ADR: Model reasoning effort override

• Status: Proposed
• Date: 2026-10-23
• Related: issue #1821, issue #1701

Context

DeepSeek-family Chat Completions compatibility disables reasoning_effort for endpoints that reject that field. Model-name detection also covers custom relays that do support it, so choosing High can silently omit the parameter. Changing the automatic default would regress the endpoints protected by #1701.

Decision

Add optional supportsReasoningEffort: boolean | null to the persisted model binding. Missing/null retains automatic compatibility; true/false explicitly allows/disallows the parameter. Keep this provenance separate from catalog compatibility and apply it last, only when building a Chat Completions model. The setting does not turn reasoning on or override the selected thinking level.

Update (2026-10-23)

The automatic default for official deepseek.com endpoints changed after this ADR landed: those rows now leave supportsReasoningEffort to pi-ai detection, so the selected level reaches the wire by default (D672, #1821). The explicit per-model override described here is unchanged and still takes precedence in both directions.

Expose Auto / Send / Do not send in the existing per-model settings. Host normalization, configuration copying and synchronization preserve the value. Other APIs hide the setting and ignore it during request construction.

Consequences

Legacy JSON needs no migration: an absent field retains its prior behavior. Clearing a choice restores automatic compatibility. A user enabling Send is responsible for selecting a gateway that accepts the parameter; provider errors remain visible. No retry, silent protocol change or compatibility fallback is introduced. The alternative of globally re-enabling the field is rejected because it would undo the existing DeepSeek protection.