Follow-up to #561 (F1 caches probes per authored run).
A body using K distinct (cli, source, model, execution) keys still synchronously blocks the shared worker event loop K times. One hung probe blocks for the model-readiness timeout: 60 s. That exceeds the 30 s lease and will still kill a concurrent step.
Move authored preflight probing off the event loop without changing the published synchronous preflight overload or duplicating authentication decisions. reviewed-plan.md recommends one pure generator for the adaptive identify/model/auth sequence with thin synchronous and asynchronous drivers. Keep flows check synchronous. Regression coverage should exercise distinct models and a slow probe while an existing worker lease heartbeats.
Follow-up to #561 (F1 caches probes per authored run).
A body using K distinct (cli, source, model, execution) keys still synchronously blocks the shared worker event loop K times. One hung probe blocks for the model-readiness timeout: 60 s. That exceeds the 30 s lease and will still kill a concurrent step.
Move authored preflight probing off the event loop without changing the published synchronous preflight overload or duplicating authentication decisions. reviewed-plan.md recommends one pure generator for the adaptive identify/model/auth sequence with thin synchronous and asynchronous drivers. Keep flows check synchronous. Regression coverage should exercise distinct models and a slow probe while an existing worker lease heartbeats.