Proven on the arc board runs: loop-trace over bridge-driven opencode cells shows 0 tool spans (1 llm + evaluator spans only), so any client idle-watchdog sees total silence during tool execution. On five-surface extreme leaves whose forge/pytest phases exceed 10 minutes, VB's stream idle-cap false-killed 9 cells across 5 leaves (idle-cap 600000ms exceeded (no event since last tool/text)), each scored on severed partial work until quarantined.
Ask: the opencode backend should forward tool start/finish (or at minimum a keepalive/progress frame) on the chat-completions SSE stream, so downstream idle detection measures agent liveness rather than transport chatter. Same gap also blinds trace analytics (0 tool spans in every bridge-driven cell trace).
Workaround in use: VB_STREAM_IDLE_MS=2700000 (45 min) on bench runs — blunt, and it delays true-stall detection by 4.5×.
Proven on the arc board runs: loop-trace over bridge-driven opencode cells shows 0 tool spans (1 llm + evaluator spans only), so any client idle-watchdog sees total silence during tool execution. On five-surface extreme leaves whose forge/pytest phases exceed 10 minutes, VB's stream idle-cap false-killed 9 cells across 5 leaves (
idle-cap 600000ms exceeded (no event since last tool/text)), each scored on severed partial work until quarantined.Ask: the opencode backend should forward tool start/finish (or at minimum a keepalive/progress frame) on the chat-completions SSE stream, so downstream idle detection measures agent liveness rather than transport chatter. Same gap also blinds trace analytics (0 tool spans in every bridge-driven cell trace).
Workaround in use:
VB_STREAM_IDLE_MS=2700000(45 min) on bench runs — blunt, and it delays true-stall detection by 4.5×.