@modelzen/feishu-codex-bridge 0.3.4 → 0.3.6

package/README.md

@@ -165,6 +165,7 @@ feishu-codex-bridge bot rm <名>     # 移除一个机器人配置
   - **💬 单会话群**：整群就是**一条连续会话**（全程**免 @**、消息按序排队、无 `/resume`）。适合**个人单线深入**、像私聊一样直接聊。
 - **干活**：在项目群里 **@机器人** 描述需求；机器人在该群绑定的目录里跑 Codex，流式卡片回结果。
 - **话题 = 会话**：对某条消息开话题后，话题内可**免 @** 连续对话，是一条连贯的 Codex 会话。
+- **发图 / 发附件**：直接在消息里**发图片**，Codex 能看到（多模态读图）；**发文件附件**（日志 / PDF / 代码等），桥会把它下载到本地并把**绝对路径**告诉 Codex，让它用工具直接打开分析。⚠️ 附件落在桥的全局临时目录（`~/.feishu-codex-bridge/inbound`，1h 后自动清），**只有「完全访问」档**能读到——「项目内只读 / 读写」档的沙箱把读取锁在项目目录内，读不到该目录。单文件上限 50MB、单条消息最多 9 个；合并转发里的附件飞书官方不支持取，故不支持。
 - **文档评论 @机器人**：在飞书文档评论里 @ 它就回（前提：已开通文档评论权限 + 订阅 `drive.notice.comment_add_v1`，且机器人对该文档有访问权限）。只支持 doc/docx/sheet/file；评论框不渲染 markdown，回复为纯文本，超长会截断。
 - **终止**：卡片上的 **⏹** 随时终止当前轮；卡死超过 watchdog 阈值（默认 120s）自动中止并回收进程。
 - **私聊控制台**：项目列表、设置（模型 / 推理强度 / 免 @ / watchdog / 管理员）、用量、诊断、重连，全在私聊菜单里。

package/dist/cli.js

@@ -68,7 +68,10 @@ var paths = {
    * passes lark-cli's AssertSecurePath audit.
    */
   secretsGetterScript: join(appDir, "secrets-getter"),
-  mediaDir: join(appDir, "media")
+  mediaDir: join(appDir, "media"),
+  /** Inbound file attachments downloaded from chat, handed to codex by absolute
+   * path (codex has no native file input). TTL-pruned like {@link mediaDir}. */
+  inboundDir: join(appDir, "inbound")
 };
 
 // src/config/bots.ts
@@ -5404,7 +5407,7 @@ async function collectInboundImages(channel, msg) {
     return [];
   }
   if (refs.length === 0) return [];
-  await pruneOldMedia();
+  await pruneOldMedia(paths.mediaDir);
   try {
     await mkdir11(paths.mediaDir, { recursive: true });
   } catch {
@@ -5507,16 +5510,16 @@ function readHeader(headers, name) {
 function safeName(fileKey) {
   return fileKey.replace(/[^a-zA-Z0-9_-]/g, "").slice(-40) || "img";
 }
-async function pruneOldMedia() {
+async function pruneOldMedia(dir) {
   let entries;
   try {
-    entries = await readdir2(paths.mediaDir);
+    entries = await readdir2(dir);
   } catch {
     return;
   }
   const cutoff = Date.now() - MEDIA_TTL_MS;
   for (const name of entries) {
-    const file = join14(paths.mediaDir, name);
+    const file = join14(dir, name);
     try {
       const st = await stat3(file);
       if (st.mtimeMs < cutoff) await rm4(file, { force: true });
@@ -5524,6 +5527,277 @@ async function pruneOldMedia() {
     }
   }
 }
+var MAX_FILES = 9;
+var MAX_FILE_BYTES = 50 * 1024 * 1024;
+function messageHasFiles(msg) {
+  return (msg.resources ?? []).some((r) => r.type === "file");
+}
+async function collectInboundFiles(channel, msg) {
+  const refs = [];
+  const seen = /* @__PURE__ */ new Set();
+  for (const r of msg.resources ?? []) {
+    if (r.type === "file" && r.fileKey && !seen.has(r.fileKey)) {
+      seen.add(r.fileKey);
+      refs.push({ messageId: msg.messageId, fileKey: r.fileKey, fileName: r.fileName });
+    }
+  }
+  if (refs.length === 0) return [];
+  await pruneOldMedia(paths.inboundDir);
+  try {
+    await mkdir11(paths.inboundDir, { recursive: true });
+  } catch {
+  }
+  const out = [];
+  for (const ref of refs.slice(0, MAX_FILES)) {
+    const f = await downloadOneFile(channel, ref);
+    if (f) out.push(f);
+  }
+  log.info("intake", "files", { found: refs.length, downloaded: out.length });
+  return out;
+}
+async function downloadOneFile(channel, ref) {
+  try {
+    const res = await channel.rawClient.im.v1.messageResource.get({
+      path: { message_id: ref.messageId, file_key: ref.fileKey },
+      params: { type: "file" }
+    });
+    const declared = Number(readHeader(res.headers, "content-length"));
+    if (Number.isFinite(declared) && declared > MAX_FILE_BYTES) {
+      log.warn("intake", "file-too-large", { fileKey: ref.fileKey.slice(0, 24), bytes: declared });
+      return void 0;
+    }
+    const name = cleanFileName(ref.fileName) || "attachment";
+    const onDisk = `${safeName(ref.fileKey)}-${name}`;
+    const file = join14(paths.inboundDir, onDisk);
+    await res.writeFile(file);
+    try {
+      const st = await stat3(file);
+      if (st.size > MAX_FILE_BYTES) {
+        await rm4(file, { force: true });
+        log.warn("intake", "file-too-large", { fileKey: ref.fileKey.slice(0, 24), bytes: st.size });
+        return void 0;
+      }
+    } catch {
+    }
+    return { path: file, name };
+  } catch (err) {
+    log.warn("intake", "file-download-failed", { fileKey: ref.fileKey.slice(0, 24), err: String(err) });
+    return void 0;
+  }
+}
+function cleanFileName(name) {
+  if (!name) return "";
+  const base = name.split(/[/\\]/).pop() ?? name;
+  const cleaned = base.replace(/[\x00-\x1f<>:"|?*]/g, "_").replace(/\s+/g, " ").trim().slice(0, 100);
+  return cleaned === "." || cleaned === ".." ? "" : cleaned;
+}
+function stripFileTokens(text) {
+  return text.replace(/<file\b[^<]*\/>/g, "").replace(/[ \t]+\n/g, "\n").trim();
+}
+function weaveFileManifest(text, files) {
+  const stripped = stripFileTokens(text);
+  if (files.length === 0) return stripped;
+  const lines = files.map((f) => `- ${f.name} \u2192 ${f.path}`).join("\n");
+  const head = stripped ? `${stripped}
+
+` : "";
+  return `${head}[\u7528\u6237\u4E0A\u4F20\u4E86 ${files.length} \u4E2A\u9644\u4EF6\uFF0C\u5DF2\u4FDD\u5B58\u5230\u672C\u5730\uFF0C\u53EF\u7528 shell / \u8BFB\u53D6\u5DE5\u5177\u6309\u4E0B\u9762\u7684\u7EDD\u5BF9\u8DEF\u5F84\u76F4\u63A5\u6253\u5F00\uFF1A
+${lines}
+]`;
+}
+
+// src/bot/context-weave.ts
+var QUOTE_MAX = 800;
+var LINE_MAX = 280;
+var THREAD_WEAVE_MAX = 20;
+var THREAD_PAGE_SIZE = 50;
+async function fetchQuotedMessage(channel, messageId) {
+  try {
+    const res = await channel.rawClient.im.v1.message.get({ path: { message_id: messageId } });
+    const items = res.data?.items ?? [];
+    const item = items[0];
+    if (!item || item.deleted) return void 0;
+    const cm = toContextMessage(item);
+    return cm.text.trim() ? cm : void 0;
+  } catch (err) {
+    log.warn("intake", "quote-fetch-failed", { messageId, err: String(err) });
+    return void 0;
+  }
+}
+async function fetchThreadContext(channel, threadId, opts = {}) {
+  const limit = opts.limit ?? THREAD_WEAVE_MAX;
+  const since = opts.sinceTime ?? 0;
+  try {
+    const res = await channel.rawClient.im.v1.message.list({
+      params: {
+        container_id_type: "thread",
+        container_id: threadId,
+        sort_type: "ByCreateTimeDesc",
+        page_size: THREAD_PAGE_SIZE
+      }
+    });
+    const items = res.data?.items ?? [];
+    const picked = items.filter((it) => !it.deleted).map(toContextMessage).filter(
+      (m) => m.fromUser && // drop the bot's own replies, other apps, system notices
+      m.messageId !== opts.excludeMessageId && // drop the triggering @ message
+      (since === 0 || m.createTime > since) && // delta only for existing sessions
+      m.text.trim().length > 0
+    );
+    picked.sort((a, b) => a.createTime - b.createTime);
+    const out = picked.slice(-limit);
+    if (picked.length > out.length) {
+      log.info("intake", "thread-context-truncated", { threadId, kept: out.length, total: picked.length });
+    }
+    return out;
+  } catch (err) {
+    log.warn("intake", "thread-context-failed", { threadId, err: String(err) });
+    return [];
+  }
+}
+function toContextMessage(item) {
+  const id = item.sender?.id ?? "";
+  const name = item.sender?.sender_name || (id ? `\u7528\u6237${id.slice(-4)}` : "\u67D0\u4EBA");
+  return {
+    messageId: item.message_id ?? "",
+    senderName: name,
+    text: extractMessageText(item.msg_type, item.body?.content, item.mentions),
+    fromUser: item.sender?.sender_type === "user",
+    createTime: Number(item.create_time) || 0
+  };
+}
+function extractMessageText(msgType, content, mentions) {
+  if (!content) return placeholderFor(msgType);
+  let parsed;
+  try {
+    parsed = JSON.parse(content);
+  } catch {
+    return placeholderFor(msgType);
+  }
+  switch (msgType) {
+    case "text":
+      return replaceMentions(parsed?.text ?? "", mentions);
+    case "post":
+      return replaceMentions(extractPostText(parsed), mentions);
+    case "image":
+      return "[\u56FE\u7247]";
+    case "audio":
+      return "[\u8BED\u97F3]";
+    case "media":
+      return "[\u89C6\u9891]";
+    case "file": {
+      const name = parsed?.file_name;
+      return name ? `[\u6587\u4EF6\uFF1A${name}]` : "[\u6587\u4EF6]";
+    }
+    case "sticker":
+      return "[\u8868\u60C5]";
+    case "interactive":
+      return "[\u5361\u7247\u6D88\u606F]";
+    case "share_chat":
+      return "[\u5206\u4EAB\u7FA4\u540D\u7247]";
+    case "share_user":
+      return "[\u5206\u4EAB\u4E2A\u4EBA\u540D\u7247]";
+    case "merge_forward":
+    case "forward":
+      return "[\u5408\u5E76\u8F6C\u53D1\u6D88\u606F]";
+    default:
+      return placeholderFor(msgType);
+  }
+}
+function placeholderFor(msgType) {
+  return msgType ? `[${msgType} \u6D88\u606F]` : "[\u6D88\u606F]";
+}
+function extractPostText(parsed) {
+  if (!parsed || typeof parsed !== "object") return "";
+  const obj = parsed;
+  let title = obj.title;
+  let blocks = obj.content;
+  if (!Array.isArray(blocks)) {
+    for (const v of Object.values(obj)) {
+      if (v && typeof v === "object" && Array.isArray(v.content)) {
+        title = v.title;
+        blocks = v.content;
+        break;
+      }
+    }
+  }
+  const parts = [];
+  if (typeof title === "string" && title.trim()) parts.push(title.trim());
+  if (Array.isArray(blocks)) {
+    for (const line of blocks) {
+      if (!Array.isArray(line)) continue;
+      const lineText = line.map(nodeToText).join("");
+      if (lineText) parts.push(lineText);
+    }
+  }
+  return parts.join("\n");
+}
+function nodeToText(node) {
+  if (!node || typeof node !== "object") return "";
+  const n = node;
+  switch (n.tag) {
+    case "text":
+      return typeof n.text === "string" ? n.text : "";
+    case "a":
+      return typeof n.text === "string" ? n.text : typeof n.href === "string" ? n.href : "";
+    case "at": {
+      const name = typeof n.user_name === "string" ? n.user_name : "";
+      return name ? `@${name}` : "@\u67D0\u4EBA";
+    }
+    case "img":
+      return "[\u56FE\u7247]";
+    case "media":
+      return "[\u89C6\u9891]";
+    case "emotion":
+      return "[\u8868\u60C5]";
+    default:
+      return typeof n.text === "string" ? n.text : "";
+  }
+}
+function replaceMentions(text, mentions) {
+  if (!text || !mentions?.length) return text;
+  let out = text;
+  for (const m of mentions) {
+    if (!m.key) continue;
+    out = out.split(m.key).join(m.name ? `@${m.name}` : "@\u67D0\u4EBA");
+  }
+  return out;
+}
+function sanitizeContext(s, maxLen, oneLine2) {
+  if (!s) return "";
+  let out = s.replace(/[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]/g, "").replace(/\r\n?/g, "\n");
+  out = oneLine2 ? out.replace(/\s+/g, " ") : out.replace(/[ \t]+\n/g, "\n").replace(/\n{3,}/g, "\n\n");
+  out = out.trim();
+  return out.length > maxLen ? `${out.slice(0, maxLen)}\u2026` : out;
+}
+function weaveQuote(text, quoted) {
+  if (!quoted) return text;
+  const who = sanitizeContext(quoted.senderName, 40, true) || "\u67D0\u4EBA";
+  const body = sanitizeContext(quoted.text, QUOTE_MAX, true);
+  if (!body) return text;
+  const block = `[\u7528\u6237\u5F15\u7528\u4E86\u4E00\u6761\u6D88\u606F\uFF08\u6765\u81EA ${who}\uFF09\uFF1A
+${body}
+]`;
+  const base = text.trim();
+  return base ? `${block}
+
+${base}` : block;
+}
+function weaveThreadHistory(text, msgs) {
+  if (msgs.length === 0) return text;
+  const lines = msgs.map((m) => {
+    const who = sanitizeContext(m.senderName, 40, true) || "\u67D0\u4EBA";
+    const body = sanitizeContext(m.text, LINE_MAX, true);
+    return body ? `${who}\uFF1A${body}` : "";
+  }).filter((l) => l.length > 0);
+  if (lines.length === 0) return text;
+  const block = `[\u8BDD\u9898\u4E2D\u5728\u6B64\u4E4B\u524D\u5DF2\u6709\u7684\u6D88\u606F\uFF08\u6309\u65F6\u95F4\u5148\u540E\u6392\u5217\uFF0C\u4F9B\u4F60\u7406\u89E3\u4E0A\u4E0B\u6587\uFF09\uFF1A
+${lines.join("\n")}
+]`;
+  const base = text.trim();
+  return base ? `${block}
+
+${base}` : block;
+}
 
 // src/bot/comments.ts
 var SUPPORTED_FILE_TYPES = /* @__PURE__ */ new Set(["doc", "docx", "sheet", "file"]);
@@ -6010,20 +6284,36 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
     const perm = turnPerm(project, senderId);
     return { sessionKey: perm.roleSuffix ? `${baseKey}#${perm.roleSuffix}` : baseKey, ...perm };
   }
+  async function ingestContext(msg, text) {
+    let body = text;
+    if (messageHasFiles(msg)) {
+      const files = await collectInboundFiles(channel, msg);
+      body = weaveFileManifest(text, files);
+      if (!body.trim()) {
+        body = "\u7528\u6237\u53D1\u6765\u4E00\u4E2A\u9644\u4EF6\uFF0C\u4F46\u6865\u6CA1\u80FD\u4E0B\u8F7D\u5B83\uFF08\u53EF\u80FD\u8D85\u8FC7 50MB \u4E0A\u9650\u6216\u88AB\u98DE\u4E66\u62D2\u7EDD\uFF09\u3002\u8BF7\u544A\u8BC9\u7528\u6237\u9644\u4EF6\u6CA1\u8BFB\u5230\uFF0C\u53EF\u4EE5\u91CD\u53D1\uFF0C\u6216\u6539\u4E3A\u7C98\u8D34\u6587\u672C / \u53D1\u56FE\u7247\u3002";
+      }
+    }
+    if (msg.replyToMessageId) {
+      const quoted = await fetchQuotedMessage(channel, msg.replyToMessageId);
+      body = weaveQuote(body, quoted);
+    }
+    return body;
+  }
   async function handleTurn(msg, text, sessionKey, flat, project, perm) {
     const existing = active.get(sessionKey);
     if (existing) {
       const images = messageHasImages(msg) ? await collectInboundImages(channel, msg) : void 0;
+      const woven = await ingestContext(msg, text);
       const cur = active.get(sessionKey);
       if (!cur) {
-        startReservedRun(msg, text, sessionKey, flat, project, perm, images);
+        startReservedRun(msg, woven, sessionKey, flat, project, perm, images, true, text);
         return;
       }
       if (getPendingPolicy(cfg) === "steer" && cur.run && cur.thread) {
         const tid = cur.run.turnId();
         if (tid) {
           try {
-            await cur.thread.steer({ text, images }, tid);
+            await cur.thread.steer({ text: woven, images }, tid);
             log.info("intake", "steer", { tid, images: images?.length ?? 0 });
             return;
           } catch (err) {
@@ -6031,13 +6321,13 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
           }
         }
       }
-      cur.queue.push({ text, images });
+      cur.queue.push({ text: woven, images });
       log.info("intake", "queued", { depth: cur.queue.length });
       return;
     }
     startReservedRun(msg, text, sessionKey, flat, project, perm);
   }
-  function startReservedRun(msg, text, sessionKey, flat, project, perm, preloadedImages) {
+  function startReservedRun(msg, text, sessionKey, flat, project, perm, preloadedImages, preIngested, summaryText2) {
     const existing = active.get(sessionKey);
     if (existing) {
       existing.queue.push({ text, images: preloadedImages });
@@ -6050,7 +6340,15 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
       const reaction = runReaction(msg.messageId, !sema.hasFree());
       try {
         const images = preloadedImages ?? (messageHasImages(msg) ? await collectInboundImages(channel, msg) : void 0);
-        let thread = await resolveThread(sessionKey, msg.chatId, { mode: perm.mode, network: perm.network });
+        let firstText = preIngested ? text : await ingestContext(msg, text);
+        const { thread: resolved, recreated } = await resolveThread(sessionKey, msg.chatId, {
+          mode: perm.mode,
+          network: perm.network
+        });
+        let thread = resolved;
+        const neverSeen = !thread;
+        const codexEmpty = neverSeen || recreated;
+        const prior = neverSeen ? void 0 : await getSession(sessionKey);
         if (!thread) {
           const cwd = project?.cwd ?? fallbackCwd;
           thread = await backend.startThread({ cwd, mode: perm.mode, network: perm.network });
@@ -6060,11 +6358,23 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
             chatId: msg.chatId,
             cwd,
             codexThreadId: thread.codexThreadId,
-            summary: text.slice(0, 80),
+            // `text` is already file-woven when preIngested; use the raw
+            // `summaryText` (handleTurn's original) so the session label isn't
+            // manifest boilerplate + a temp path.
+            summary: stripFileTokens(summaryText2 ?? text).slice(0, 80),
+            lastSeenAt: msg.createTime,
             createdAt: Date.now(),
             updatedAt: Date.now()
           });
         }
+        if (msg.threadId && (codexEmpty || prior?.lastSeenAt !== void 0)) {
+          const history = await fetchThreadContext(channel, msg.threadId, {
+            sinceTime: codexEmpty ? 0 : prior?.lastSeenAt ?? 0,
+            excludeMessageId: msg.messageId
+          });
+          firstText = weaveThreadHistory(firstText, history);
+        }
+        if (!neverSeen) void patchSession(sessionKey, { lastSeenAt: msg.createTime }).catch(() => void 0);
         reserved.thread = thread;
         await launchRun(
           {
@@ -6073,7 +6383,7 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
             replyInThread: !flat,
             flat,
             thread,
-            firstText: text,
+            firstText,
             images,
             knownThreadId: sessionKey,
             requesterOpenId: msg.senderId
@@ -6090,9 +6400,9 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
   }
   async function resolveThread(threadId, chatId, perm) {
     const live = sessions.get(threadId);
-    if (live) return live;
+    if (live) return { thread: live, recreated: false };
     const rec = await getSession(threadId);
-    if (!rec) return void 0;
+    if (!rec) return { thread: void 0, recreated: false };
     try {
       const resumed = await backend.resumeThread({
         cwd: rec.cwd,
@@ -6103,7 +6413,7 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
         network: perm?.network
       });
       sessions.set(threadId, resumed);
-      return resumed;
+      return { thread: resumed, recreated: false };
     } catch (err) {
       log.fail("agent", err, { phase: "resume-on-turn", threadId });
       const project = await getProjectByChatId(chatId);
@@ -6116,7 +6426,8 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
         network: perm?.network ?? project?.network
       });
       sessions.set(threadId, fresh);
-      return fresh;
+      await patchSession(threadId, { codexThreadId: fresh.codexThreadId }).catch(() => void 0);
+      return { thread: fresh, recreated: true };
     }
   }
   async function evictLiveSessionsForChat(chatId) {
@@ -6147,8 +6458,8 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
         await channel.send(msg.chatId, { markdown: `\u274C \u542F\u52A8\u5931\u8D25\uFF1A${err instanceof Error ? err.message : String(err)}` }, { replyTo: msg.messageId }).catch(() => void 0);
         return;
       }
-      const firstText = text || "\u4F60\u597D\uFF0C\u6211\u4EEC\u5F00\u59CB\u5427\u3002";
       const images = messageHasImages(msg) ? await collectInboundImages(channel, msg) : void 0;
+      const firstText = await ingestContext(msg, text) || "\u4F60\u597D\uFF0C\u6211\u4EEC\u5F00\u59CB\u5427\u3002";
       log.info("card", "start", { project: project?.name ?? "(unregistered)", model, effort, images: images?.length ?? 0 });
       await launchRun(
         {
@@ -6161,7 +6472,7 @@ function createOrchestrator(channel, cfg, fallbackCwd) {
           model,
           effort,
           cwd,
-          summary: text.slice(0, 80) || "(\u7A7A)",
+          summary: stripFileTokens(text).slice(0, 80) || "(\u7A7A)",
           requesterOpenId: msg.senderId,
           roleSuffix: perm.roleSuffix
         },

package/dist/index.d.ts

@@ -42,6 +42,9 @@ declare const paths: {
      */
     secretsGetterScript: string;
     mediaDir: string;
+    /** Inbound file attachments downloaded from chat, handed to codex by absolute
+     * path (codex has no native file input). TTL-pruned like {@link mediaDir}. */
+    inboundDir: string;
 };
 
 export { log, newTraceId, paths, withTrace };

package/dist/index.js

@@ -47,7 +47,10 @@ var paths = {
    * passes lark-cli's AssertSecurePath audit.
    */
   secretsGetterScript: join(appDir, "secrets-getter"),
-  mediaDir: join(appDir, "media")
+  mediaDir: join(appDir, "media"),
+  /** Inbound file attachments downloaded from chat, handed to codex by absolute
+   * path (codex has no native file input). TTL-pruned like {@link mediaDir}. */
+  inboundDir: join(appDir, "inbound")
 };
 
 // src/core/logger.ts

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@modelzen/feishu-codex-bridge",
-  "version": "0.3.4",
+  "version": "0.3.6",
   "description": "Bridge Feishu/Lark messenger with local Codex via app-server (project=group, thread=session)",
   "type": "module",
   "bin": {