feat: 支持 agent 发送图片给用户 (feishu_send_image)

[lishuceo]

背景

此前飞书集成只能读用户发来的图（feishu_download_message_image），不能发图给用户——client.ts 的发送方法只有 text/post/interactive，没有图片上传/发送能力。

本 PR 实现方案 A（独立 msg_type:'image' 消息），让 agent 把工作区里生成的图片（截图、图表、渲染结果）发回当前会话。

改动

src/feishu/client.ts — 三个新方法：

• uploadImage(buf|stream) → 调 im.image.create 拿 image_key（该接口返回已解包的 { image_key }，与其它 message 接口的 {code,msg,data} 不同）
• sendImage(chatId, imageKey) → msg_type:'image' 发到会话
• replyImageInThread(messageId, imageKey) → reply_in_thread 回复到话题

src/feishu/tools/send-image.ts（新）— feishu_send_image MCP 工具：

• 入参 file_path（本地绝对路径）
• 校验：存在性、isFile、大小 ≤ 10MB、非空、magic bytes 为图片（png/jpg/gif/webp/bmp）
• 上传后按 threadReplyMsgId 路由：话题内回复到话题，否则发到会话

接线 — tools/index.ts 增加 threadReplyMsgId 参数并 gate on chatId 注册；executor.ts 透传 ExecuteInput.threadRootMessageId。

测试

• send-image.test.ts：11 个用例（成功会话/话题、文件不存在/过大/为空、非图片、上传失败、发送失败、异常、jpg 识别）
• index.test.ts:更新工具计数（chatId 存在=12，缺省不注册 send-image）
• npm run typecheck 通过；npx vitest run 全量 1696 通过（唯一失败的 memory/quality.test.ts 为数据依赖的检索质量用例，基线同样失败，与本 PR 无关）

范围说明

方案 B（在 post/card 里图文混排）作为后续增强，本 PR 不含。

🤖 Generated with Claude Code

[github-actions]

Claude finished @lishuceo's task in 2m 46s —— View job

---

Code Review: feishu_send_image

• Read CLAUDE.md & PR context
• Check for previous review comments (none found)
• Review client.ts (upload/send/reply methods)
• Review send-image.ts (MCP tool + validation)
• Review wiring (tools/index.ts, executor.ts)
• Review tests

---

✅ Approved — clean, well-scoped, well-tested

This is a solid implementation of 方案 A. I verified the key risk areas and found no blocking issues.

Correctness verified:

• uploadImage unpacked return — Confirmed against the SDK type. node-sdk types im.image.create as Promise<{ image_key?: string } | null> (no {code,msg,data} wrapper), unlike the message APIs. So resp?.image_key is correct, and the | null case is safely handled by optional chaining + the if (!imageKey) guard in the tool. The inline comment documenting this quirk is appreciated.
• Thread routing — replyImageInThread(threadRootMessageId, …) is consistent with how replyTextInThread/replyPostInThread are called elsewhere (cron scheduler, restart notifications) using the thread root message id. ✔
• Wiring — feishu_send_image correctly gated on chatId (mirrors feishuMainChatTool), and threadRootMessageId is threaded through executor.ts → createFeishuToolsMcpServer. Tool count test updated to 12 with chatId / absent without. ✔
• Validation — exists → isFile → empty → size ≤ 10MB → magic-bytes is a sensible order, fails fast before upload, and every branch returns isError. The 11 test cases cover all branches. ✔

Non-blocking observations (no change required):

1. send-image.ts:77 — resolve(args.file_path) resolves relative paths against the server process cwd, not the agent's isolated workspace cwd. The tool description does say "绝对路径", and a relative path just yields a clear "不存在" error, so this is benign — but if you later see agents passing workspace-relative paths, that's the reason.
2. Supported formats (png/jpg/gif/webp/bmp) are a subset of what Feishu accepts (also TIFF/ICO per the SDK docs). Already noted as intentional scope; fine.
3. No path sandboxing on the read — but the agent already has filesystem read access via Read/Bash, so this introduces no new capability. isInsideSourceRepo only guards writes, which is unaffected here.

Note: I couldn't execute npx vitest run in this environment (command needs approval), so I relied on static analysis plus the PR's reported green run (1696 passing). The test structure looks correct.