Both halves done in 00acd283, merged.
Part 1 — preserve_thinking is written up in 03-neurons.md under a
new Reasoning section: what it is (the model's own chat-template
control,…
Fixed in 43e873f8, deployed and verified (batch engine started max_slots=2 on beast). 00acd283 corrects the docs.
beast now runs max_in_flight = 2. Each slot gets ~59,000 tokens of…
Fixed in 2c09f871, merged.
The invented ladder is gone. Rungs are now probed from the model's own
chat template — for Qwen3.8-27B that is low, medium, xhigh, with
xhigh as the…
Fixed in b0da2d35, merged; 9fa3fcea handles the trend-line
discontinuity.
Each concurrent stream now carries its own principal identity, so per-principal fair-share stops treating a single…
cargo clippy --workspace without --all-targets misses #[cfg(test)]
cargo clippy --workspace without --all-targets misses #[cfg(test)]
Fixed in 03fe47e9, merged.
CI now runs cargo clippy --workspace --all-targets -- -D warnings, so
test and bench targets are linted at the same severity as the library.
The gap was not…
Fixed in 45e03893, merged.
Non-streaming completions now run through the batch engine, the same producer as the streaming path. Measured on beast:
Fixed in b403a510 and a1b4584f, both merged.
All three defects closed:
incomplete_details.reasonis emitted on the streaming payload and the non-streaming body, for aLength…
Fixed across all five surfaces — 039a353a and eb783c34.
Fixed in 82a7b842 — and the cause was ours, not the model's.
What was actually happening
Responses splits one agentic step into several input items (reasoning,
message, function_call)…