取消 DirectProject 成功回复脱敏

DirectProject 流式文本和最终回复只移除 thinking 包装，成功正文按用户项目内容原样展示。

删除路径、链接和凭据占位符替换以及回复长度截断逻辑。

同步删除最终回复必须脱敏的文档要求，并更新回归测试验证项目内容保持原样。

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

@@ -14,7 +14,6 @@ const DIRECT_AGC_ENGINEERING_GUIDANCE: &str = "AGC 工程合同（仅说明项
 const DIRECT_CODEX_ART_SPEC_ASSET_PATH: &str = "assets/art-spec.png";
 const DIRECT_CODEX_BACKGROUND_ASSET_PATH: &str = "assets/direct-game-background.png";
 const DIRECT_CODEX_SPRITESHEET_ASSET_PATH: &str = "assets/art-spritesheet.png";
-const MAX_DIRECT_VISIBLE_REPLY_CHARS: usize = 16 * 1024;
 const PLATFORM_GENERATION_SOURCE_PRESERVED_NO_RETRY_PREFIX: &str =
     "platform-generation-source-preserved-no-retry:";
 const DIRECT_TAONIER_LOCAL_RECONCILIATION_PREFIX: &str =
@@ -3661,24 +3660,15 @@ fn sync_direct_codex_project_outputs_at(
     sync_direct_codex_project_file_projection_at(root, previous_output_fingerprint)
 }
 
-/// Project Codex text into the only form that may cross the DirectProject UI
-/// boundary. The app-server stream can contain reasoning blocks, URLs,
-/// credentials, or host paths before the final reply is known; those values
-/// must never be emitted as an intermediate chat message or persisted as the
-/// user-visible assistant turn.
-fn project_direct_codex_visible_text(root: &Path, value: &str) -> Option<String> {
+/// Project Codex text for the user-visible DirectProject stream and reply.
+/// Reasoning wrappers are still removed because they are not reply text, but
+/// the user owns the project and the resulting reply is not redacted here.
+fn project_direct_codex_visible_text(value: &str) -> Option<String> {
     let stripped = strip_incomplete_direct_thinking_marker(&strip_llm_thinking_blocks(value));
     if stripped.trim().is_empty() {
         return None;
     }
-    let redacted = redact_agent_runtime_error(root, &stripped, MAX_DIRECT_VISIBLE_REPLY_CHARS)
-        .replace("<redacted-url>", "（链接已隐藏）")
-        .replace("$PROJECT_ROOT", "（项目路径已隐藏）")
-        .replace("<absolute-path>", "（路径已隐藏）")
-        .replace("[redacted-secret]", "（敏感信息已隐藏）")
-        .replace("[redacted-sensitive-field]", "（敏感字段已隐藏）")
-        .replace("[redacted sensitive context]", "（内部信息已隐藏）");
-    let visible = redacted.trim().to_string();
+    let visible = stripped.trim().to_string();
     (!visible.is_empty()).then_some(visible)
 }
 
@@ -3695,14 +3685,13 @@ fn strip_incomplete_direct_thinking_marker(value: &str) -> String {
 }
 
 fn project_direct_codex_accumulated_text(
-    root: &Path,
     stream_enabled: bool,
     accumulated_text: &str,
 ) -> Option<String> {
     if !stream_enabled {
         return None;
     }
-    project_direct_codex_visible_text(root, accumulated_text)
+    project_direct_codex_visible_text(accumulated_text)
 }
 
 pub(crate) fn build_direct_codex_system_prompt(root: &Path) -> Result<String, String> {
@@ -3915,7 +3904,7 @@ async fn run_direct_game_creator_turn_inner(
         let mut observer = move |observation: DirectCodexTurnObservation| match observation {
             DirectCodexTurnObservation::AccumulatedText(accumulated_text) => {
                 let visible_text =
-                    project_direct_codex_accumulated_text(root, stream_enabled, &accumulated_text);
+                    project_direct_codex_accumulated_text(stream_enabled, &accumulated_text);
                 if visible_text.is_none() {
                     return;
                 }
@@ -3926,7 +3915,7 @@ async fn run_direct_game_creator_turn_inner(
             }
             DirectCodexTurnObservation::IntermediateText(intermediate_text) => {
                 let visible_text = if stream_enabled {
-                    project_direct_codex_visible_text(root, &intermediate_text)
+                    project_direct_codex_visible_text(&intermediate_text)
                 } else {
                     None
                 };
@@ -3963,7 +3952,7 @@ async fn run_direct_game_creator_turn_inner(
         .await
     }
     .map_err(|error| DirectCodexTurnFailure::new(DirectCodexFailureStage::CodeGeneration, error))?;
-    let visible_reply = project_direct_codex_visible_text(root, &reply).ok_or_else(|| {
+    let visible_reply = project_direct_codex_visible_text(&reply).ok_or_else(|| {
         DirectCodexTurnFailure::new(
             DirectCodexFailureStage::CodeGeneration,
             "陶泥儿未返回可展示的回复".to_string(),
@@ -4662,58 +4651,48 @@ mod tests {
     }
 
     #[test]
-    fn direct_visible_stream_projection_hides_internal_content() {
-        let root = tempfile::tempdir().expect("direct stream root");
-        let project_file = root.path().join("game/index.html");
+    fn direct_visible_stream_projection_keeps_user_project_reply_content() {
+        let project_file = std::path::Path::new("game/index.html");
         let raw = format!(
             "先说一句\n<think>内部推理不应显示</think>\n来源 https://example.test/a\n路径 {}\nauthorization: Bearer secret-value-123",
             project_file.display()
         );
-        let visible =
-            project_direct_codex_visible_text(root.path(), &raw).expect("safe visible stream text");
+        let visible = project_direct_codex_visible_text(&raw).expect("visible stream text");
         assert!(visible.contains("先说一句"), "{visible}");
         assert!(!visible.contains("内部推理"), "{visible}");
-        assert!(!visible.contains("https://example.test"), "{visible}");
+        assert!(visible.contains("https://example.test"), "{visible}");
         assert!(
-            !visible.contains(project_file.to_string_lossy().as_ref()),
-            "{visible}"
-        );
-        assert!(!visible.contains("secret-value-123"), "{visible}");
-        assert!(visible.contains("链接已隐藏"), "{visible}");
-        assert!(
-            visible.contains("项目路径已隐藏") || visible.contains("路径已隐藏"),
+            visible.contains(project_file.to_string_lossy().as_ref()),
             "{visible}"
         );
+        assert!(visible.contains("secret-value-123"), "{visible}");
     }
 
     #[test]
     fn direct_visible_stream_projection_drops_unclosed_thinking_only_delta() {
-        let root = tempfile::tempdir().expect("direct stream root");
         assert_eq!(
-            project_direct_codex_visible_text(root.path(), "<think>secret reasoning"),
+            project_direct_codex_visible_text("<think>secret reasoning"),
             None
         );
     }
 
     #[test]
     fn direct_visible_stream_projection_hides_partial_thinking_tag() {
-        let root = tempfile::tempdir().expect("direct stream root");
         assert_eq!(
-            project_direct_codex_visible_text(root.path(), "已公开内容\n<thi"),
+            project_direct_codex_visible_text("已公开内容\n<thi"),
             Some("已公开内容".to_string())
         );
     }
 
     #[test]
     fn direct_accumulated_text_respects_the_explicit_stream_setting() {
-        let root = tempfile::tempdir().expect("direct stream root");
         assert_eq!(
-            project_direct_codex_accumulated_text(root.path(), false, "阶段性回复"),
+            project_direct_codex_accumulated_text(false, "阶段性回复"),
             None,
             "stream=false 只能保留阶段状态，不能向聊天窗口发增量文本"
         );
         assert_eq!(
-            project_direct_codex_accumulated_text(root.path(), true, "阶段性回复"),
+            project_direct_codex_accumulated_text(true, "阶段性回复"),
             Some("阶段性回复".to_string())
         );
     }

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

@@ -206,7 +206,7 @@ Supervisor 认领该回执后，由父 run 自己为每个原 delivery 逐一创
 - 安装包侧车：Windows x64 release 固定随 Tauri resource 打包 `@openai/codex@0.147.0` 的原生 `codex.exe`；Rust build script 从 AGC 子包锁定依赖 stage 到 resource，并写入版本与 SHA-256 清单。Windows 侧车映射只写入 `tauri.windows.conf.json`，通用 `tauri.conf.json` 不得让 Linux / macOS 构建依赖未生成的 Windows 二进制。运行时只在文件摘要和 `codex-cli` 版本同时匹配清单时优先选内置侧车；缺失、损坏或版本漂移时跳过它，按既有 npm 安装、PATH 顺序回退。安装包同时携带 Apache-2.0 第三方声明；API Key、`auth.json`、Cookie、Token、用户 `CODEX_HOME`、用户配置和项目数据绝不打包。
 - Windows x64 release 安装包只生成 NSIS，不生成 MSI：`tauri.windows.conf.json` 的 `bundle.targets` 固定为 `["nsis"]`，通用配置继续保留其它平台的默认打包目标。
 - CLI 安全边界：CLI 固定使用 argv 启动，禁止 shell 拼接；工作目录使用本次请求专用的空临时目录，不把游戏项目绝对路径写入 prompt、stdout、stderr 或持久记录。调用固定使用 ephemeral、忽略用户配置和 exec rules、read-only sandbox、never approval，并关闭 Codex shell tool；只继承 CLI 运行和认证所需的最小环境，显式移除宿主 `CODEX_API_KEY`。用户级 Codex 登录态继续由本机 Codex 自己读取，API Key、auth 文件、Cookie、Token、`CODEX_HOME` 私有内容不得复制到项目配置、Runtime sidecar、Agent DB、conversation 或日志；stdout / stderr 无换行时也受硬上限约束，stderr 诊断只记录固定分类、字节数和 SHA-256。
-- 协议边界：Runtime 把既有 `LlmRunRequest` 的消息和当前函数目录编码为有界 prompt，并从同一函数 JSON Schema 生成 Codex structured-output schema。CLI 输出转换为现有 `LlmRunResponse / LlmToolCall` 后，继续经过 native tool / MCP 参数校验、动作上限、权限、pending、receipt、验证与格式修复链；最终回复仍走现有脱敏和唯一提交路径，不新增平行响应协议。
+- 协议边界：Runtime 把既有 `LlmRunRequest` 的消息和当前函数目录编码为有界 prompt，并从同一函数 JSON Schema 生成 Codex structured-output schema。CLI 输出转换为现有 `LlmRunResponse / LlmToolCall` 后，继续经过 native tool / MCP 参数校验、动作上限、权限、pending、receipt、验证与格式修复链；最终回复仍走唯一提交路径，不新增平行响应协议。
 - 取消与恢复：Codex 子进程绑定当前 Provider request lifecycle，取消、暂停、Runner draining 或 GUI owner 丢失时终止并回收当前进程；started 后没有可信终态仍沿现有 Provider reconciliation 处理。`agentMode`、CLI 可执行身份和影响输出的 Codex 参数进入 `providerConfigFingerprint`，模式切换不得消费另一模式遗留的 retry/handoff。
 - 可用性与兼容：`check_game_creator_llm_config` 在 `codex_cli` 模式检查 Codex 可执行文件和版本，不再要求 HTTP API Key；在 `provider` 模式保持原有 API Key、base URL、model 和逐 Agent override 校验。设置面板始终保留旧 Provider 配置，模式切回后不得丢失既有密钥或逐 Agent设置。配置向导若写入 HTTP Provider 凭据，必须显式把模式切到 `provider`。
 - 验收：覆盖缺省模式、显式 Provider 回退、配置 round-trip、CLI 缺失/非零退出/损坏 JSONL/超限输出、structured tool calls、普通最终回复、进程取消、模式指纹漂移和旧 Provider 回归；同时执行 AGC 定向 Rust 测试、前端配置测试与 typecheck、配置门禁、编码检查和 `git diff --check`。