Send the room's reasoning setting as OpenRouter's unified parameter, with an explicit off value

[jurgenwerk]

Kimi K2.7 Code thought for ten minutes before writing anything, and setting its ModelConfiguration card to medium effort changed nothing. The room carried the setting and the bot sent it, as the OpenAI-style reasoning_effort field. OpenRouter lists reasoning and include_reasoning among the parameters that model accepts, not reasoning_effort, so the setting never reached the provider.

card.reasoningEffort ──► active-llm event ──► bot request
  ''  / null (unset)                          (no parameter: provider default)
  'none'                                      reasoning: { enabled: false }
  'low' | 'medium' | ...                      reasoning: { effort }

• The bot sends the unified reasoning parameter, which OpenRouter translates per provider: an effort level where the model grades its thinking, a budget where it budgets it, on or off where that is all it has.
• A card with no setting serializes as null, so every unset model already carried null and it could not mean "off". Null and undefined both leave the parameter out, which is what the old field did in effect.
• The field gains none, labelled "Off (no thinking)", as the explicit off switch, and the empty option is relabelled "Not specified". The bot accepts none alongside the levels.

The boxel-catalog side, the labels and the Kimi card set to none, is a separate PR there. A one-run smoke on Kimi after both land tells whether the provider honors the switch.

🤖 Generated with Claude Code

[github-actions]

Host Test Results

    1 files      1 suites   2h 7m 50s ⏱️
4 803 tests 4 788 ✅ 14 💤 0 ❌ 1 🔥
4 818 runs  4 802 ✅ 14 💤 1 ❌ 1 🔥

Results for commit 9662abb.

For more details on these errors, see this check.

Realm Server Test Results

    1 files    223 suites   1h 19m 14s ⏱️
3 002 tests 3 002 ✅ 0 💤 0 ❌
3 041 runs  3 041 ✅ 0 💤 0 ❌

Results for commit 9662abb.