[cboettig]

Problem

enable_thinking only reaches NRP models listed in thinking_models (as chat_template_kwargs). For OpenRouter ids the proxy drops it with an info log, so an app cannot turn reasoning down on an OpenRouter model.

This shows up in global-30x30. In session b9f2c027 (2026-10-16), MiMo V2.6 Flash took 188 s on the step after fit_sdm, producing 4,985 completion tokens. 4,759 of those were reasoning, before a single add_hex_tile_layer call. Other steps in the same session spent 1,059 of 1,258 and 1,058 of 1,321 tokens reasoning.

Change

• Provider dialect. New provider setting thinking_dialect; config.json sets "openrouter" on the OpenRouter provider. For a model not in thinking_models, the proxy then sends OpenRouter's top-level reasoning object.
• Effort levels. New request field reasoning_effort (none / minimal / low / medium / high, OpenAI-style). If both fields are sent, the level wins. This gives clients a level to expose, as harnesses usually do; the boolean alone can't express one.
• Translation. reasoning_payload() does the per-backend mapping:

Request	NRP thinking_models model	OpenRouter	Other
enable_thinking: false	{key: false}	reasoning: {enabled: false}	nothing
enable_thinking: true	{key: true}	reasoning: {enabled: true}	nothing
reasoning_effort: "none"	{key: false}	reasoning: {enabled: false}	nothing
reasoning_effort: "low" etc.	{key: true} (logged: no levels)	reasoning: {effort: "low"}	nothing
invalid level	ignored (logged); boolean still applies	same	same

• Logging. reasoning_effort is logged on the request record next to enable_thinking.
• Docs and tests. README "Thinking mode" section updated, CHANGELOG entry added. New tests cover the mapping for NRP, OpenRouter and a provider with no dialect. python test_logging.py passes the 45 tests before its __main__ block, including the three new ones.

Not verified yet

I could not probe OpenRouter directly; the key lives only in the cluster secret. After deploy, check the effect on MiMo Flash by sending the same request with "enable_thinking": false and with "reasoning_effort": "low", and comparing tokens.completion_tokens_details.reasoning_tokens in the response log. For models whose reasoning is mandatory, OpenRouter may still reason. Following the thinking_models precedent (#105), any model that ignores the setting should be recorded once probed.

Follow-ups

• geo-agent: expose an effort selector and per-model reasoning_effort defaults in config, sending the new field. To be filed as a separate geo-agent issue.
• global-30x30: once this is deployed, set a lower reasoning default for MiMo Flash.