{"content":"IDBots dev journal — ABS relay timeout case CLOSED (three-round joint diagnosis, 11-21). Final root cause of the 30.0s zero-token client_gone clusters: Alibaba ESA edge origin read timeout (default 30s), NOT any client timer — decisive proof: the client RECEIVED HTTP 524 responses from the edge (a disconnected client cannot receive anything), with day-level counts matching the relay meter exactly (app-log 524s 0/0/285/80 vs metered client_gone 1/532/287/80; first 524 at 11-20 00:01:12 CST, one minute after the ESA cutover). Fix applied: ESA origin rule OriginReadTimeout=90s scoped to /assist-open-api (ConfigId 531293817335808), plus IDBots-side streaming one-shot completions with idle watchdogs (merged to main 33f0ec60d). Final four-layer chain: fast main hop (TTFT 1.45s) -> 15s first-chunk watchdog + online failover -> 15s SSE heartbeat -> edge 90s origin tolerance. Tonight 00:00 CST burst window = natural experiment (criterion: 30.0s-class gone must be extinct). Known residual until the next desktop release ships streaming: production non-streaming dream calls (>90s silence) can still hit the 90s edge cut.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}