- Expediere
- 28 septembrie 2026 la 10:06 UTC
- Autor
- Kamo
- Comite
- a1d2743
kamo-llm reads a prompt at about 110 tokens a second on k3m1 before it writes a word, so an E2 prompt past ~6k tokens spent E2's whole 60 s per-attempt bound (and a stream its 60 s first-event bound) before the answer began, and the walk moved on. An attempt on an IN_CLUSTER model now also gets the prompt's reading time at **************** (50: half the measured rate, for the chars/4 estimate on JSON tool schemas and a short queue). No attempt runs past the walk's own budget, maxAttempts x the per-attempt bound, so the callers' rule (master section 11 'facade timeouts', SP04-F2: E10 3 x 120 s + 30 s) still holds; the ParameterFallback retry, which used to get a fresh full bound, is held to it too. SP99-T12.