feat(session): 新增 replyDelivery=transcript 转写回复模式，并桥接 Claude Code statusline 配额

docs-site/docs/en/bots-json.md

@@ -240,6 +240,31 @@ This option addresses one narrow gap: Codex running through Botmux's app-server
 |-------|-------------|
 | `senderTag` | Boolean, default `true` (on). Whether each turn forwarded to the CLI carries a `<sender type="user\|bot" open_id="ou_…" name="…" email="…" />` tag naming who spoke. Only an explicit `false` is persisted and disables it; absent or `true` both keep injecting, leaving the prompt byte-for-byte identical to historical behavior |
 | `thinkingCardToolResult` | Boolean, default `true` (on). Whether tool nodes in the native thinking bubble (bot-level master switch `thinkingCard`, default on) carry the command output / file content code block. `false` keeps only thinking paragraphs and tool node titles (tool · command / path) and degrades the result to a single `✓ Done` line (a tool node only leaves the “running” state once a result event arrives, so the event cannot simply be dropped), matching Claude Code's own UI; toggle via `/botconfig set thinkingCardToolResult off` or the dashboard card sub-switch, effective immediately |
+| `replyDelivery` | `"transcript"` or `"send"`; the default depends on the CLI: `claude-code` defaults to `transcript`, every other CLI to `send`. How the final reply reaches Feishu: `transcript` = the daemon takes the last assistant text of the turn from the CLI transcript and posts it as the final reply card, and the system prompt no longer mentions `botmux send`; `send` = the model must run `botmux send` itself (historical behavior). An explicit `"send"` is the only way to put claude-code back on the old behavior; both `send` and `transcript` are persisted, `unset` returns to the CLI default |
+
+### `replyDelivery: "transcript"`
+
+`claude-code` defaults to `transcript`; the other supported CLIs need it set explicitly. Once active it changes three things for that bot's sessions:
+
+1. **The system prompt never mentions `botmux send`**: the intro becomes "your final assistant message is automatically forwarded back to Lark by botmux — just answer directly"; the heredoc rule, the @ decision gate, the attachment usage and the `<identity>` rule "collaboration requires `botmux send --mention`" are all dropped, leaving only `botmux history` / `botmux bots list` and the `BOTMUX_NOTHING_TO_SEND` silence sentinel. For the cases that genuinely need `botmux send` (attachments, cross-bot @) the model can discover the built-in skill (`botmux-send` under `--plugin-dir`) on its own;
+2. **The per-turn `<botmux_reminder>` is no longer injected** (one less reminder block per prompt);
+3. **Solo sessions are unwrapped**: in a DM, or a plain 1:1 group whose only participants are the owner and this bot, each turn drops the `<user_message>` wrapper and the `<sender/>` tag, so the model sees bare text. Topic groups, multi-member groups, and turns spoken by anyone other than the owner never count as solo; wrapper and tag stay as before.
+
+Supported CLIs: `claude-code`, plus the structured-transcript bridge CLIs `codex` / `traex` / `coco` / `hermes` / `mtr` / `pi` / `oh-my-pi` / `ebsd` / `grok`. Other CLIs (e.g. `cursor`, `gemini`) have no transcript capture, so both `/botconfig set` and the dashboard reject the value (`reply_delivery_unsupported`); if the field is already persisted and `cli` is later switched to an unsupported CLI, the runtime falls back to `send` (one warn in the log) rather than losing replies.
+
+Hot-updatable by the owner / `allowedUsers` via `/botconfig`:
+
+```text
+/botconfig set replyDelivery transcript   # enable explicitly on the other supported CLIs
+/botconfig set replyDelivery send         # put claude-code back on the old behavior (model runs botmux send itself)
+/botconfig unset replyDelivery            # back to the CLI default
+```
+
+- **Two activation points**: the per-turn envelope (reminder / wrapper / `<sender/>`) applies from the next turn; the system prompt is injected at spawn time, so a running session needs `/restart` to pick up the new value, while new sessions use it directly.
+- **Observability cost**: the bare-text shape of a solo session has no `<user_message>` / `<sender>` structure, so `/adopt` no longer recognizes such sessions as botmux's own (the same class of cost as `senderTag: false`).
+- The dashboard "Reply Delivery → Transcript reply mode" toggle saves this field; it is disabled with an explanation when the current CLI does not support it.
+
+### `senderTag: false`
 
 With it off the model cannot see speaker identity: in a multi-person chat it cannot tell participants apart or address them by name. Useful for a CLI whose model copies the tag into its reply body (e.g. cursor — see the `<sender_note>` anti-echo hint, which disappears together with the tag), or when you do not want per-message identity written into the CLI transcript.
 

docs-site/docs/zh/bots-json.md

@@ -243,6 +243,31 @@ Dashboard 的「Bot 配置 → 消息卡片 → 实时卡片按钮」提供同
 |------|------|
 | `senderTag` | 布尔，默认 `true`（开）。每轮转发给 CLI 的消息是否附带一个 `<sender type="user\|bot" open_id="ou_…" name="…" email="…" />` 标签，告诉模型这句话是谁说的。只有显式 `false` 会写盘并关闭；缺省或 `true` 都保持注入，prompt 与历史行为逐字节一致 |
 | `thinkingCardToolResult` | 布尔，默认 `true`（开）。思考气泡（bot 级总开关 `thinkingCard`，默认开）的工具节点是否附带命令输出 / 文件内容代码块。设为 `false` 后气泡只保留思考段落与工具节点标题（工具名 · 命令 / 路径），结果退化成一行 `✓ 已完成`（工具节点在飞书端要收到结果事件才会从「执行中」落定，所以不能干脆不发），与 Claude Code 自身界面一致；`/botconfig set thinkingCardToolResult off` 或 dashboard「卡片」子开关切换，立即生效 |
+| `replyDelivery` | `"transcript"` 或 `"send"`，缺省按 CLI：`claude-code` 默认 `transcript`，其它 CLI 默认 `send`。最终回复怎么送到飞书：`transcript` = daemon 从 CLI 转写自动取本轮最后的 assistant 文本发最终回复卡，系统提示不再提及 `botmux send`；`send` = 模型必须自己 `botmux send`（历史行为）。显式写 `"send"` 才让 claude-code 退回旧行为；`send` / `transcript` 都会写盘，`unset` 回各 CLI 默认 |
+
+### `replyDelivery: "transcript"`
+
+`claude-code` 缺省即 `transcript`；其它支持的 CLI 需显式设置。生效后对该 bot 的会话有三条效果：
+
+1. **系统提示不再提及 `botmux send`**：开场改为「最终 assistant message 由 botmux 自动转发回飞书，直接作答即可」，heredoc / @ 决策 / 附件用法、`<identity>` 里「协作必须 `botmux send --mention`」的规则一并去掉，只保留 `botmux history` / `botmux bots list` 与 `BOTMUX_NOTHING_TO_SEND` 沉默哨兵。附件、跨 bot @ 等确实需要 `botmux send` 的场景，模型可通过内置 skill（`--plugin-dir` 里的 `botmux-send`）按需自行发现；
+2. **不再逐轮注入 `<botmux_reminder>`**（每轮 prompt 少一段提醒）；
+3. **solo 会话去壳**：私聊、或只有 owner 和本 bot 两个参与者的 1v1 普通群，每轮消息去掉 `<user_message>` 壳与 `<sender/>` 标签，模型看到的就是裸文本。话题群、多人群、非 owner 发言的一律不算 solo，壳与标签照旧。
+
+支持的 CLI 白名单：`claude-code`，以及走结构化转写桥的 `codex` / `traex` / `coco` / `hermes` / `mtr` / `pi` / `oh-my-pi` / `ebsd` / `grok`。其它 CLI（如 `cursor`、`gemini`）没有转写采集通道，`/botconfig set` 与 dashboard 都会拒绝（`reply_delivery_unsupported`）；已写盘后再把 `cli` 切到不支持的 CLI，运行时自动回落 `send`（日志 warn 一次），不会丢回复。
+
+可由 owner / `allowedUsers` 通过 `/botconfig` 热更新：
+
+```text
+/botconfig set replyDelivery transcript   # 其它支持的 CLI 显式开启
+/botconfig set replyDelivery send         # claude-code 退回旧行为（模型自己 botmux send）
+/botconfig unset replyDelivery            # 回各 CLI 默认
+```
+
+- **生效时机分两段**：逐轮信封（reminder / 壳 / `<sender/>`）从下一轮起生效；系统提示是 spawn 时注入的，已在跑的会话要 `/restart` 才换新值，新会话直接用新值。
+- **观测代价**：solo 会话的裸文本形态没有 `<user_message>` / `<sender>` 结构，`/adopt` 不再把这类会话识别为 botmux 自产会话（与 `senderTag: false` 同类代价）。
+- dashboard「回复投递 → 转写回复模式」开关保存的就是这个字段；当前 CLI 不支持时开关禁用并说明。
+
+### `senderTag: false`
 
 关掉后模型看不到发言人身份：多人会话里无法区分谁说的、也无法按人称呼。适合模型会把标签内容抄进回复正文的 CLI（如 cursor，见 `<sender_note>` 反抄写提示——标签关掉后该提示也一并消失），或不希望把每条消息的身份写进 CLI 记录的场景。
 

src/adapters/cli/claude-code.ts

@@ -20,7 +20,7 @@ import {
 import { homedir } from 'node:os';
 import { basename, dirname, isAbsolute, join, relative, sep } from 'node:path';
 import { resolveCommand } from './registry.js';
-import { sessionReadyHookCommand, userPromptHookCommand } from '../hook-command.js';
+import { sessionReadyHookCommand, statuslineHookCommand, userPromptHookCommand } from '../hook-command.js';
 import type { CliAdapter, CliId, PtyHandle } from './types.js';
 import { findJsonlContainingFingerprint, jsonlContainsFingerprint, normaliseForFingerprint } from '../../services/claude-transcript.js';
 import { CLAUDE_REASONING_EFFORTS } from '../../services/codex-reasoning-effort.js';
@@ -715,6 +715,56 @@ function resolveClaudeChatKeybindings(keybindingsPath: string): ClaudeChatKeybin
  *  across multiple adapter instances shares the warmup state. */
 const claudeFirstWriteSeen = new WeakSet<PtyHandle>();
 
+/** 用户自己配置的 statusLine（被 botmux 进程级 --settings 遮蔽的那一条）。 */
+export interface ShadowedStatusLine {
+  command?: string;
+  padding?: number;
+  refreshInterval?: number;
+}
+
+/**
+ * 找回被 botmux 进程级 `--settings` 遮蔽的用户 statusLine。
+ *
+ * 背景：Claude 的 settings 里 `statusLine` 是**单值**（不像 hooks 按事件合并数组），
+ * 而 --settings 优先级最高，所以 botmux 一注入，用户在项目 / 用户 settings 里配的
+ * statusline 命令就再也不会被 Claude 调用。为了不吞掉它，worker 在 spawn 前按 Claude
+ * 自己的优先级找到那条命令，经 `BOTMUX_STATUSLINE_CHAIN` 交给 `botmux statusline`：
+ * 落盘之后把**原始 stdin 字节**转发给它并透传其 stdout / 退出码——对用户的终端来说
+ * 状态栏行为不变。
+ *
+ * 优先级（高 → 低，取第一个 `type === 'command'` 且 command 非空的）：
+ *   `<cwd>/.claude/settings.local.json` > `<cwd>/.claude/settings.json` > `userSettingsPath`
+ * （后者通常是 `~/.claude/settings.json`；read-isolation 下是 `<BOT_HOME>/claude/settings.json`）。
+ * 不看 managed / enterprise 策略层：那一层 botmux 本来就无权覆盖，Claude 会自行处理。
+ *
+ * 纯函数、fail-open：任何读 / parse 失败视为该层无配置，继续向下找；全部没有 ⇒ `{}`。
+ * 不做全局 settings 兜底写入——全局只能有一个 statusLine，写进去就覆盖用户自己的。
+ * wrapperCli=aiden 会把 --settings 整个剥掉，此时 Claude 直接用用户自己的 statusLine，
+ * `botmux statusline` 不会被调用，worker 照常算出的 BOTMUX_STATUSLINE_CHAIN 只是闲置无害
+ * （cjadk / ccr / ttadk 会透传 --settings，沙盒开启时 wrapperCli 又被整体忽略，都需要链）。
+ */
+export function resolveShadowedStatusLine(opts: { workingDir: string; userSettingsPath?: string }): ShadowedStatusLine {
+  const candidates = [
+    join(opts.workingDir, '.claude', 'settings.local.json'),
+    join(opts.workingDir, '.claude', 'settings.json'),
+    ...(opts.userSettingsPath ? [opts.userSettingsPath] : []),
+  ];
+  for (const path of candidates) {
+    let parsed: unknown;
+    try { parsed = JSON.parse(readFileSync(path, 'utf-8')); } catch { continue; }
+    if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) continue;
+    const sl = (parsed as Record<string, unknown>).statusLine;
+    if (!sl || typeof sl !== 'object' || Array.isArray(sl)) continue;
+    const o = sl as Record<string, unknown>;
+    if (o.type !== 'command' || typeof o.command !== 'string' || o.command.trim() === '') continue;
+    const out: ShadowedStatusLine = { command: o.command };
+    if (typeof o.padding === 'number' && Number.isFinite(o.padding)) out.padding = o.padding;
+    if (typeof o.refreshInterval === 'number' && Number.isFinite(o.refreshInterval)) out.refreshInterval = o.refreshInterval;
+    return out;
+  }
+  return {};
+}
+
 /** A member of the Claude-family CLIs: Claude Code itself and forks that share
  *  its on-disk session layout (per-project JSONL transcripts, `sessions/<pid>.json`
  *  pid-state, `tasks/` fd locks, keybindings.json, settings.json hooks) but
@@ -852,7 +902,7 @@ export function createClaudeFamilyAdapter(variant: ClaudeFamilyVariant, rawBin:
       return discoverClaudeFamilySessions(variant.dataDir, limit, exclude);
     },
 
-    buildArgs({ sessionId, resume, resumeSessionId, forkSession, botName, botOpenId, locale, model, reasoningEffort, disableCliBypass, skillPluginDir, noTransport, triggerUserAuth }) {
+    buildArgs({ sessionId, resume, resumeSessionId, forkSession, botName, botOpenId, locale, model, reasoningEffort, disableCliBypass, skillPluginDir, noTransport, triggerUserAuth, replyDelivery, solo }) {
       const args: string[] = [];
       if (resume) {
         args.push('--resume', resumeSessionId ?? sessionId);
@@ -908,7 +958,22 @@ export function createClaudeFamilyAdapter(variant: ClaudeFamilyVariant, rawBin:
         inlineSettings.skipDangerousModePermissionPrompt = true;
         inlineSettings.permissions = { defaultMode: 'bypassPermissions' };
       }
-      // 仅在有内容（bypass 键）时才传 --settings；disableCliBypass 下没东西可传就不传。
+      // statusLine（仅 claude-code）：Claude 把 context_window / rate_limits 等 JSON 喂给
+      // 这条命令的 stdin，`botmux statusline` 落盘到 `<DATA_DIR>/statusline/<sid>/`，
+      // 卡片用量段据此渲染 `ctx 23% · 5h 18% · 7d 5%`。它**必须**走进程级 --settings 而
+      // 不能像就绪 hook 那样写全局：settings 里 statusLine 只能有一个（不是 hooks 那样按
+      // 事件合并的数组），写全局会覆盖用户自己的 statusline。进程级这份优先级最高，会
+      // **遮蔽**用户在项目 / 用户 settings 里的 statusLine——worker 用
+      // resolveShadowedStatusLine 找回它并经 BOTMUX_STATUSLINE_CHAIN 交给 `botmux
+      // statusline` 转发，用户终端里的状态栏不受影响。wrapperCli=aiden 会剥掉本
+      // --settings ⇒ 无数据 ⇒ 卡片省略配额段（fail-open），不做全局兜底。
+      // refreshInterval=60：实测冷启动 0.24–0.34s，每分钟一次可承受，且能在无消息时
+      // 跟上 5h/7d 窗口滚动；快照 10 min 陈旧自动失效（STATUSLINE_STALE_MS）。
+      if (variant.id === 'claude-code') {
+        inlineSettings.statusLine = { type: 'command', command: statuslineHookCommand(), refreshInterval: 60 };
+      }
+      // claude-code 恒传 --settings（statusLine 总在）；其它 variant 仅在有内容（bypass 键）
+      // 时才传，disableCliBypass 下没东西可传就不传。
       // （读隔离由 worker 的整进程 Seatbelt wrapper 强制，这里不注入任何 sandbox 设置——
       // 注入内置 sandbox 会嵌套沙箱且 permissions deny>allow 会挡掉 memory carve-out。）
       if (Object.keys(inlineSettings).length > 0) {
@@ -924,7 +989,10 @@ export function createClaudeFamilyAdapter(variant: ClaudeFamilyVariant, rawBin:
       // `claude` never surfaces/mis-fires `botmux send` etc.
       args.push('--plugin-dir', CLAUDE_PLUGIN_DIR);
       if (skillPluginDir) args.push('--plugin-dir', skillPluginDir);
-      args.push('--append-system-prompt', buildBotmuxSystemPromptText({ locale, botName, botOpenId, noTransport, triggerUserAuth }));
+      // replyDelivery=transcript：系统提示改口为「最终回复由 botmux 自动转发」。v3 workflow
+      // 子会话（GOAL_ENV.V3_MARKER）的收口靠 botmux send，强制保持 send 措辞。
+      const effectiveReplyDelivery = process.env[GOAL_ENV.V3_MARKER] === '1' ? 'send' : replyDelivery;
+      args.push('--append-system-prompt', buildBotmuxSystemPromptText({ locale, botName, botOpenId, noTransport, triggerUserAuth, replyDelivery: effectiveReplyDelivery, solo }));
       return args;
     },
 

src/adapters/cli/fs-policy.ts

@@ -901,6 +901,13 @@ export function buildFsPolicy(ctx: FsPolicyContext): FsPolicy {
   // worker PRE-CREATES this file before spawn so it survives the existence
   // filter and bwrap can bind it (bwrap cannot bind a nonexistent source).
   if (ctx.sessionId) push([`${sd}/turn-sends/${ctx.sessionId}.jsonl`], 'readWrite', 'internal');
+  // statusline: `botmux statusline` (Claude's statusLine.command, run INSIDE the
+  // sandbox) atomically writes `statusline/<sessionId>/latest.json`. Atomic
+  // write = tmp + rename in the parent dir, so a single-file grant (as for
+  // turn-sends) cannot work — grant the per-session DIRECTORY instead. Still
+  // session-scoped: sibling sessions' dirs are not exposed. The worker
+  // pre-creates the dir so bwrap has a bind source.
+  if (ctx.sessionId) push([`${sd}/statusline/${ctx.sessionId}`], 'readWrite', 'internal');
   // (schedules: stored PER BOT inside each BOT_HOME — the owner's dir is
   // already readWrite above and siblings' stores are denied by construction,
   // so the old shared data/schedules.json grant (and the cross-bot task-prompt
@@ -1013,6 +1020,8 @@ export function buildFsPolicy(ctx: FsPolicyContext): FsPolicy {
       `${ctx.sessionDataDir}/bin`,
     ], 'readOnly', 'internal');
     if (ctx.sessionId) push([`${ctx.sessionDataDir}/turn-sends/${ctx.sessionId}.jsonl`], 'readWrite', 'internal');
+    // statusline snapshot dir (see the larkTransport branch for why a dir, not a file).
+    if (ctx.sessionId) push([`${ctx.sessionDataDir}/statusline/${ctx.sessionId}`], 'readWrite', 'internal');
     // NOTE: dashboard-daemons (sibling IPC port table) and .dashboard-secret/-token
     // are deliberately NOT re-allowed — a no-transport turn has no business
     // reaching sibling daemons, and the secret is the escalation vector.