lark-remote 0.5.12 → 0.5.13

package/CHANGELOG.md

@@ -5,6 +5,17 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.5.13] - 2026-10-26
+
+### 修复
+
+- **长回合的最终回答不再被进度内容挤掉**：工具输出、思考、通知此前与最终回答共用同一份卡片预算，回合越长越容易把回答截断甚至整段挤没。现在最终回答独立保存，不受进度文本窗口和工具块上限淘汰；终态渲染先排状态、审批与操作控件，再放完整回答，剩余空间才从新到旧回填进度内容。协议带 `final_answer` 标记的（Codex app-server）直接采用，没有标记的 agent 在收尾时把最后一条非空 assistant 文本提升为最终回答，Claude、opencode、Kimi、Pi、DSH 共用同一套保护
+- **回答放不下时随后补发完整文本**：单张卡片连「最终回答 + 必要控件」都装不下时，原卡提示「最终回答较长，完整内容已随后发送」，完整回答按 UTF-8 字节边界拆成多条纯文本消息补发（优先落在换行处，单条上限 20000 字节），拼接后与原文逐字一致，不再静默截断；补发失败会在原卡给出「完整回答补发失败」的可见回执
+
+### 维护
+
+- **入口与路由/桥接层拆分为独立模块**：消息入站装配、命令实现（`ls` / `order` / `resume` / `ws`）、分页与路径校验、审批转发、用量聚合、会话回写、DSH 多路复用流各自成模块，`src/index.ts`、`src/bridge/index.ts`、`src/router/index.ts` 只保留编排职责；同时为拆分中暴露的状态竞争（中断与自然结束并发写同一张卡的终态、审批重投递、bash 卡片收尾）补上回归用例。无行为变化
+
 ## [0.5.12] - 2026-10-24
 
 ### 修复
@@ -562,8 +573,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### 修复
 
 - `displayTitle`/`summary` 跳过 task-notification 注入
-- 修复 `scanClaudeSessions` 对迁移到其他目录的会话的遗漏（S2）
-- 修复 probe 生命周期、工厂回退、卡片回调等评审发现的问题（F1-F8）
+- 修复 `scanClaudeSessions` 对迁移到其他目录的会话的遗漏
+- 修复 probe 生命周期、工厂回退、卡片回调等问题
 
 ### 变更
 

package/README.en.md

@@ -14,19 +14,6 @@ Designed for a single-user, peer-to-peer private chat. The agent's system prompt
 
 > **New here? Start with [docs/en/getting-started.md](docs/en/getting-started.md).**
 
-## ⚠️ Security Warning (read first)
-
-This tool turns Feishu messages directly into local agent executions. **Out of the box, agents run without approval prompts**: Claude defaults to `bypassPermissions` (no approval card), but that permission mode is **configurable** — set it to anything other than `bypassPermissions` (`default` / `acceptEdits` / `auto` / `dontAsk` / `plan`) to enable interactive approval, where high-risk actions wait for your confirmation on the Feishu card. Codex uses the **app-server approval mode** (default `approvalPolicy=on-request` — command execution requires your confirmation on the Feishu card, sandbox defaults to `workspace-write`); Kimi defaults to `manual` (per-action approval; note that since kimi 0.40, `yolo` means Ask-When-Needed rather than allow-everything — use `auto` for fully unattended, no-approval execution); other agents run with their CLI defaults. **In any mode**, agents can read and write the local directories you specify. You MUST:
-
-- **Use it only in your own private (p2p) chat.** Never add the bot to any group chat; never let anyone else talk to it.
-- Restrict the app's visibility to **yourself only** in the Feishu Open Platform.
-- Don't run it on a machine holding data you can't afford to lose; prefer a dedicated user / machine / container.
-- Keep credentials such as `appSecret` in your local `config.yaml` only — **never commit them anywhere**.
-
-In approval mode (Claude, Codex, Kimi), an approval card arrives before a command runs — approve or reject it right in Feishu:
-
-![Command approval card example](docs/images/approval-card.png)
-
 ## Prerequisites
 
 - OS: macOS / Linux / Windows 10+ (native Windows support is verified on real hardware, with behavior matching macOS/Linux)
@@ -70,7 +57,7 @@ defaultAgent: claude
 claude:
   model: claude-opus-4-8
   effort: medium            # low | medium | high | xhigh | max
-  permissionMode: bypassPermissions  # Claude official --permission-mode: default | acceptEdits | auto | bypassPermissions | manual | dontAsk | plan (switchable via /config card; anything other than bypassPermissions enables approval)
+  permissionMode: auto  # Claude official --permission-mode: default | acceptEdits | auto | bypassPermissions | manual | dontAsk | plan (switchable via /config card; auto is the default: low-risk actions run, high-risk actions ask)
   stopGraceMs: 5000
 
 logging:

package/README.md

@@ -14,19 +14,6 @@
 
 > **第一次接触本项目？先看 [docs/zh/getting-started.md](docs/zh/getting-started.md)。**
 
-## ⚠️ 安全警告（必读）
-
-本工具会把飞书消息直接转为本地 agent 执行。**开箱默认是不弹审批的**：Claude 默认以 `bypassPermissions` 运行（无审批卡），但该权限模式**可配置**——把它改成 `bypassPermissions` 以外的值（`default` / `acceptEdits` / `auto` / `dontAsk` / `plan`）即启用交互式审批，高风险操作会在飞书卡片上等你确认；Codex 采用 **app-server 审批模式**（默认 `approvalPolicy=on-request`，执行命令需在飞书卡片上确认，沙箱默认 `workspace-write`）；Kimi 默认 `manual`（逐条审批；注意 kimi 0.40+ 起 `yolo` 已改为「必要时询问」而非全部放行，要完全不弹审批需选 `auto`）；其他 agent 以各自 CLI 的默认权限运行。**无论哪种模式**，agent 都能读写你指定的本地目录，请务必：
-
-- **仅限你自己的私聊（p2p）使用**。不要把机器人拉进任何群聊，不要让任何其他人能与它对话。
-- 在飞书开放平台把应用可见范围限制到**只有你自己**。
-- 运行本工具的机器上不要存放与损失承受力不符的数据；建议在独立用户/机器/容器里运行。
-- `appSecret` 等凭据只在本地 `config.yaml`，**不要提交到任何仓库**。
-
-审批模式下（Claude、Codex、Kimi），执行命令前会收到审批卡片，可直接在飞书里允许或拒绝：
-
-![命令审批卡片示例](docs/images/approval-card.png)
-
 ## 前置条件
 
 - 操作系统：macOS / Linux / Windows 10+（原生 Windows 支持已完成真机验证，行为与 macOS/Linux 对齐）
@@ -70,7 +57,7 @@ defaultAgent: claude
 claude:
   model: claude-opus-4-8
   effort: medium            # low | medium | high | xhigh | max
-  permissionMode: bypassPermissions  # Claude 官方 --permission-mode：default | acceptEdits | auto | bypassPermissions | manual | dontAsk | plan（`/config` 卡片可切换；非 bypassPermissions 会启用审批）
+  permissionMode: auto  # Claude 官方 --permission-mode：default | acceptEdits | auto | bypassPermissions | manual | dontAsk | plan（`/config` 卡片可切换；auto 为默认：低风险放行、高风险审批）
   stopGraceMs: 5000
 
 logging:

package/dist/bridge/approval-coordinator.d.ts

@@ -179,6 +179,7 @@ export declare class ApprovalCoordinator {
     /** 记录问题答案到 tracked.answers（懒初始化 Map）。 */
     private recordAnswer;
     private getPendingQuestion;
+    private getPendingQuestionAt;
     private allQuestionsAnswered;
     /** 推送审批卡视图更新（approval_view_updated 重渲染；失败仅记日志）。 */
     private pushApprovalViewUpdate;

package/dist/bridge/approval-coordinator.js

@@ -178,12 +178,7 @@ export class ApprovalCoordinator {
      *   submitAnswers() 显式提交。
      */
     async toggleAnswer(action, ctx) {
-        const tracked = this.getPendingQuestion(ctx.requestId);
-        const questions = tracked.view.questions;
-        const q = questions[action.questionIndex];
-        if (!q) {
-            throw new Error(`Question ${action.questionIndex} not found in request ${ctx.requestId}`);
-        }
+        const { tracked, question: q } = this.getPendingQuestionAt(ctx.requestId, action.questionIndex);
         const option = q.options.find((o) => o.label === action.option);
         if (!option) {
             throw new Error(`Option "${action.option}" not found in question ${action.questionIndex}`);
@@ -217,12 +212,7 @@ export class ApprovalCoordinator {
      * 提交该问题当前勾选的选项；全部问题答完时通过 responder 提交。
      */
     async submitAnswers(action, ctx) {
-        const tracked = this.getPendingQuestion(ctx.requestId);
-        const questions = tracked.view.questions;
-        const q = questions[action.questionIndex];
-        if (!q) {
-            throw new Error(`Question ${action.questionIndex} not found in request ${ctx.requestId}`);
-        }
+        const { tracked, question: q } = this.getPendingQuestionAt(ctx.requestId, action.questionIndex);
         if (!q.multiSelect) {
             throw new Error('该问题为单选，请直接点击选项');
         }
@@ -240,12 +230,7 @@ export class ApprovalCoordinator {
      * 单选问题的答案。与单选 toggle 同语义：全部问题答完自动提交。
      */
     async answerCustom(action, ctx) {
-        const tracked = this.getPendingQuestion(ctx.requestId);
-        const questions = tracked.view.questions;
-        const q = questions[action.questionIndex];
-        if (!q) {
-            throw new Error(`Question ${action.questionIndex} not found in request ${ctx.requestId}`);
-        }
+        const { tracked, question: q } = this.getPendingQuestionAt(ctx.requestId, action.questionIndex);
         if (q.multiSelect) {
             throw new Error('自定义答案仅支持单选');
         }
@@ -266,12 +251,7 @@ export class ApprovalCoordinator {
      * 后回流视图（卡片回显），同一请求可多次修改（新 nonce 来自新渲染）。
      */
     async answerNote(action, ctx) {
-        const tracked = this.getPendingQuestion(ctx.requestId);
-        const questions = tracked.view.questions;
-        const q = questions[action.questionIndex];
-        if (!q) {
-            throw new Error(`Question ${action.questionIndex} not found in request ${ctx.requestId}`);
-        }
+        const { tracked, question: q } = this.getPendingQuestionAt(ctx.requestId, action.questionIndex);
         const text = action.text.trim();
         if (!text) {
             throw new Error('请输入补充说明');
@@ -443,6 +423,14 @@ export class ApprovalCoordinator {
         }
         return tracked;
     }
+    getPendingQuestionAt(requestId, questionIndex) {
+        const tracked = this.getPendingQuestion(requestId);
+        const question = tracked.view.questions?.[questionIndex];
+        if (!question) {
+            throw new Error(`Question ${questionIndex} not found in request ${requestId}`);
+        }
+        return { tracked, question };
+    }
     allQuestionsAnswered(tracked) {
         const questions = tracked.view.questions ?? [];
         if (questions.length === 0)

package/dist/bridge/index.d.ts

@@ -369,6 +369,12 @@ export declare class Bridge {
     private interruptActiveRun;
     /** Send a result message to the user (§7). Returns true on success, false on failure. */
     sendResult(result: CommandResult, ctx: CommandContext): Promise<boolean>;
+    /**
+     * Deliver a final answer that cannot fit in the run card. Plain text avoids
+     * markdown parser limits and preserves every character; sequence labels make
+     * multi-message delivery unambiguous.
+     */
+    private sendFinalAnswerOverflow;
     /**
      * Update an existing card in place (原地更新) instead of sending a new one.
      *
@@ -567,6 +573,7 @@ export declare class Bridge {
         text: string;
         nonce: string;
     }): Promise<void>;
+    private dispatchQuestionAction;
     /**
      * 记下一次「可压缩 run」：终态卡上的 Compact 按钮点击时据此反查
      * sessionId + agentKind。runId 是按钮自带的唯一键，同 cwd 内唯一。