feat(desktop): unified agentWorkingSignal for all working indicators

[tellaho]

Stacked on #1493 (tho/observer-owner-global), which stacks on #1492. Merge order: #1492 → #1493 → this.

Overview

Category: improvement
User Impact: "Agent is working" indicators (left-nav channel badges, composer activity bar, profile badges, activity panel, stop-turn) now agree with each other and appear reliably, instead of some surfaces showing an agent as working while others show nothing.

Problem: The app had two independent "working" signals: typing indicators (kind 20002, 8s TTL) drove the composer bar and the activity panel's stop-turn state, while observer-derived active turns (kind 24200) drove sidebar working badges and profile badges. Depending on build, harness, and timing, one pipe could be live while the other was silent — so a user could clearly see an agent working in chat while the sidebar badge, activity panel, or composer bar disagreed.

Solution: One reusable signal module, agentWorkingSignal, becomes the single source of truth: observer-derived turns are primary (they carry channel scope and a start anchor), bot typing is the fallback when no observer data exists for that scope. It supports both the channel-scoped rule and the all-channels rule, exposes reference-stable hooks, and is deliberately generalized so future surfaces (e.g. thread ingresses) consume the same hooks. Every current working-affordance surface is migrated onto it.

File changes

desktop/src/features/agents/agentWorkingSignal.ts
New module: typing registry (fed per-channel by the typing hooks), observer-primary merge logic, cached snapshot selectors, and useAgentWorking / useWorkingChannels / useChannelWorkingAgentPubkeys hooks.

desktop/src/features/agents/agentWorkingSignal.test.mjs
15 tests: scope rules, typing fallback, observer-over-typing precedence, TTL clearing, anchor stability across re-reports, channel merging, snapshot reference stability, notify dedupe, workspace reset.

desktop/src/features/channels/ui/useChannelActivityTyping.ts
Mirrors botTypingEntries into the signal via reportChannelBotTyping; entries follow the typing TTL because the effect re-reports on every change and clears on unmount.

desktop/src/features/sidebar/lib/useActiveWorkingChannelsById.ts
Sidebar working badges read useWorkingChannels() — badges now also appear for typing-only agents (previously invisible in the nav when observer data was absent).

desktop/src/features/channels/ui/ChannelPane.tsx
Composer activity bar reads the unified per-channel working set instead of typing-only pubkeys; drops the isWorking prop drilling into the session panel.

desktop/src/features/channels/ui/BotActivityBar.tsx
Prop renamed typingBotPubkeys → workingBotPubkeys; internals renamed to match — the bar now represents working agents, not just typing ones.

desktop/src/features/channels/ui/AgentSessionThreadPanel.tsx
Computes isWorking internally via useAgentWorking(agent.pubkey, channel?.id ?? null) — the panel's working/stop-turn state now clears on turn_completed instead of waiting out the typing TTL.

desktop/src/features/agents/ui/ManagedAgentRow.tsx, desktop/src/features/profile/ui/UserProfilePanelSections.tsx, desktop/src/features/profile/ui/UserProfilePopover.tsx
"Working in #channel" badges read useAgentWorking(...).channels (observer channels plus typing-only channels, deduped observer-first).

desktop/src/features/workspaces/useWorkspaceInit.ts
resetAgentWorkingSignal() added to resetWorkspaceState() per the workspace-switching singleton contract.

Reproduction Steps

1. Run the desktop app with a managed agent and give it a task in a channel.
2. Watch the sidebar working badge, composer activity bar, and activity panel header — all three now show working together and clear together when the turn completes.
3. To see the fallback direction: with an agent whose observer stream is absent (e.g. a harness without relay observer), the sidebar badge and activity panel still show working from the typing signal — previously only the composer bar did.
4. To see the primary direction: with an agent whose typing signal is absent, the composer bar still shows the agent working from observer turns — previously it stayed empty.
5. Automated: cd desktop && pnpm test (1573/1573, 15 new).

Note: ModelPicker intentionally stays on raw useActiveAgentTurns — it needs actual observer turn targets for control frames, not a display signal.

[wpfleger96]

🟡 [IMPORTANT · correctness] Thread-only bot typing marks the whole channel as working.

This new reportChannelBotTyping mirror effect builds botTypingPubkeyKey from the unfiltered botTypingEntries. The loop above (lines 82–89) splits agent-vs-human only — it never distinguishes thread-scoped from channel-scoped entries. ChannelPane.tsx:332 then reads this back via useChannelWorkingAgentPubkeys(activeChannel?.id) to drive the main composer activity bar.

Result: an agent typing only in an open thread lights the main channel composer bar and feeds false channel-level activity into sidebar/profile surfaces. Note the sibling thread path (ChannelPane.tsx threadComposerBotTypingPubkeys) correctly applies a threadHeadId filter — this new channel mirror was added without the equivalent scope guard.

Suggested fix — filter to channel-scoped entries before building the key:

const botTypingPubkeyKey = botTypingEntries
  .filter((entry) => entry.threadHeadId === null)
  .map((entry) => entry.pubkey.toLowerCase())
  .sort()
  .join(",");

Plus a regression test: "thread-only bot typing does not mark channel-level working."

Everything else here is solid — the observer-primary / typing-fallback merge is well-factored and covered by focused tests, and I found no subscription/memory leak across the 8 consumers.

[tellaho]

Updated by Ned 🤖 in f4456ff.