空或保留已累计部分。
+- Provider 瞬态重试：reasoning 与正文使用同一回合、同一响应槽，新的 attempt 替换临时值。
+- 流中断：沿用现有 Provider 错误和策划会话恢复规则，不把未完成 reasoning 误判为正式消息。
+- UI 刷新或项目恢复：只从当前事件/状态恢复 reasoning，不隐式唤醒 Provider。
+
+## 风险与回滚点
+
+| 风险 | 控制措施 | 回滚点 |
+| --- | --- | --- |
+| 共享结构体新增字段导致构造点遗漏 | 先补齐所有构造点和编译检查 | 回退共享字段提交 |
+| Provider 把 reasoning 混入正文 | 保留独立提取器和正文过滤测试 | 关闭 reasoning 捕获开关 |
+| Responses summary 事件重复累计 | 以增量事件为主，终态仅做快照/兜底 | 关闭对应事件解析 |
+| GameAgent 意外展示 reasoning | 捕获默认关闭，调用方只读正文字段 | 回退策划开关，不影响共享解析 |
+| 重试残留旧 reasoning | 按回合和响应槽清理/替换 | 回退 UI 事件消费 |
+
+## 验收命令
+
+代码实现阶段按里程碑执行，不在本计划阶段运行业务测试。预计命令：
+
+```text
+cargo test -p platform-llm
+cargo test -p ai-game-creator-shell
+npm run typecheck
+npm run check:encoding
+git diff --check
+```
+
+文档阶段已要求补充运行：
+
+```text
+npm run check:doc-index
+npm run check:encoding
+git diff --check
+```
+
+## 当前状态与下一步
+
+当前仅完成问题定位和方案设计，未修改业务代码。进入实现前应先评审本里程碑的字段语义、默认关闭策略、Responses 事件覆盖范围和 reasoning 是否进入 debug 记录；评审通过后再为单个里程碑建立对应的 `【实施计划】` 文档。

docs/project-memory/plans/【里程碑】项目客户端占用锁收敛-2026-09-14.md

@@ -7,24 +7,22 @@ Parent Spec: `docs/technical/【技术方案】AI游戏创作智能体App实施
 
 ## 目标
 
-项目只保留一个面向客户端占用的项目级跨进程锁，防止多个客户端同时打开同一项目；同一客户端进程内**同一条写调用链（同一线程）的嵌套调用**复用既有项目锁，不因自身持锁进入等待。
+项目只保留一个面向客户端占用的项目级跨进程锁，防止多个客户端同时打开同一项目；同一客户端进程内的嵌套调用复用既有项目锁，不因自身持锁进入等待。
 
 ## 边界
 
 - 项目客户端占用锁与项目写入调用的职责统一，跨进程竞争仍返回占用语义。
 - Agent DB、session lane、manifest 原子写和 Git 自身的底层一致性机制不在本里程碑删除范围内。
-- 不改变项目 revision、权限、幂等、恢复和数据格式合同。**本进程其它线程的并发写入必须继续串行化**：按 `pid` 一律返回 advisory guard 会放过并行写，直接违反本边界（见验收标准第 2 条）。
+- 不改变项目 revision、权限、幂等、恢复和数据格式合同。
 
 ## 验收标准
 
-- 同一线程（同一条写调用链）嵌套取得项目锁立即返回 advisory guard，不等待、不删除真实持有者锁。
+- 同一进程内嵌套取得项目锁立即返回 advisory guard，不等待、不删除真实持有者锁。
-- 本进程另一条线程持锁（模拟“另一个写通道/另一个客户端”的既有用例形态）时仍保持等待与终态占用：项目 revision 侧车、steer 序号分配、一致快照读、pending sidecar 复核和恢复安装不得被复用判据放过。
 - 不同进程持有项目锁时仍保持占用失败与残留回收判据。
 - 客户端项目占用入口与 Runtime 写入入口不会各自维护第二个项目级锁文件。
 - 锁释放后下一客户端可重新取得锁。
-- 定向 Rust 锁测试、`cargo fmt --check`、`npm run check:encoding`、`git diff --check` 通过；锁语义变更必须跑 `cargo test --locked --manifest-path apps/ai-game-creator-shell/src-tauri/Cargo.toml -- --test-threads=1` 全量，定向用例覆盖不到 `project_tools` / `command_runtime` / `parallel_actions` / `runtime_state` / `response_stream` / `direct_tool_bridge` / `ui_editor::persistence` 里的锁不变量。
+- 定向 Rust 锁测试、`cargo fmt --check`、`npm run check:encoding`、`git diff --check` 通过。
 
 ## 未决事项
 
 - Runner 的 `execution-owner.lock` 如何迁移到统一客户端占用锁，需要补充跨进程启动、恢复和诊断测试后再落地。
-- 同进程**跨线程**重入（持锁调用链在 `await` / `spawn_blocking` 之后于其它线程再次取锁）仍会走有界等待，预算耗尽时报“项目正在被其他写操作占用”。发现这类现场时按 2026-08-27 的既有处置改用 `*_locked` 入口复用已有 guard（`project-memory/shared-memory/pitfalls.md`「持锁调用链二次取锁」），不放宽整条锁的串行化语义。

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

@@ -3,15 +3,6 @@
 > 用途：记录已经确认、会影响后续开发的长期技术/产品/协作决策。短期讨论不要写在这里。
 > 当前口径：历史条目的旧路径、旧版本和已退役对象只用于追溯，不构成现行实现依据；如与当前代码或 `docs/README.md` 冲突，以当前代码和最新专题文档为准。
 
-## 2026-09-14 客户端 CI 按门禁组拆成三个 job，AGC 的 web / rust 两段并行
-
-- 背景：`Project CI / Native shell tests` 把微信壳、Expo 移动壳、Tauri 桌面壳、H5 HostBridge 与 AI 游戏创作壳的全部门禁串在一个 job 里，实测 18 分 37 秒；同一次运行的 Repository / Frontend / Backend 分别只要 3 分 21 秒、4 分 16 秒、6 分 14 秒，其余三个 job 结束后客户端 job 还要再跑十几分钟。日志时间戳显示门禁段 932 秒里：AGC `ai-game-creator-shell:check` 占 654 秒（其中壳内 Rust 套件 2451 个用例 `--test-threads=1` 单跑 441.58 秒、编译 79 秒），AGC vitest 75 秒，两个发布构建 smoke 加落盘断言 230 秒，而 h5 / 微信 / 移动 / 桌面壳的全部运行时门禁加起来不到 50 秒。
-- 决策：`scripts/check-native-shells.mjs` 引入 `--groups=`，把门禁分成 `contract`（静态契约断言）、`shells`（H5 / 微信 / Expo / 桌面壳运行时门禁）、`agc-web`（AGC typecheck 与壳内测试）、`agc-rust`（共享 / 平台 crate 测试、AGC 串行壳测试、agent-run smoke）、`release`（AGC 与桌面壳发布构建 smoke、落盘产物断言）五组，每组暴露一个 `check:native-shells:<group>` 根脚本；不带 `--groups=` 时仍然串行跑全部分组，本地 `npm run check:native-shells` 语义不变。CI 据此把原客户端 job 拆成 `Native shell tests`（contract + shells + release）、`AI game creator shell web tests`（agc-web）、`AI game creator shell Rust tests`（agc-rust）三个 job，并把最长的 AGC Rust job 声明在最前，使 runner 领取顺序与关键路径一致。
-- 命令等价：`npm run ai-game-creator-shell:check` 拆成 `:check:web`（typecheck + 壳内测试）与 `:check:rust`（agent-runtime 两个独立 crate + `platform-llm` + `shared-contracts` + AGC 壳串行测试），聚合脚本仍是 `web && rust && agent-run:smoke` 同序同命令，本地与文档入口不变。`agent-run:smoke` 会用 `src-tauri/Cargo.toml` spawn `cargo`，因此归入 `agc-rust` 分组，与 AGC 依赖预热同 job。
-- 影响范围：`.gitea/workflows/project-ci.yml`（六个 job）、`scripts/check-native-shells.mjs`、根 `package.json` 门禁脚本、`scripts/project-ci-workflow.test.ts`（校验分组清单、根脚本内容与 job 覆盖，防止新增分组时静默漏跑）、开发运维文档与开发流程记忆。门禁覆盖不变，只有执行位置改变；Gitea `master` 分支保护的 required context 是追加式的（旧四个继续上报，需补上两个新 AGC context）。
-- 验证方式：`npx vitest run scripts/project-ci-workflow.test.ts`（11 条）；`node scripts/check-native-shells.mjs --groups=contract` 本地 0.6 秒通过；`--groups=` 未知组与空组都要报错关闭。实测耗时按拆分前同一 run 的日志时间戳折算：关键路径从 18 分 37 秒收敛到 AGC Rust job 的约 13 分钟量级（若 runner 并发槽位 ≥ 6，可压缩到约 10.5 分钟）。
-- 关联文档：[开发运维](../../【开发运维】本地开发验证与生产运维-2026-05-15.md)、[踩坑记录](pitfalls.md)。
-
 ## 2026-09-10 策划 Agent 迁移只复用生产基建
 
 - 决策：待实施的生产迁移以自由协作策划原型为行为基线，仅复用 Provider、恢复、文件操作、审计和 UI 通信；不继承旧 Planning V2 的强制工具、问询轮数、GDD 内容校验和版本审批。保留五阶段与顾问态、当前阶段资源注入和产物存在性检查，系统阶段空必需清单不增加解析或登记功能。
@@ -8646,10 +8637,3 @@ CI 上 `background_agent_runtime_recovers_stale_running_before_pending_task` 在
 - 决策：DirectProject app-server thread 改为 `sandbox="danger-full-access"`，turn 改为 `sandboxPolicy.type="dangerFullAccess"`，不再发送 `writableRoots` 或 workspace 网络开关，原生命令网络随完整 sandbox 开放；app-server 交互请求不再按 grant root 做白名单裁剪，直接项目会话统一接受文件变更、命令执行和权限请求。首页只读对话、AGC `agc_tools` 业务授权、Provider 凭据隔离、Runtime 审计和客户端受控文件工具合同继续保留。
 - 提示词同步：DirectProject 不再把路径范围描述成 Codex 原生能力禁区，但仍禁止主动输出 Token、Cookie、auth.json、`.env` 和 Runtime 私有控制面。
 - 验证：Rust 定向单测覆盖 `danger-full-access` / `dangerFullAccess`、无 `writableRoots`、外部 grant root 仍接受，以及 DirectHome 继续只读拒绝。
-
-## 2026-09-14 项目写锁的同进程复用收窄为同线程重入
-
-- 背景：`write_lock.rs` 的 advisory 复用判据曾放宽为「`.agent/project.lock` 的 `pid` 等于当前进程」，使本进程所有写通道都不再等待。`Project CI` 的 Rust 全量门禁因此出现 12 条失败：另一线程持锁时一致快照读 / `project.diff` / `action_history` / `command.output_read` / steer 不再等待，4 路并行直写撞项目 revision 侧车（`File exists (os error 17)`），8 线程并发 steer 拿到重复序号，`file.write` 锁失败脱敏与恢复安装的失败关闭变成成功。
-- 决策：复用判据收窄为**同一条写调用链（同一线程）重入**——按锁路径登记真实持锁线程，只有当前线程就是持锁线程时才返回 advisory guard；本进程其它线程的争用继续走有界等待与终态占用。自主游戏构建流水线的并行专家动作豁免保持不变；跨进程占用、残留回收、权限分类、等待预算和错误文案不变。
-- 边界：锁定这些不变量的既有用例（`project_tools` / `command_runtime` / `parallel_actions` / `runtime_state` / `response_stream` / `direct_tool_bridge` / `ui_editor::persistence`）不得为了让锁语义通过而改写；用「同线程自持锁」模拟「另一个写者」的两条用例改为**在另一条线程持锁**，断言语义不变。同进程跨线程重入（持锁链在 `await` / `spawn_blocking` 后于其它线程再取锁）仍会等满预算，出现现场时按 2026-08-27 的既有处置改用 `*_locked` 入口，不放宽判据。
-- 关联文档：[项目客户端占用锁收敛里程碑](../plans/【里程碑】项目客户端占用锁收敛-2026-09-14.md)、[踩坑记录](pitfalls.md)。

docs/project-memory/shared-memory/development-workflow.md

@@ -74,4 +74,4 @@ SpacetimeDB 任务统一先读取 `.codex/skills/genarrative-spacetimedb/SKILL.m
 
 ## Gitea CI 依赖闭合
 
-`.gitea/workflows/project-ci.yml` 的客户端门禁拆成三个 job，每个 job 只预热自己会构建的那几份依赖：`AI game creator shell Rust tests` 用 `cargo fetch --locked` 预取 `server-rs/Cargo.toml` 与 AGC 壳 manifest（`agent-run` smoke 会用 `src-tauri/Cargo.toml` spawn `cargo`，因此必须同 job），`Native shell tests` 预取桌面壳与 AGC 壳 manifest，`AI game creator shell web tests` 不触碰 Cargo，不预热。两个被 `server-rs/Cargo.toml` 排除、且没有提交 `Cargo.lock` 的独立 crate（`agent-runtime-core`、`agent-runtime-orchestration`）只能在 `AI game creator shell Rust tests` 里用不带锁标志的 fetch。Backend host workspace tests 使用 `cargo test --locked --workspace --exclude spacetime-module --no-fail-fast`，避免 `spacetime-module` 的 `spacetime-types` feature 统一污染普通领域 crate 的 host 测试；随后单独执行 `cargo test --locked -p spacetime-module --no-fail-fast`，由 `spacetime-module/src/active.rs` 在 host 测试构建期间提供仅测试期的 SpacetimeDB ABI 链接支持，使该 crate 的纯单元测试也纳入 Backend 门禁。`spacetime-module` 的 reducer / procedure 运行时行为仍必须通过真实 SpacetimeDB runtime/integration harness 验证，host 链接支持不得被当作运行时替身。Backend 另外执行 `cargo check --locked -p spacetime-module` 验证模块源码。AGC 壳检查还会运行 `platform-llm` 与 `shared-contracts` 的 server-rs workspace 测试，这些命令以及 AGC 壳测试必须带 `--locked`，避免在测试阶段重新解析 registry index；锁文件发生变化时应先更新受信任 CI 镜像缓存，再重跑门禁。
+`.gitea/workflows/project-ci.yml` 的 `Native shell tests` 在运行原生壳门禁前，必须使用 `cargo fetch --locked` 预取 `server-rs/Cargo.toml`、桌面壳和 AGC 壳三份依赖。Backend host workspace tests 使用 `cargo test --locked --workspace --exclude spacetime-module --no-fail-fast`，避免 `spacetime-module` 的 `spacetime-types` feature 统一污染普通领域 crate 的 host 测试；随后单独执行 `cargo test --locked -p spacetime-module --no-fail-fast`，由 `spacetime-module/src/active.rs` 在 host 测试构建期间提供仅测试期的 SpacetimeDB ABI 链接支持，使该 crate 的纯单元测试也纳入 Backend 门禁。`spacetime-module` 的 reducer / procedure 运行时行为仍必须通过真实 SpacetimeDB runtime/integration harness 验证，host 链接支持不得被当作运行时替身。Backend 另外执行 `cargo check --locked -p spacetime-module` 验证模块源码。AGC 壳检查还会运行 `platform-llm` 与 `shared-contracts` 的 server-rs workspace 测试，这些命令以及 AGC 壳测试必须带 `--locked`，避免在测试阶段重新解析 registry index；锁文件发生变化时应先更新受信任 CI 镜像缓存，再重跑门禁。