feat(tui): enriched switchable agent list — status/elapsed/tokens + arrow-open (agent switcher PR1)

[andrei-hasna]

Agent switcher PR1 — enriched, switchable live agent list (in-thread "FleetView")

Codewith already has the agent switching engine (the /agent · /session picker in
multi_agents.rs / agent_navigation.rs, session_lifecycle::select_agent_thread,
AppEvent::SelectAgentThread, and the app-server attach/detach). This PR only enriches the
picker rows, adds arrow-open at an empty composer, and introduces a live per-thread metrics
cache. No new protocol / RPC.

Target UX (in-thread FleetView)

Picker rows now render:

<status-dot> <agent-name>  <task>   <elapsed> · ↓ <tokens> tokens

• ● filled for the selected/active thread, ○ hollow otherwise; terminal states use a distinct
  glyph, all colored from the existing CollabAgentStatus vocabulary:
  running ●/○ green · completed ✓ green · interrupted ("wrapped") ✓ yellow ·
  stopped ■ dim · errored ✗ red.
• First row is always main; remaining rows keep earliest-launched-first spawn order.
• At an empty composer (no overlay/modal), a bare ↓ or ← opens the picker. ↑/↓ select
  (existing ListSelectionView), Enter switches into that agent's live window (existing
  select_agent_thread), Esc returns to the prompt, and main + Enter goes back to main.
• Alt+Left / Alt+Right quick-cycle is unchanged.
• Footer hint copy updated (Enter now "to switch"; subtitle mentions switching into the live window).
• ↓ tokens is the thread's cumulative total_tokens.

Why the metrics cache lives in AgentNavigationState (not ChatWidget)

Switching into an agent rebuilds the ChatWidget from scratch (select_agent_thread →
replace_chat_widget). If elapsed/tokens/status lived on the widget they would reset every time
you glanced at a different agent. So a new AgentLiveMetrics { status, started_at, last_task_message, token_total } is kept per ThreadId in AgentNavigationState, in a separate map from the
picker-entry metadata so it survives upsert liveness refreshes and thread switches (only clear
and remove drop it). Mutators: note_turn_started / note_turn_completed / note_error /
note_status / note_token_usage, plus a metrics() getter.

One shared feed for main + every sub-agent

app_server_events::handle_server_notification_event folds each per-thread notification into the
cache before it forks to primary-vs-non-primary enqueue, so main and every sub-agent update
uniformly. The mapping mirrors core/src/agent/status.rs::agent_status_from_event, adapted to the
v2 app-server stream: TurnStarted → Running (+ started_at, + task from the turn's first user
input) · TurnCompleted → {Completed|Interrupted|Errored} (by turn status) · non-retrying
Error → Errored · ThreadClosed → Shutdown · ThreadStatusChanged → status ·
ThreadTokenUsageUpdated → token_total.

Rendering

• multi_agents::agent_picker_status_dot_spans(status, is_active) — status-driven ●/○/✓/✗/■ + color.
• multi_agents::agent_picker_row_status(is_closed, live_status) — resolves the row's status (live
  telemetry authoritative; a closed entry with a non-terminal live status downgrades to Shutdown).
• multi_agents::format_agent_picker_metrics(task, elapsed_secs, token_total) — builds the
  description column, reusing status_indicator_widget::fmt_elapsed_compact and
  status::format_tokens_compact.
• session_lifecycle::agent_picker_selection_items — Main hoisted first, status-driven dot,
  metrics description replacing the raw-UUID description (the UUID stays in search_value).

Files changed

• codex-rs/tui/src/app/agent_navigation.rs — AgentLiveMetrics + cache + mutators/getter + tests.
• codex-rs/tui/src/app/app_server_events.rs — shared metrics-feed hook + status mappers.
• codex-rs/tui/src/multi_agents.rs — status-dot / row-status / metrics formatters + tests + snapshot.
• codex-rs/tui/src/app/session_lifecycle.rs — enriched picker rows, Main-first, footer copy.
• codex-rs/tui/src/app/input.rs — empty-composer Down/Left → OpenAgentPicker (gated).
• codex-rs/tui/src/app.rs — new multi_agents fn imports.
• snapshots: new agent_picker_rows; updated agent_picker_selection_items_with_new_agent.

Verification (Blacksmith, warm box)

• cargo nextest run -p codex-tui: 3485 passed, 4 skipped, 0 failed (exit 0) — includes the
  new reducer / formatter / row-status tests and both picker snapshots.
• cargo clippy -p codex-tui --tests --all-targets -- -D warnings: clean (exit 0). The only
  note is a pre-existing transitive proc-macro-error2 future-incompat warning, unrelated.
• cargo fmt --all -- --config imports_granularity=Item: clean.
• No *.md changed, so prettier is a no-op here.

Out of scope — follow-up increments (PR2–6)

1. PR2 — persistent agent rail (always-visible side list, not just the modal picker).
2. PR3 — live-transcript fidelity when watching a sub-agent (streaming deltas parity).
3. PR4 — richer per-agent widget (model/effort, progress, last message preview).
4. PR5 — inline steering / stop from the list (send input, interrupt an agent).
5. PR6 — snapshot RPC (authoritative server-side agent fleet snapshot instead of client-folded
   metrics; also freezes elapsed at completion and sources the task more precisely).

Do not merge/tag — handing back for review.