Block a user
Reasoning blocks repeat themselves: the repetition penalty sees a 64-token window inside a 28k think block
helexa-bench should measure only quiet neurons — and stamp every sample with the contention it saw
A restart collapses the advertised context window to a sixth, until real traffic warms the prefill EMA
The guard already exists — it is just defeated
Found while auditing undocumented operator toggles (#297).
crates/neuron/src/config.rs:
/// Cold-start prefill speed (tokens/sec)…
Document preserve_thinking for operators, and audit the ~31 other undocumented neuron toggles
grenade
created branch feat/stamp-preserve-thinking-state in helexa/helexa
2026-08-26 13:41:43 +00:00
grenade
created branch fix/296-one-step-one-assistant-turn in helexa/helexa
2026-08-26 12:52:04 +00:00
Deep-session stall: the model announces a tool call, then emits EOS instead of making it