@unifan/pi-unifan-zh 1.0.16 → 1.0.17

package/README.md

@@ -85,20 +85,26 @@ pi install npm:@unifan/pi-commit-zh
 
 ---
 
-### 2. 🔍 `review` (AI Code Review Suite)
-- **Review Modes**:
-  - **`/review-lite`** ⭐ **(Recommended for daily development)**: Fast single-agent review pass in 2~3s, highly token-efficient.
-  - **`/review-perf`** 🔥 **(Performance & Benchmark Reviewer)**: Deep-dive inspection for GC allocation, hot loops, algorithmic complexity, and benchmark probes.
-  - **`/review`** 🔍 **(Daily 3-Expert Review)**: Bugbot + Security + Compliance + Gatekeeper with low latency.
-  - **`/review-full`** 🏥 **(Full 6-Expert Consultation)**: Full multi-agent parallel audit (ideal for major releases and PR merges).
-- **Utility Commands**:
-  - **`/review-show`**: Re-render the most recent review report.
-  - **`/review-agents`**: View reviewer agents and models status.
-  - **`/review-config`**: Edit review configuration.
-- **Key Features**:
-  - **Model Inheritance**: Gate automatically inherits the active session model (`inherit`), avoiding missing model errors.
-  - **Fault Tolerance & Auto-Retry**: Automatically retries on transient stream drops and preserves all completed reviewer findings even if the gate fails.
-  - **100% Fluent Chinese Output**.
+### 2. 🔍 `review` (Interactive AI Code Review Suite)
+- **Architecture**:
+  - Re-architected following the battle-tested **Codex & pi-agent-extensions** design (by Armin Ronacher @mitsuhiko).
+  - **Native Session Tree Branch Isolation**: Completely eliminates multi-subagent WebSocket dropouts, proxy streaming errors, and timeouts. Rock-solid stability on any model (GPT, Claude, DeepSeek, GLM, etc.).
+- **Commands**:
+  - **`/review`** 🔍: Launches interactive Chinese TUI menu with 6 review modes:
+    - ① `Review uncommitted changes` (staged + unstaged working tree, smart default)
+    - ② `Review against base branch` (auto merge-base calculation against main/master/dev)
+    - ③ `Review specific commit` (search and pick from recent commits)
+    - ④ `Review GitHub Pull Request` (auto local checkout via gh CLI)
+    - ⑤ `Review folder/files snapshot` (full snapshot audit, non-diff)
+    - ⑥ `Custom review focus` (concurrency, GC, security, API contracts)
+  - **`/review-lite`** ⭐: Directly audits current uncommitted changes without menu prompts.
+  - **`/end-review`** 🏁:
+    - Automatically structures review findings into P0~P3 action items.
+    - Seamlessly jumps back to the original working branch and pre-fills the editor with fix instructions!
+- **Standards & Guidelines**:
+  - **P0~P3 Strict Severity Tags**: Blocker, Urgent, Normal, Low.
+  - **Zero Speculation & Grounded in Diff**: Every finding is backed by source evidence.
+  - **Custom Guidelines**: Automatically respects repo-level `REVIEW_GUIDELINES.md`, `AGENTS.md`, or `CLAUDE.md`.
 
 ---
 

package/README_zh.md

@@ -84,20 +84,26 @@ pi install npm:@unifan/pi-commit-zh
 
 ---
 
-### 2. 🔍 `review`（AI 并发代码审查套件·中文版）
-- **四大审查命令模式**：
-  - **`/review-lite`** ⭐ **（极速单兵审查，强烈推荐日常使用）**：单 AI 快速体检，2~3 秒出结果，极度省时省 Token。
-  - **`/review-perf`** 🔥 **（专项性能审查与基准测试）**：专项深度排查 GC 堆内存分配、Update/帧循环热点、算法复杂度与 Benchmark 探针。
-  - **`/review`** 🔍 **（日常 3 大核心专家审查）**：Bug 猎手 + 安全审查 + 规范合规 + 门禁裁判长（低延迟抗超时）。
-  - **`/review-full`** 🏥 **（全量 6 专家深度会诊）**：全套专家并发会诊（适合重大 PR 与版本合并，支持 `--full`）。
-- **辅助命令**：
-  - **`/review-show`**：重新查看最近一次的代码审查报告。
-  - **`/review-agents`**：查看各专家代理状态与模型分配。
-  - **`/review-config`**：编辑审查配置文件。
-- **核心亮点**：
-  - **门禁自适应继承**：门禁裁判长（Gate）自适应继承当前主会话模型，绝不报错 `Unknown subagent model`。
-  - **自动网络重试与熔断保护**：并发网络抖动自动重试，即使门禁异常也能完整保留前面所有专家的审查成果，绝不浪费 Token！
-  - **报告 100% 纯正中文输出**。
+### 2. 🔍 `review`（AI 交互式代码审查套件·中文版）
+- **架构升级**：
+  - 参考并融合了 **Codex 与 pi-agent-extensions**（Armin Ronacher @mitsuhiko）的经典架构。
+  - **原生会话分支隔离技术**：彻底告别脆弱的 6 子代理并发通信、网络断流与超时风险，支持任意模型（GPT/Claude/DeepSeek/免费模型）。
+- **核心审查命令**：
+  - **`/review`** 🔍 **（交互式全功能审查）**：唤起全中文 TUI 选择弹窗，支持 6 种审查模式：
+    - ① `审查未提交改动` (工作区已暂存 + 未暂存代码，智能默认推荐)
+    - ② `与基准分支对比审查` (如与 main / master / dev 自动计算 merge-base 对比)
+    - ③ `审查指定提交` (从最近提交历史中搜索挑选)
+    - ④ `审查 Pull Request` (输入 PR 编号或 GitHub URL 本地自动检出审查)
+    - ⑤ `审查指定目录/文件` (全量静态快照审查，非 diff)
+    - ⑥ `自定义审查要求` (针对并发安全、GC/性能、指定接口等重点排查)
+  - **`/review-lite`** ⭐ **（极速单兵审查）**：无需弹窗，直接对当前工作区未提交改动做闪电审查。
+  - **`/end-review`** 🏁 **（结束审查并返回）**：
+    - 自动将审查分支中发现的 `[P0~P3]` 缺陷与待办事项生成结构化中文报告。
+    - 一键平滑跳回主开发会话分支，并自动在输入框回填 `根据上述代码审查发现进行修改修复`！
+- **审查准则与标准**：
+  - **P0~P3 严格分级**：致命阻塞、紧急待修、普通建议、细节优化。
+  - **严格过滤误报**：基于代码事实与真实 diff，杜绝主观臆测，重点排查并发竞态、内存开销、外部未受信输入与边界异常。
+  - **支持项目定制规范**：自动识别并加载项目根目录的 `REVIEW_GUIDELINES.md`、`AGENTS.md` 或 `CLAUDE.md`。
 
 ---