OpenRouter provider hardcodes reasoning:{enabled:true} — wedges non-thinking models into token-repetition garbage on some backends

[Sunster1970]

providers/openrouter.py unconditionally sends:

body = {"reasoning": {"enabled": True, "max_tokens": max_tokens, "exclude": True}}

for every model. Reasoning-token mode is an endpoint-specific capability; for non-thinking models (e.g. deepseek-chat-v3-0324) support depends on which upstream provider OpenRouter routes to.

Measured effect (A/B, frozen identical prompt, temperature 0.1, 400 max_tokens):

Request shape	Backend A (DeepInfra)	Backend B (SiliconFlow)
WITHOUT reasoning field	22/22 clean	all clean
WITH reasoning field	25/26 garbage (~96% fail, token-repetition, finish_reason: length)	all clean

The garbage output is decoder collapse: repeated token ladders, foreign-script tokens not present anywhere in the prompt, whole output budget burned. Combined with the session_id sticky routing (which pins a session to one backend), an agent can be glued to a failing backend for hours — the operator sees intermittent multi-hour "the model went insane" episodes that are actually request-shape × backend interactions.

Fix we run locally: make the reasoning block opt-in per model/config (send it only when explicitly enabled; log the decision each call). One screen of code; happy to PR it. Also worth noting for operators: provider: {"order": [...], "allow_fallbacks": false} gives determinism when a specific backend is misbehaving.