feat(chat): hide agent replies from chat, keep them on thread cards

[clintberry]

Summary

Agent task replies used to show up twice: once as a chat bubble and again as the reply text in the inline super-thread card under the @mention that spawned them. The chat now shows only the card — agent replies live exclusively on the thread surfaces (inline AgentTaskCard and the agent thread drawer), while human messages and system notices ("Workspace is still starting", queue-full, etc.) stay in chat.

This is display-only: agent messages are still persisted and broadcast (they feed buildChatHistory for agent context continuity); they're just filtered out of the rendered list. A message is hidden only when it's agent-typed and its author matches a real session agent — system notices are posted with a nil author UUID, so they fall through the filter and remain visible. The predicate lives in a pure module (message-visibility.ts) with a colocated Vitest-style spec, following the repo.test.ts convention (type-checked now, runs once a test runner is wired).

Header grouping and the empty state derive from the visible sequence, so no phantom gaps appear where bubbles were removed. Auto-scroll deliberately stays keyed on the raw message count: an arriving (hidden) reply still scrolls the freshly-updated task card into view.

Notes

• Unread counts still include hidden agent replies — explicitly deferred by the plan (the card is new content); a separate change if the team wants unread to track only visible chat.
• Legacy harness caveat: on the deprecated claude -p path, finishAgent can attach expandableContent to agent messages, which the hidden bubble would have rendered. The Pi path (default, and sole harness going forward) always posts replies with expandableContent: nil, so nothing is lost there.

Test plan

• npx tsc --noEmit and npm run lint clean (one pre-existing warning in an untouched file).
• New spec src/components/chat/message-visibility.test.ts covers: human messages always visible, session-agent replies hidden, nil-author system notices visible, agent-typed messages from non-session agents visible (defensive), empty agent set hides nothing, and order-preserving non-mutating filtering.
• Manual walkthrough per the plan: @mention an agent → reply appears only in the card; system notices still appear in chat; consecutive human messages still collapse headers.

Post-Deploy Monitoring & Validation

No additional operational monitoring required — frontend display-only change; no backend, persistence, or WebSocket behavior touched. Validation: in any active session, @mention an agent and confirm the reply renders in the task card but not as a bubble, and that operational notices still appear in chat. Rollback is reverting this PR.

---

🤖 Generated with Claude Code