UI Streaming and Rendering

How TUI and web render harness events, and how to debug “stuck” or garbled UI.

Stream Normalization

Core streaming sanitization removes common malformed glyph patterns from token deltas before UI rendering.

Event Ordering

High-throughput streams can interleave text and structural events (tool cards, traces, retries).

Mitigation:

• queue text/trace deltas briefly
• flush buffered deltas before structural events (tool_start, tool_result, etc.)
• maintain channel separation (user vs trace)

Web: SSE and busy state

Client: packages/web/client/useSSE.ts. Server: packages/web/server/sse.ts.

Transport vs API vs HUD “STANDBY”

Treat these separately:

Layer	Signal	Typical meaning
REST /api/status	busy, lastTurnEndedAt, clients	Express up; harness may be mid-turn even when SSE hiccups.
SSE /api/stream	connected, sseTransport, sseTransportDetail	Live event delivery; reconnects briefly during API restarts or web:dev proxy ECONNRESET.
HUD orb “STANDBY”	idle, not necessarily broken	Idle between turns — not diagnostics for crashes.

Client behaviour (web):

• CONNECTING / WAIT API appears before the first successful SSE handshake (sseHasOpenedOnce); REST may already be reachable.
• DEGRADED means /api/status OK while the EventSource layer is reconnecting / offline, including when the harness is busy (copy notes that events may pause).
• OFFLINE means /api/status could not be reached — kill :3001, wrong proxy target, VPN, firewall, etc.
• Busy stall watchdog: if busy and there are no streamed assistant/tool events (tokens, tool cards, subtask lines, etc.) for ~52s, a [UI] trace fires; it repeats about every 4 minutes while the stall continues. Periodic harness_running alone does not reset this timer — it only proves the SSE socket is warm. Use RAW for trace lines when AGENT_UI_VERBOSITY=quiet.
• web:dev (Vite on :5173 proxying to Express :3001) often logs http proxy error ... ECONNRESET when Express restarts; the UI probes /api/status on SSE interruption so SIGNAL does not jump straight to misleading OFFLINE when the API is still healthy.

turn_end and “stuck processing”

Symptom: trace shows [UI] Server reports idle — cleared stuck "processing" or the composer stays disabled after a long turn.

Causes:

• Very large RAW transcripts delaying turn_end delivery
• SSE reconnect missing a single turn_end event
• Heavy parallel tools (e.g. several web_fetch or hung read_artifact attempts)

Mitigations (client):

• Track expected turn_end after each user send
• Poll GET /api/status with consecutive idle samples + grace period before forcing idle
• Compare server lastTurnEndedAt (set on turn_end in agentBridge) before unlocking UI

Mitigations (operator):

• Rebuild core/tools and restart web after harness changes
• Avoid calling read_artifact when distill/elide are off — see Troubleshooting
• Cap parallel slow web_fetch — see Harness protocol — Web research

Session bootstrap

GET /api/config, POST /api/message, and POST /api/persona/bootstrap await whenSessionReady() so early requests do not race harness init. See Web API.

TUI Rendering Model

Reducer-driven message list, bounded window, stable modal heights — packages/tui/src/useAgent.ts.

Common Artifact Patterns

1. Channel bleed: trace/user text inside tool lines
2. Fragment splits: malformed glyph separators
3. Repaint flicker: frequent tiny updates
4. Stale tool cards: tool_start without matching tool_result after crash — refresh session

Debug flow

1. Confirm artifact origin (model text vs UI composition)
2. Inspect SSE event sequence around corruption (RAW toggle in web)
3. Verify turn_end and /api/status alignment
4. Test with AGENT_UI_VERBOSITY=quiet to reduce trace noise

API reference: Web API.