Files
ssmithxandClaude Opus 5.5 c48c88955a fix(routstr): serve a model that fits AIUI, and say why a peer refused
AIUI sends ~2.4k tokens of instructions with every message. TinyLlama is
trained on 2048 tokens, llama-server silently capped the configured 4096 to
that, and every AIUI request failed with Routstr's "Error forwarding request
to upstream" (the token was refunded each time).

- llama-server: 8192-token context for Qwen2.5-1.5B-Instruct (pinned
  revision and sha256 in the README); threads default to physical cores.
- archy-start.py re-states the model (name, context) on every start, keeping
  an existing price and on/off choice, so buyers see the real limit.
- Timeouts nest: Routstr upstream 240s < gateway 260s < buyer 280s < nginx
  300s. Routstr v0.5.0 keeps an X-Cashu token without refund when its caller
  hangs up, so the seller side must never be the one cut off.
- The buyer turns a peer refusal into a message naming the likely cause and
  what happened to the payment.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 13:39:56 +00:00
..