ShellFrame v0.40.1

Fixes

• The Cursor Agent install button works on Windows.
  On Windows the install action opened a PowerShell tab and typed the macOS and
  Linux installer, a curl download piped into bash, which PowerShell cannot run.
  The provider table held one Unix command for every platform. Windows now gets
  the official PowerShell installer (irm 'https://cursor.com/install?win32=true' | iex);
  macOS and Linux keep the curl command. Covered by tests_cursor_provider.py.

  Cursor Agent 的安裝按鈕在 Windows 上可以用了。
  Windows 按安裝時，會開一個 PowerShell 分頁，貼上 macOS／Linux 的安裝指令，
  也就是 curl 下載後交給 bash，PowerShell 跑不起來。原因是 provider 表對所有
  平台只放一條 Unix 指令。現在 Windows 改用官方的 PowerShell 安裝指令
  （irm 'https://cursor.com/install?win32=true' | iex），macOS 與 Linux 維持
  curl。由 tests_cursor_provider.py 守著。

Changes

• The route hint sends a large implementation to cursor-agent first, and names the other half of the pair.
  In v0.40.0 an implementation task was suggested to Grok Build before
  cursor-agent, and the hint did not say who should plan or review the work.
  The hint now prefers cursor-agent on the subscription model composer-2.5 for
  implementation (Grok Build stays the fallback, also inside the subscription),
  keeps planning on Claude Opus, and adds a review tier that also goes to
  Claude Opus. A task that says it is a review counts as review even when it
  also mentions design. Each hint names the counterpart: the implementer for a
  plan or review, the planner and reviewer for an implementation. It is still
  advice only; nothing is retargeted, started, or billed. Covered by
  tests_agent_route.py.

  路由建議把大量實作先交給 cursor-agent，並點出配對的另一側。
  v0.40.0 的實作任務會先建議 Grok Build，再來才是 cursor-agent，建議也沒說
  誰該規劃或審查。現在實作優先建議訂閱內的 cursor-agent composer-2.5（Grok
  Build 是備選，同樣在訂閱內），規劃仍給 Claude Opus，另外新增審查這一檔，
  同樣給 Claude Opus。任務寫明是審查時，即使提到設計也算審查。每則建議都會
  點出配對的另一側：規劃或審查配實作方，實作配規劃與審查。這仍然只是建議，
  不改派、不啟動、不花額度。由 tests_agent_route.py 守著。

Added

• A handoff can come back as a diff summary, and the requester can send the next step to the same agent.
  Claude Code and cursor-agent are meant to work as a pair, one planning or
  reviewing and the other implementing, in either direction. After the first
  handoff there was no way to say the result was a set of file changes, and the
  next step had to be addressed again by hand, which could land on a different
  tab. A result can now be marked as a diff summary, and a new command sends
  the next step as a fresh handoff to the same target, on this machine or the
  same paired peer, starting with the previous id. Only the requester can
  continue, and only after the previous result is in. Covered by
  tests_agent_link.py.

  交接結果可以是 diff 摘要，交辦方也能把下一步送回同一個 agent。
  Claude Code 與 cursor-agent 要配成一對，一邊規劃或審查、一邊實作，方向可以
  互換。第一次交接之後，沒辦法標明結果是一組檔案改動，下一步也得重新手動指定
  目標，可能送到別的分頁。現在結果可以標成 diff 摘要，新的指令把下一步當成一筆
  新的交接，送回同一個目標（同一台，或同一台已配對的機器），本文以上一筆的
  編號開頭。只有交辦方能接續，而且要等上一筆回報之後。由 tests_agent_link.py
  守著。

• A roster role can be the coordinator, which hands out work but is not handed implementation.
  With several machines handing work around, the agent that splits and checks
  the work could itself receive an implementation task and stop coordinating.
  A roster entry can now set a coordinator flag. A handoff aimed at that role,
  by role name or by its session id, is refused. Messages and results still
  reach it. Covered by tests_agent_link.py.

  roster 可以指定一個總管角色，只派工與驗收，不接實作。
  多台機器互相交接時，負責拆工作與驗收的那個 agent 也可能被交接實作，結果
  不再協調。現在 roster 項目可以設總管旗標。交接到那個角色，不論用角色名或
  它的 session id，都會被拒。一般訊息與結果仍然送得到。由
  tests_agent_link.py 守著。

• A host without ShellFrame can run a headless cursor-agent task over SSH, read-only by default on production.
  Some machines that should take work do not run ShellFrame, so they cannot be
  Frame Link peers. A paired ShellFrame host can now use SSH only as an
  executor: it runs cursor-agent in print mode on the configured host and sends
  the output back to the requesting tab. This is not a second transport and
  nothing is queued when the host is down. Hosts are listed in the local config
  only, never in the defaults. A host marked production starts in plan or ask
  mode without force; writing needs an explicit flag on that one task, and a
  task that restarts or stops a service is refused before any connection, even
  with that flag. The model flag is limited to subscription models. The command
  returns at once and the output, capped at 4000 characters, or a failure
  notice arrives in the tab later, so a slow host does not hold up other
  commands. On a PowerShell host the prompt is reduced to characters that the
  Windows launcher cannot re-split. Covered by tests_a2a_ssh.py and
  tests_agent_link.py, with a fake runner; no real host is contacted in tests.

  沒有 ShellFrame 的主機可以透過 SSH 跑 headless cursor-agent 任務，正式環境預設唯讀。
  有些該接工作的機器沒有跑 ShellFrame，所以不能成為 Frame Link 的配對機器。
  現在已配對的 ShellFrame 主機可以把 SSH 只當執行器：在設定好的主機上以列印
  模式跑 cursor-agent，把輸出送回交辦的分頁。這不是第二條傳輸通道，主機離線
  也不排隊。主機只寫在本機設定，不進預設值。標成正式環境的主機以 plan 或 ask
  模式啟動，不加 force；要寫入得在那一筆任務明確加旗標，而重啟或停止服務的
  任務在連線前就拒絕，加了旗標也一樣。模型旗標只允許訂閱內的模型。指令會立刻
  返回，輸出（上限 4000 字）或失敗通知稍後送進分頁，慢的主機不會卡住其他指令。
  PowerShell 主機上的提示會先換掉 Windows 啟動器會重新切開的字元。由
  tests_a2a_ssh.py 與 tests_agent_link.py 用假的執行器守著，測試不連任何真實
  主機。