Summary

Fixes #36777. After /compress forks the TUI session (ends parent, creates continuation), the gateway emits session.info with the new session_id — but the TUI handler was dropping the update on the floor, leaving ctx.sid pointing at the closed parent session. All subsequent RPCs (message submission, /image, /model, etc.) then targeted the dead session.

Root cause

Two interlocking bugs in ui-tui/src/app/createGatewayEventHandler.ts:

1. Guard discarded the event. The top-level guard early-returned any event whose ev.session_id differed from state.sid (except gateway.*). The post-compression session.info — which by definition carries a different sid — was silently discarded before any case ran.
2. Handler ignored sid. Even with the event delivered, the session.info case only patched info / status / usage — never sid.

The gateway side (tui_gateway/server.py::_sync_session_key_after_compress, _emit("session.info", new_sid, ...)) is correct; the bug was purely on the TS receiver.

Fix

• Allow session.info past the sid mismatch guard so the compression-fork event reaches the handler.
• In the session.info case, read ev.session_id (set by the gateway _emit envelope) and update state.sid when it differs.

case 'session.info': {
  const info = ev.payload
  const incomingSid = ev.session_id
  patchUiState(state => {
    const sidChanged = Boolean(incomingSid) && incomingSid !== state.sid
    return {
      ...state,
      info,
      sid: sidChanged ? (incomingSid as string) : state.sid,
      status: state.status === 'starting agent…' ? 'ready' : state.status,
      usage: info.usage ? { ...state.usage, ...info.usage } : state.usage
    }
  })
  ...
}

Testing

• Added two vitest cases in ui-tui/src/__tests__/createGatewayEventHandler.test.ts:
  • updates sid when ev.session_id differs (compression fork)
  • leaves sid unchanged when it matches or is absent
• npx vitest run src/__tests__/createGatewayEventHandler.test.ts → 52/52 pass
• npx tsc --noEmit → no new errors in touched files

Notes

• Issue reporter's suggested fix read info.session_id from the payload, but the gateway puts session_id on the event envelope (_emit line params = {"type": event, "session_id": sid}), not in the payload dict — so this PR uses ev.session_id, which is the source of truth and is what the existing guard also reads.
• Scope is intentionally minimal: only the handler + tests; no refactors.