Commit Graph
9 Commits
Author SHA1 Message Date
host 125b436ab3 browser: click-path allowlist, safe tenant script upgrade, allowlist activation plan
- click/click_ref: pre-click allowlist gate on resolvable anchor href
  (same check as navigate; miss refuses before the click executes);
  click/click_ref/click_at: post-click landing URL check.
  click_at has no pre-click gate by design (coordinates name no
  element); JS-driven navigation, evaluate, meta refresh and popups
  are documented as unintercepable with the Tinyproxy 403 backstop.
- Tenant script upgrade is now safe: only proven-unmodified scripts
  (manifest match, 2a569ac known-shipped SHA-256, missing, or
  current) auto-merge; tenant-modified scripts are never overwritten,
  are reported as skipped, and get one deduped audit record. The
  manifest never learns a tenant hash (else the next upgrade would
  mistake a modified file for untouched). Idempotent reruns.
- Activation candidate = deployed Tinyproxy programming set encoded
  with leading-dot entries (plain entries would regress
  docs.python.org); npmjs.org per the deployed filter, not npmjs.com.
  Host activation itself is NOT performed this round: only the plan,
  validation tests and rollback are delivered.
Tests: 28 new (click 13, upgrade 8, activation 7), all fake/no-network;
98 browser-related tests pass including prior suites.

(cherry picked from commit e4a93f5748ec815f163d0bc096ff4dc4b8c651f5)
2026-11-01 09:06:28 +00:00
host da7a9950fb feat: platform browser multi-tab, navigation allowlist, close cleanup, egress filter, allowlist worker wiring
- Multi-tab: browser_new_page / browser_list_pages /
  browser_switch_page / browser_close_page in playwright_mcp.py;
  each new page gets a fresh Playwright context, so cookies and
  storage are isolated between pages. Single-page tools act on
  the current page; switching to a missing index is an error,
  not a crash; closing the last page is an error (use
  browser_close).
- Navigation allowlist: BROWSER_ALLOWED_HOSTS (comma list,
  leading-dot suffix match, case/port-insensitive) checked in
  _check_allowed_url before launching Chromium; unset (or empty)
  means no check, the old behaviour. file:// always allowed;
  localhost/127.0.0.1 have no special treatment.
- Close cleanup: browser_close closes all pages and best-effort
  removes cwd browser-live.png / browser-live.json; custom
  screenshot files are untouched. list_pages needs no launch.
- Egress filter: 10 programming-source Tinyproxy BASIC rules
  (python.org, nodejs.org, go.dev, pkg.go.dev, rust-lang.org,
  developer.mozilla.org, crates.io, stackoverflow.com,
  cve.mitre.org, nvd.nist.gov); already-present github.com /
  pypi.org / npmjs.org not duplicated; searx.diante.me stays
  exact-only, no diante.me wildcard.
- Allowlist worker wiring: _docker_task_context now passes a
  set, non-empty BROWSER_ALLOWED_HOSTS from the worker env into
  [mcp_servers.playwright.env] in the tenant config, so the MCP
  process actually sees it (container -e alone does not reach
  MCP subprocesses, 2026-10-30). Unset/empty writes nothing and
  never overwrites; the value never goes into logs; it is not
  added to the websearch env block or the container env.
  ensure_playwright_mcp_env now returns False without rewriting
  when the result would be identical (canonical join: one blank
  line before the block), so a second prep leaves an unchanged
  tenant config alone; other sections untouched.
- Skill + docs: browser-automation SKILL.md multi-tab/allowlist
  section; docs/platform-browser.md comparison doc (new file).
- Tests (mock/fake only, no network, no real Chromium, no real
  credentials): cloud.tests.test_playwright +
  cloud.tests.test_websearch_worker, 48 tests OK, including 6
  new wiring tests (switch off passes no variable, switch on
  reaches MCP env verbatim, empty never written, old-tenant
  merge idempotent + no-rewrite, ensure no-op when identical,
  allowlist enforces only after wiring).
- NOT done (see PR body): no persistent profile, no credential
  vault, no approval card, no CAPTCHA bypass, in-page link
  navigation does not pass the allowlist, no real-Chromium live
  verification, host tinyproxy filter not reloaded.
2026-11-01 08:34:13 +00:00
host cbf5e65a7a perf: 小优化组 + 恢复浏览器页签
1. MCP 空闲睡眠 50ms -> 10ms:#187 后单轮 drain 约 0.3ms,
   100 次/s 空闲迭代约 3% 单核,请求与远控输入拾取延迟
   从平均 25ms 降到 5ms。
2. stream 轮询 50ms -> 20ms。
3. stream payload 加 url/title;面板 stream 处理合并 url/
   title,5 秒 state 轮询只在未观看时跑(观看时 stream 已
   全量覆盖,轮询纯冗余;不观看时保留轮询,避免会话开始
   后页眉 stale)。
4. 恢复浏览器页签:#176(ccb5a86)改写 TaskDetailPanel
   页签时顺带删了 BrowserPanel 的挂载,导致线上控制台
   自 #176 起就没有浏览器页签(面板文件一直在,只是不
   被 import,构建产物因此不变)。本次把 import、页签
   与 Card 原样挂回;构建产物含“开始观看”标记可证。
console.html 由 npm run build:console 重打(342484 字节)。
console-redesign/ 为构建脚本审阅产物,不进仓库。
2026-10-30 14:09:02 +08:00
host 910168126d perf: 桥变化才写、stdin 优先、stream 50ms 轮询
1. 桥写放大:旧代码每 ~50ms 轮询都无条件重写 state 和帧
   文件(约 40 次原子 temp+replace/s 走 runsc gofer,
   泵持久返回同一帧、updated_at 每轮都变,字节其实不变)。
   现帧字节不变不写;state 只在 url/title/started/holder/
   streaming 变化时写,或 0.5s 心跳一次(判死阈值 30s,
   绰绰有余);静态页写量从 ~40次/s 降到 ~2次/s。
2. MCP 主循环 stdin 优先:旧循环先 drain 再读 stdin,
   每次工具调用都先付一轮文件 IO。现 pending 请求先处理,
   空闲轮询 drain,忙时 200ms 保底 drain 一次防饿死。
3. stream 轮询 200ms → 50ms,画面延迟下限随之下降。
附回归测试:无变化轮询不重写、帧变化立即写、holder
变化绕过节流立即写。
2026-10-30 13:36:24 +08:00
host f85de63120 fix: MCP 主循环 stdin 批量行饥饿
select + buffered readline 会丢批量到达的行:一行 readline
可能把多行一起读进用户态 buffer,之后 select 再也看不到
空 fd,buffer 里的行永远不被读。客户端在 initialize 后
习惯把 notification + 下一请求连写,MCP 于是只答 initialize
就卡死(生产 2ff8975 活体验证两跑复现,faulthandler 证实
主线程停在 select)。改用非阻塞 fd + 自有 buffer 的行泵,
每行按到达顺序必达;附带回归测试(批量行全达 + EOF 残行)。
2026-10-30 12:47:23 +08:00
host 6bcdeb8da1 fix: 浏览器远控传输层改文件桥(Unix socket 在 runsc 下宿主不可见)
根因(2026-10-30 活体实证):#181 的 .browser-remote.sock 假设
宿主能连容器内 socket,但生产 Docker 任务走 runsc/gVisor,
其 gofer 挂载下容器内建的 socket 只在容器内可见,宿主目录
里只有普通文件(browser-live.png 一直能透传)。后果:server
代理对真实任务一律 503,会话根本没活过。

修法:传输层改四个普通文件(worktree 内,原子写 + 追加日志):
- .browser-remote-state.json:MCP 每轮重写,mtime 即存活信号,
  server 判 30 秒无心跳为会话死亡(503);
- .browser-remote-frame.jpg:最新 screencast 帧(base64 文本);
- .browser-remote-input.jsonl:server 追加,MCP 按偏移量消费
  完整行,执行时再判 holder,半行等下一轮;
- .browser-remote-control.json:{holder, seq},MCP 首轮即应用,
  接管可先于浏览器启动。
- holder 竞态:control 写下后 state 文件约 50ms 才刷新,server
  的 input/stream 改判 effective_holder(控制文件 mtime 新者胜),
  否则接管后立刻点会误 409。
- 线程契约与控制权模型不变;MCP 主循环仍是唯一触页线程。
- 播种刷新(#181)让老租户自动拿到新版脚本。

测试:socket 测试改文件桥测试(桥轮询/偏移量/半行/陈旧态/
effective_holder + server 代理全链),全量 1357 项通过(skipped 4)。
2026-10-30 07:28:32 +08:00
host 6f39ab3945 feat: 浏览器二期后端:远控服务 + ref 快照 + server 代理
- playwright_remote.py(新):MCP 进程内 Unix socket HTTP 服务,
  监听 worktree/.browser-remote.sock;GET /state、GET /stream
  (SSE 画面流)、POST /input、POST /control(接管/交还)。
  线程契约(实测):Playwright 同步调用只能在创建页面的线程,
  所以 HTTP 线程只读属性/读最新帧/入队,MCP 主循环用
  select 轮询(50ms) draining 输入并 pump screencast 帧
  (无 ack 则 CDP 首帧后停流)。
- 画面流走 CDP screencast 帧(学 chrome-devtools-mcp),不用
  定时截图;ffmpeg 录像暂不做。
- snapshot_refs/click_ref(学 playwright-mcp 的 ref 机制)+
  mouse_move/key_press/type_text;agent 工具在 holder=user 时
  直接报错,等用户交还。
- server.py:/tasks/{id}/browser/{state,stream,input,control}
  代理到 worktree socket,复用任务鉴权与 worktree 解析;
  会话不存在时 503(轮次结束即消失是设计)。
- 播种器:MCP 脚本是系统文件,版本变化时刷新老租户旧脚本
  (script_refreshed 打日志),否则远控只对新项目生效;
  新增 playwright_remote.py 进播种清单。
- 队列遮蔽修复:cloud/queue.py 会遮蔽 stdlib queue,改用
  SimpleQueue(deque+锁),仓库模式与 flat-copy 两路都通。
- 测试:新 20 项(remote 服务、server 代理、ref/输入/
  screencast fake),全量 1354 项通过(skipped 4)。
2026-10-30 06:44:24 +08:00
host 814fe8f7c2 feat: 浏览器一期:坐标点击 + live 截图
- browser_click_at(x, y):视口坐标点击(1280x800、左上角
  原点),越界与非数字拒绝;live 视图里用户在截图上的
  点击落点就走这个工具。
- browser_live_shot:固定文件 browser-live.png + 状态
  sidecar browser-live.json(URL/标题/时间戳),watcher
  轮询一个稳定路径即可;技能要求每次动作后都调一次,
  保持画面最新。
- MCP 新增 browser_click_at / browser_live_shot 两个工具;
  技能补 live 视图一节;测试 9 项(含 fake mouse、越界、
  live 文件与状态读回)。
2026-10-30 06:19:55 +08:00
host c9a483282d feat: 浏览器自动化 MCP(Playwright)+ browser-automation 技能
- playwright_browser.Browser:单页 Chromium,navigate /
  snapshot(正文 8000 字截断标记)/ click / fill / evaluate /
  screenshot(工作区内)/ close;启动参数 --no-sandbox +
  --disable-dev-shm-usage(runsc + 1003 用户实测必需),
  代理显式传给 launch(Chromium 在 Linux 下不稳定地读 env
  代理),浏览器路径 PLAYWRIGHT_BROWSERS_PATH=/opt/ms-playwright。
- playwright_mcp:stdio MCP,7 个 browser_* 工具,同一进程
  单 Browser 实例;测试用 fake browser / fake factory。
- worker:--playwright-mcp flag,_tenant_dirs 播种两个脚本 +
  [mcp_servers.playwright],_prepare_tenant 写 env(代理 +
  浏览器路径,MCP 子进程不继承 Codex 进程 env,镜像 ENV
  在此必须重写,缺了会报 Executable doesn't exist)。
- 技能 browser-automation:JS 页、翻页、表单、截图时用,
  静态页优先 web_fetch / web_search;截图须 PIL 读回验证。
- 镜像(层构建,非本 PR):FROM latest + fonts-noto-cjk +
  playwright + chromium(npmmirror)+ install-deps,3.77GB;
  runsc + 1003 + docker-codex 条件下 JS 执行、example.com
  标题、截图均已实测通过。
2026-10-30 05:14:09 +08:00