Roadmap proposal: supervising agents from anywhere

[powerfooI]

Roamgate has grown quickly since August, with about 300 merged PRs covering
terminals, the Inspector, worktrees, review annotations, Ranger, mobile, SSH,
and Windows. Before adding more, I'd like to set a clearer direction and hear
from people who use it.

The proposal

Roamgate should be the best place to supervise coding agents from anywhere.
Start work on any computer. When an agent needs you, deal with it from your
phone in under a minute, then put the phone down.

The loop I want to make fast and reliable:

agent blocks or finishes -> notification -> open -> understand -> respond -> leave

Most of the pieces already exist (push notifications, agent status, History
chat with a message box, agent commands, Ranger tasks). The work now is joining
them into one flow.

Three tracks

1. Attention loop

• Triage view from a notification: tapping a push shows the agent's status,
  its last message or prompt, a diff summary, and a reply box, without hunting
  through tabs.
• "Needs you" inbox: one list across all workspaces of blocked and recently
  finished agents, kept in sync across your devices.
• Quick responses: one-tap replies for each agent type (continue, approve,
  deny, "show me the diff first"), using the existing "Send anyway" check.

2. Core hardening

• Clear the terminal correctness backlog: IME duplication on Linux/Wayland
  (Duplicated IME commit in a terminal pane on Linux/Wayland（终端面板中 IME 提交重复） #376) and the stale strip left after switching agents (Switching agents leaves a stale strip of the previous agent's output rendered on the right side of the terminal pane #171).
• Turn multi-client, reconnect, and "last known" behavior into tested
  contracts instead of fixing those cases one at a time.
• Publish a Herdr compatibility table showing which features need which Herdr
  version.
• Alternate feature releases with fix-only releases.

3. Ranger, with a narrower job

• Short summaries of inbox items: why an agent needs you, with sources and
  freshness. Only for workspaces you allow; it never acts on its own.
• A "what happened while I was away" digest built on the existing Tasks
  scheduler.
• Keep the actions people really use and drop the "experimental" label for
  those.

Later: dispatch agents from tasks

After the above is solid: a board where a card becomes a worktree, an agent,
and a branch, and the card's status comes from agent and PR/MR status (#174).

What I'm deliberately not prioritizing

• More depth in the desktop Inspector; your IDE already covers that.
• New agent integrations beyond keeping current ones working.
• Offline PWA behavior.

Questions for you

1. Do you mostly use Roamgate on your phone or on desktop? This decides
   how hard to push the attention loop.
2. When an agent needs you, what do you actually do? Approve, type a
   reply, read the diff, open the terminal?
3. Should Ranger's per-item summaries be on by default or opt-in, given
   model cost and what gets sent to the provider?
4. Anything that would make you stop trusting the remote view (terminal glitches,
   stale state, reconnect issues)? Please link issues or describe them here.
5. Is a task board (Add Kanban/Linear style support #174) something you'd use, and should it sync with
   Linear/GitHub Issues or stay local?

Replies, objections, and "you're missing X" are all welcome. I'll turn the
outcome into tracked issues.