fix(tui): name agents in the spawn tree and keep teammates past their turn

[ericleepi314]

Summary

This fixes two problems in the TUI's agents overlay (the "Spawn tree"). Both came up while running a math team of named, persistent teammates.

1. Rows didn't name the agent

Each row showed only the task description, so a team of named specialists looked like a list of anonymous tasks.

• Cause: the backend sends name, description and subagent_type as separate fields. The TUI lost the name in three places:
  • gatewayClient used name only as a fallback for the goal (goal: description || name). The Agent tool always requires a description, so the name never survived.
  • SubagentProgress had no field for the name or the type.
  • The overlay rendered only goal.
• Fix: the TUI now stores name and agentType.
  • Each row starts with the spawn's name, or else the agent type. The default types (general-purpose, worker, fork) stay hidden, following the reference's userFacingName.
  • The detail pane shows both.
  • Replayed trees keep the label.
• Backend: agent.py's per-message progress now sends the resolved agent_def.agent_type. It used to send the raw tool input, which was None when the call didn't specify a type. The sync end-of-run frame now sends the spawn's own name; it used to put the type there.

2. Teammates vanished after the turn that spawned them

• Cause: turnState.subagents is turn-scoped. startMessage() and idle() both clear it, and the idle layout depends on that. gatewayClient.seenSubagents is never cleared, though. So a teammate's later frames arrive as subagent.progress / subagent.complete, which only update existing rows and never create them. A persistent teammate therefore disappeared from the overlay while it kept working.
• Fix: a session-scoped roster, $sessionAgents, that every subagent event updates.
  • The overlay adds agents that are still running but that the current turn no longer lists (withCarriedAgents).
  • Turn state and the idle layout are unchanged.
  • Finished entries are pruned at each turn start.
  • The roster survives /clear, /new and resume, because those keep the same agent-server process and its teammates keep running. It is cleared only when the backend process exits (forgetSessionAgents()).
• Interrupted agents now end their row. gatewayClient published subagent.complete only for completed, failed and killed. An interrupted frame fell through, and the row stayed running forever; with the roster, it would also have stayed pinned in the overlay. The backend reports interrupted for a foreground subagent stopped with ESC or from the overlay (agent.py), a killed teammate (team_runtime.py), and an aborted workflow agent (workflow/runner.py). Any status other than running or queued now ends the run.

Before / after

The real built TUI was run in a PTY against a fake agent-server.

	installed build (before)	this branch
Rows	● Pin environment encode statement	● fl-formalizer · Pin environment encode statement
Idle after turn 1	frozen snapshot, stale counts	the live teammate, fresh counts
During turn 2	teammate gone	● Explore · Find cutoff… and ● fl-formalizer · …

Test plan

• vitest:
  • overlay rendering: identity, a carried teammate, no Ink error screen
  • the gateway passing name and type through
  • the gateway ending a run on any non-live status, including interrupted
  • the handler keeping identity across partial events
  • a full turn cycle: the teammate stays in the roster, while finished and interrupted agents are pruned; the roster survives fullReset() and is emptied by forgetSessionAgents()
  • the spawn-history disk read
  • unit tests for subagentIdentity, subagentTitle, subagentAgentLabel, withCarriedAgents and isSubagentAlive
• pytest: the running emit sends the resolved type and the name, and an unnamed spawn stays unnamed when it ends.
• Mutation-checked: removing each part of the fix makes its tests fail. That includes restoring the old completed/failed/killed gate and moving the roster clear back into fullReset().
• Full vitest: the same failures as main (7 pre-existing, plus 1 timing-sensitive test under load). Typecheck is clean. ESLint shows only pre-existing warnings, none on changed lines.
• Full pytest: 10,812 passed. The one local failure (test_system_prompt_full::…damped_wording…) also fails on main; it comes from the developer's memory text.
• Critic review loop.

Follow-ups (not in this PR)

• Overlay:
  • When idle, the live list shows only carried teammates; the last turn's finished rows stay one [ away in history.
  • Once a carried agent finishes, its final state isn't shown.
  • The roster isn't reconciled with delegation.status.
• Transcript rows for Agent calls still read Agent(<description>), not the agent's name.
• ui-web's labelOf still uses name only as a fallback.
• Background agents (both behaviors predate this PR):
  • One resumed after it finished keeps its terminal status (keepTerminalElseRunning).
  • One killed from the overlay reports killed. normalizeSubagentStatus doesn't recognize that status and falls back to completed, so the row shows ✓ instead of interrupted.

🤖 Generated with Claude Code

[github-actions]

Test Results

     5 files   1 021 suites   20m 38s ⏱️
16 083 tests 16 061 ✅ 22 💤 0 ❌
32 136 runs  32 065 ✅ 71 💤 0 ❌

Results for commit c3fcdb0.