feat(agent-activity): replace the popup with a full-screen fleet view

[Juliusolsson05]

Fixes #1170

Replaces the Agent Activity popup with a full-screen view that answers three questions for every agent in the window, lane or pool: who needs me, who is working and on what, and what can I clean up. This PR started as Stage 1 (evidence and plan only). It now carries the whole redesign.

The problem

The owner's verdict on the old modal: "that modal is ages and just shit across the board". The evidence recorded in Stage 1, from the owner's own 33–48 agent window:

• it had no "needs you" signal;
• rows read as the folder name (3 of 33 sessions had a title);
• it listed panes, not agents, so parked pool sessions were missing;
• it still spoke in tabs;
• it had no filter, and its actions were hover-only and one at a time.

Owner decisions (2026-09-24)

1. Full screen.
2. Close Old Agents and Close Idle Orchestration Agents stay as quick one-shot commands.
3. An untitled row is named by its Goal, then its folder.

What it does

• Sections, in this order: Needs you, Working, Idle (oldest first, for cleanup) and Exited. Needs-you rows say why in plain words:
  • wants permission / asked a question / trust prompt;
  • turn failed, or process error;
  • hit a usage limit (only while newer than the current turn);
  • your prompt did not send;
  • goal loop blocked or paused;
  • a queued prompt behind an idle agent (a queue behind a working agent is normal and is not flagged).
• Names: title → Goal → folder. The row labels a Goal or folder name as such, so a folder is never mistaken for a chosen name. The TLDR is the second line.
• Keyboard:
  • ↑/↓ move, Enter opens the agent, Space selects, ⌘A selects all shown, ⌫ closes.
  • Any letter jumps into the filter, and in the filter Space is just a space.
  • The first Esc clears the filter; the second dismisses.
• Close: single rows and selections both go through closeGrantedSessions. It confirms once, carries its own bulk.agent-activity kill tag, records no Undo Close entries, and refuses an agent that started working after it was approved. A working agent the user deliberately selected does close.

Design decisions

• A full-viewport Radix Dialog, not a fourth MainSurface takeover. Every hidden-workspace rule (the useKeybinds picker gate and Escape, lease hygiene, the control observation) names each takeover by hand, so a takeover would mean four more edits, each easy to forget. The Dialog already owns input ownership, the focus trap, Escape and focus restore, and no terminal resizes. The reasoning is also in the decomposition doc.
• No new rules for "needs you". Each signal reuses its existing owner: conditionRequiresAttention, terminalProviderFailure, isLimitIdle, the goal-loop chip reading, and sessionActivity. The row model is a pure function, so the bulk close can re-ask it at the kill boundary.
• Notes are read in one batched IPC call per kind (TLDR, Goal, Goal Loop), only while the view is open. A closed view reads and subscribes to nothing, and it joins closedModalDerivations.

Verification

• activityRow.test.ts replays the recorded corpus from Stage 1: 71 real runtimes plus the owner's fleet with titles verbatim.
  • The corpus corrected the first draft. It holds no permission prompt or failed turn, but it does hold two prompts the app could not deliver, and exactly those two land in Needs you.
  • The 15 recorded queues, all behind a working agent, stay Working. Compaction banners are not attention.
• agentActivityClose.renderer.test.tsx drives the real close executor.
• AgentActivityView.renderer.test.tsx covers section order, one read per kind, and the keyboard grammar (typing a filter never selects or closes; a selection survives filtering).
• I broke three rules on purpose (queue-behind-working, idle ordering, the filter keyboard guard). The tests caught all three.
• tsc -b is clean except src/main/browserPocket/*. Those errors come from a local node_modules missing playwright-core, which origin/main added; CI installs it.
• Full suite: 6419 passed. The only failure is the same browser-pocket Playwright test, for the same reason. check:keybindings and test:contract pass.

Review round (one round: Codex ×2, Pi, Grok)

None of the four reviewers found a way to close an agent without confirmation, and none found a crash. Fixed:

• a failed compaction now needs you (the unread rule leaves compaction out);
• a failed start that then exited shows its error, not "Exited";
• a usage limit shows even while the provider still looks busy (Grok);
• Enter/Space on the "Close N selected" button do what the button says;
• Tab from the filter reaches the list with the query kept;
• scrolling no longer jumps back on every runtime update, and the first highlight is pinned to its agent;
• Goal text is searchable even when a title wins the name;
• extension views are left out;
• "may not have sent" is kept separate from "did not send";
• the sections are ARIA groups.

Each of these has a test.

Limits and follow-ups

• The UI has not been looked at in the running app. Layout and spacing are reasoned from source and tests only. A quick manual look is owed before merge.
• Grok agents never read "turn failed" (Grok has no failure signal, bug(orchestration): a child whose provider turn fails with an API error reports prompt_sent/waiting forever #1018).
• The phone's agent list does not use this model yet (Unknown 6).

🤖 Generated with Claude Code