[Carltk]

The problem

The generic dialect sends no template arguments. That is correct for a plain
OpenAI surface, but it leaves no way to turn thinking off on a server that
takes the switch as a top-level request field. Ollama is the common case: its
/v1/chat/completions route turns thinking off only with
reasoning_effort: "none". think: false is ignored on that route, and
chat_template_kwargs is never sent under generic.

So a thinking model behind Ollama answers every question with its first
thinking token. Measured against Ollama 0.30.10 serving qwen3.6:27b, one
Choice question, max_tokens: 1, top_logprobs: 5:

request	latency	top tokens
no extra field	22.7 s	Here -0.76, Thinking -1.06, The -1.74
think: false	0.28 s	Here -0.76, Thinking -1.06, The -1.74
reasoning_effort: "none"	0.34 s	C -0.64, A -0.82, D -3.79

The change

• endpoint.extra_body: dict in the recipe schema, defaulting to empty.
• OpenAICompatibleClient merges it into every request body. The client is the
  one path every request takes (readout, warm, probe, rerank, embeddings), so
  every caller gets the field and the guard with no extra wiring.
• Guard at the chokepoint: a key the readout already set (max_tokens,
  logprobs, logit_bias, ...) is refused with a BackendError naming the
  key, before anything is sent. It is never silently overridden.
• Hash stability: the field is left out of the dump when empty
  (exclude_if), so every existing recipe keeps its hash and every committed
  scorecard still resolves. A test pins two pre-change hashes (the
  fake-vllm fixture and the shipped gemma-4-e4b-it.llamacpp recipe) as
  literals recorded on the parent commit.
• Docs: SKILL.md (the endpoint field list, plus the Ollama recipe) and the
  ADR-0004 D2 field table.

Tests

The fake server gains thinking_off_field, a top-level request field that
turns thinking off the way Ollama's /v1 route does. Default None, so no
existing behaviour changes.

• test_generic_dialect_cannot_turn_thinking_off_without_extra_body: reproduces
  the failure (MissingLabelsError, because the answer labels are crowded out
  by the thinking opener).
• test_extra_body_reaches_every_request_on_the_generic_dialect: with the field,
  the readout gets the answer, every request carries it, and
  chat_template_kwargs is still absent.
• test_extra_body_cannot_override_a_readout_field: max_tokens is refused and
  nothing reaches the server.
• test_extra_body_is_part_of_the_recipe_and_its_hash
• test_a_recipe_without_extra_body_keeps_the_hash_it_had_before_the_field_existed

Before the implementation, 3 of these fail. The other 2 describe existing
behaviour and pass either way.

Evidence

• uv run pytest -m "not live and not encoder": 136 passed (131 before).
• ruff check and ruff format --check are clean.
• mypy jevify: 24 errors, the same 24 as on main. This change adds none.
• Live: jevify eval from this branch against Ollama 0.30.10 /
  qwen3.6:27b, using endpoint.extra_body: {reasoning_effort: none} on a
  23-case choice suite: accuracy 1.0, median ~600 ms per question over a LAN.
  That suite is private, so no scorecard is committed here.

Not exercised

• The probe path with extra_body set was run live only indirectly (the
  probe builds its client through the same build_backend). Its thinking check
  still reports "thinking state unknown" on generic, because it has no
  server-side template to inspect. Detecting a thinking opener in the readout
  on generic would be a separate change.
• rerank and embedding endpoints receive the field through the same client.
  No live test was run against either.
• No shipped Ollama recipe is added, since there's no committed scorecard to
  back one. Happy to add one if you want it.

[fidecastro]

Thank you for your contribution! Just merged it.

[Carltk]

No worries, thanks for the project .. works great over Qwen3.8 on my local machine to allow my personal assistant to triage messages I send while it's mid-task. It reads each one as redirect / refine / queue / none off a single token's logprobs, in about half a second, and nothing leaves the Server.

I used probe, eval and calibrate to choose the model and test it. Local Qwen3.8 27B got 22/23 on held-out cases, and its confidence was good enough without calibration.

The extra_body PR came out of this setup: Ollama's thinking models need reasoning_effort: "none" on /v1, or the one-token readout picks up the first token of the reasoning instead of the answer.

Really nice tool to build on.