[fredtanxo]

Problem

Turning thinking off still leaves the model reasoning. The off path only sent parameters.enable_thinking: false, but the request kept model_config.is_reasoning: true (and chat_context.extra.modelConfig.is_reasoning), so Qoder's gateway still routed the model as reasoning-enabled.

@qoder-ai/qodercli disables thinking with three signals together (reverse-engineered from its published bundle):

t.parameters   = { ..., reasoning_effort: V, enable_thinking: V !== "none" }
t.model_config = { ..., is_reasoning: V !== "none" }

i.e. enable_thinking: false + reasoning_effort: "none" + is_reasoning: false.

Fix

• Normalize every host's "off" signal into one explicitlyOff flag: pi and OMP both resolve level "off" to reasoning: undefined; OMP's fork additionally sets disableReasoning / forceReasoningOff; false / "off" cover older hosts.
• Skip clamping when off, so an off request is never snapped back up to a concrete level (previously clampThinkingLevel could lift "off" to the model's lowest effort).
• Emit the three disable signals whenever no concrete level survives (reasoningLevel === undefined): enable_thinking: false, (reasoning_effort: "none", is_reasoning: false (both chat_context and model_config).

Tests

Three regression cases assert the request body when thinking is off: reasoning: 'off", disableReasoning: true (OMP), and omitted reasoning (pi's off signal).

Notes

• disableReasoning / forceReasoningOff are OMP-fork fields not present in @earendil-works/pi-ai's SimpleStreamOptions; they are read through a type assertion, same as the existing reasoning !== false && reasoning !== "off" guard.
• For models whose catalog lacks a disabled branch (Kimi-K3, GLM-5.3, Sonus, Cantus, ...), off now still sends reasoning_effort: "none". qodercli hides the "none" option for those models, so the gateway may accept or ignore it - but the request is at least self-consistent now, no longer "disable + is_reasoning: true".

[simonsmh]

Hi @fredtanxo — thanks for this PR! It has been merged as-is (squash: 255e07e). The diagnosis was spot on: enable_thinking: false alone really was contradicted by is_reasoning: true, and the clamp snapping "off" back up to a concrete level for models without a disabled branch was a nice catch.

Before cutting a release, I ran your change against the real Qoder gateway (a reasoning-bait prompt, thinking high as a control vs off). Sharing one edge case that turned up, since you flagged it yourself in the PR notes:

• Models with a disabled branch (e.g. Ultimate): perfect. The three signals go out together and the response comes back with zero reasoning content. ✅
• Models without one (e.g. Sonus): the upstream rejects reasoning_effort: "none" with a 400 — Unsupported value: 'none' is not supported with the 'gpt-6-astra-2026-09-03' model. Supported values are: 'low', 'medium', 'high', 'xhigh', and 'max'. Because the omitted-reasoning path lands in the same branch, those models would have failed on any request that doesn't carry a concrete level. (For the record, Kimi-K3 and GLM-5.3 accept "none" and do stop reasoning — it's specifically the strict-enum upstreams that error.)

I've landed a small follow-up on main (030a1fb) that keeps your behavior intact except for this case: when the catalog positively shows there is no disabled branch (an effort-based thinking_config without disabled, or thinkingLevelMap pinning off to null), the effort signal is omitted while enable_thinking: false and is_reasoning: false still go out. With no catalog signal either way, "none" is still sent, as in your PR. Verified live afterwards: Sonus no longer errors, Ultimate still fully disables. The trade-off is the same one qodercli makes — for those strict upstreams there is no real "off" to offer, so thinking-off stays best-effort there rather than breaking the request.

Two regression tests seeding the model cache for both catalog shapes were added alongside; the suite is at 367 passing and CI is green. This will ship in v0.6.3.

Thanks again for the careful reverse-engineering and the thorough PR description — it made the live verification straightforward.