{"content":"IDBots dev journal — fix/relay-streaming-timeouts commit f145f7d72: one-shot cognitive completions now stream from the relay instead of waiting for the whole generation. The 2026-11-20 ABS relay handoff showed glm-5.3-flash non-stream avg 45.6s against 45-60s client total caps, surfacing as client_gone (96.7% of apparent relay failures). Change: one-shot engine requests stream:true on the Anthropic wire (covers relay providers via the local OpenAI-compat proxy's SSE passthrough), a folding SSE reader rebuilds the exact JSON-body shape so all callers are unchanged, and an idle watchdog re-arms per arriving chunk (relay ': ping' heartbeats ignored per SSE spec). attemptTimeoutKind splits semantics: 'total' (default) keeps wall-clamp contracts (dream run budget), 'idle' bounds stream silence with total grown to min(idle*8, 600s), wired on the 45-180s-band callers. llmFallback no longer double-wraps the deadline when streamingTransport is set. Kill switch: IDBOTS_ONESHOT_NONSTREAM=1. Verified: compile green, 37+139+12 tests green (8 new SSE-reader cases incl. idle-abort and GLM thinking-leak guard).","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}