Managed gateway (provider "opencode") kills turns mid-stream: intermittent 503s + silent hangs with no client timeout

Summary

Over ~2 months of local logs, agent turns on the managed opencode provider (big-pickle) die mid-stream in two distinct modes. Evidence below is from ~/.local/share/opencode/log/opencode.log plus direct SQLite inspection of opencode.db (message/part/token tables).

Mode B - explicit 503

AI_APICallError: Service Unavailable at ERROR level, providerID=opencode modelID=big-pickle. Client auto-retries (~2s later); turn dies when consecutive attempts fail. Partial streamed text persists via deltas, but step-finish/usage parts are never written - user sees truncated output with no error surfaced in the UI.

Occurrences (UTC):

• 2026-08-11T08:24, 08:38 - ses_010269ddeffe5pMcJNWHi8HxVg
• 2026-08-21T21:52:20 - ses_fde976355ffeJ3zlIR4zXtArN2
• 2026-08-21T23:14:42, 23:39:08, 23:46:08 - ses_fd9a6264dffeYcBr4Swo110vQk
• 2026-08-21T23:25:09, 23:26:16 - ses_fd95cb1e1ffeD0SUa5R4z6iZ2O
• 2026-08-21T23:34:38 - ses_fd9577e34ffey9VkEl1l7aYrAP (subagent)

Mode A - silent hang (worse)

Stream opens (message=stream logged), zero parts persisted, no error line. Hangs indefinitely until user abort/retry, then logged as error=Aborted. No client-side stream timeout exists.

Example: message msg_02107484b00163Rm57QVpjKXu2, stream started 2026-08-20T21:15:20Z, marked Aborted 22:46:04Z - hung for 90 minutes.

Supporting data

• 36 dead turns (assistant messages with tokens.output=0, usage never persisted) cluster in time bursts across different sessions in the same window: Aug 10/11, Aug 18, Aug 20/21. Burst pattern across sessions = provider capacity episodes, not per-client state.
• A failed turn's immediate retry succeeded 2 minutes later at nearly identical context size (~107k -> ~108k ctx tokens) = transient upstream failure.
• Zero occurrences of ECONNRESET / fetch failed / socket hang up / ETIMEDOUT / EPIPE in the full log since June 24 = not local network.
• Deaths occur at context sizes ~100k-140k, but successful turns also occur at 139k+ = not a context-length threshold.
• Primary agent and subagent turns fail identically with the same error signature = not subagent orchestration.

Requests

1. Investigate gateway capacity during the burst windows above.
2. Add a client-side stream watchdog/timeout for Mode A silent hangs.
3. Surface swallowed stream errors in the desktop UI instead of silent truncation.

Happy to share forensics scripts and the dead-turn CSV if useful.