fix(tui): use nonblocking send for ListSubAgents refresh events (#3802)

[Hmbown]

Closes #3802 (child of #3800).

Problem

High sub-agent fanout creates status storms that put pressure on the bounded engine op channel (32) and event channel (256). Because Op::ListSubAgents sends used .send().await, the engine or TUI event loop could stall when the receiver was not draining fast enough.

Change

Switched all ListSubAgents refresh-type send paths from blocking send().await to non-blocking try_send():

1. crates/tui/src/core/engine.rs — Op::ListSubAgents handler now uses tx_event.try_send(Event::AgentList{...}) with a tracing::debug! on channel-full drop.
2. crates/tui/src/core/engine/handle.rs — Added EngineHandle::try_send() non-blocking method that delegates to self.tx_op.try_send(op).
3. crates/tui/src/tui/ui.rs — Converted 3 engine_handle.send(Op::ListSubAgents).await call sites to engine_handle.try_send(Op::ListSubAgents).

Guardrails (per the issue)

• Critical events unchanged: approval prompts, tool completions/results, fatal errors, cancellation, and parent completion all remain on send().await.
• Only refresh-type events use try_send: ListSubAgents is a sidebar refresh that converges on the next drain cycle.
• No capacity increase: channel buffer sizes unchanged.

Tests

• engine_handle_try_send_does_not_block_when_op_channel_is_full
• list_subagents_event_try_send_does_not_block_when_event_channel_full

[claude]

Claude finished @Hmbown's task in 1m 23s —— View job

---

Code Review — PR #3813: fix(tui): use nonblocking send for ListSubAgents refresh events

Summary: Correct fix for the backpressure problem. The guardrail is respected — critical event paths remain on .send().await, only the refresh ListSubAgents ops switch to try_send. No logic errors or race conditions found. Two minor observations below.

---

Findings

1. Closed and Full are conflated in the debug log (engine.rs:1534–1538)

if let Err(_e) = self.tx_event.try_send(Event::AgentList { agents }) {
    tracing::debug!(
        "Event channel full; dropping ListSubAgents refresh (will retry next drain)"
    );
}

TrySendError has two variants: Full (expected under load) and Closed (receiver dropped — TUI is gone). Both hit the same message. A Closed error will repeat every drain cycle until the engine itself shuts down and is harder to diagnose from logs. Consider including _e in the message or branching on the variant:

match self.tx_event.try_send(Event::AgentList { agents }) {
    Ok(()) => {}
    Err(tokio::sync::mpsc::error::TrySendError::Full(_)) => {
        tracing::debug!("event channel full; dropping ListSubAgents refresh");
    }
    Err(tokio::sync::mpsc::error::TrySendError::Closed(_)) => {
        tracing::warn!("event channel closed; dropping ListSubAgents refresh");
    }
}

Fix this →

---

2. Second test uses #[tokio::test] but has no await points (tests.rs:6055–6075)

#[tokio::test]
async fn list_subagents_event_try_send_does_not_block_when_event_channel_full() {

This test is purely synchronous — try_send doesn't need a runtime. Using #[test] / fn is cheaper and clearer about the non-async nature. Not a blocker, but worth tidying.

---

What's correct

• All three ListSubAgents call sites in ui.rs are correctly converted; no other Op variants are touched.
• The EngineHandle::try_send wrapper is clean and appropriately thin.
• The first test (engine_handle_try_send_does_not_block_when_op_channel_is_full) correctly exercises the handle's method and uses a synchronous #[test] as it should.
• Critical paths (approval, tool completions, fatal errors, cancellation) are verified untouched by inspection of the diff.
• The one-refresh-per-drain-batch invariant (subagent_list_refresh_requested flag) is preserved under try_send.

---