[fylorn]

Fixes two cases where set.reasoning: off left the model reasoning, and bumps the version to 0.83.1.

Refused none made reasoning go up

When an upstream rejected the none written by off (models that accept only minimal/low/medium/high), the resend dropped the effort, so the model reasoned at its default (usually medium), and that was remembered for 7 days.

• New tw_dialect::think::none_fix(status, error_body) -> Option<EffortFix>: when none_rejection holds, Use(lowest level listed in the error), else Drop. none_rejection is unchanged.
• hop.rs effort_refused uses it.

off did nothing across a format conversion

Converted hops applied off to the intermediate representation by removing the reasoning config, so an Anthropic or Gemini client routed to an OpenAI Chat/Responses upstream sent no effort at all and the upstream reasoned at its default, while the attempt chain showed sent: null.

• translate::apply_set takes the target format and reports whether it turned reasoning off. After encoding, translate::reasoning_off writes none on the converted body for Chat (reasoning_effort) and Responses (reasoning.effort), via the new tw_dialect::params::write_reasoning_none, which also fills in the field when the converted body carries no effort (an adaptive Claude request without output_config.effort).
• Gemini targets: reasoning is marked disabled instead of removed, so the existing encoder rule writes thinkingBudget: 0 on Gemini 2.5 Flash; models that cannot turn thinking off (2.5 Pro, Gemini 3) still get no thinking config, and the conversion now reports reasoning as dropped. Smallest non-zero budgets (2.5 Pro) and Gemini 3 minimal levels are not modelled.
• Anthropic/Bedrock targets are unchanged (no thinking, no effort).
• Levels across a conversion keep going through the intermediate representation, so the encoder writes them in the upstream format (including requests with reasoning on and no effort, which a wire-level rewrite would skip).
• adjust_effort, the effort memory and the resend already read the outbound body, so AttemptView.effort and remembered fixes now reflect the none really sent. Same-format hops and their byte-preserving splice are untouched.

The control-plane protocol stays at 54; no API or TS type changes.

Tests

• think.rs: none_fix with a list containing minimal → minimal; low/medium/high → low; no list → Drop; not about none → None.
• params.rs: write_reasoning_none fills in and overwrites the OpenAI fields, leaves other formats alone.
• tests/effort.rs: refused none resent with low and remembered, Drop without a list; Anthropic client + off → Chat gets reasoning_effort: "none" (also adaptive without an effort, and a refused none there is resent with minimal); → Responses gets reasoning.effort: "none"; Chat client → Anthropic upstream has no thinking; → Gemini 2.5 Flash gets thinkingBudget: 0; a rule level across conversion is written in Chat, Responses and Anthropic format; each attempt-chain effort matches what was sent.

Docs (config.md, config.zh-CN.md) and release-notes/0.83.1.md describe the change.

🤖 Generated with Claude Code