diff --git a/SKILL.md b/SKILL.md
index e21ac3a..61a2794 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -53,12 +53,18 @@ Every recipe has `schema_version: 1`, `model: {name, kind, revision}` and
 the rest.
 
 **`endpoint`** (any OpenAI-compatible server): `endpoint: {base_url,
-api_key_env, dialect: auto|vllm|llamacpp|generic, timeout_s}`, `template`
-and `answers`.
+api_key_env, dialect: auto|vllm|llamacpp|generic, timeout_s, extra_body}`,
+`template` and `answers`.
 - `template.mode: messages` sends chat messages; put thinking off in
   `template.kwargs` (`{thinking: false}` for DeepSeek, `{enable_thinking:
   false}` for Qwen3-class models; names are model-specific, the probe checks
   that no `<think>` token opens the answer).
+- `endpoint.extra_body` is merged into the top level of every request, for
+  servers that take a switch there rather than as a template argument. The
+  `generic` dialect sends no template arguments, so on Ollama's `/v1` route
+  a thinking model needs `extra_body: {reasoning_effort: none}` (`think:
+  false` is ignored there). A key jevify already sets for the readout
+  (`max_tokens`, `logprobs`, ...) is refused, not overridden.
 - `template.mode: raw` sends a completion prompt you write with the model's
   own control tokens; use it when the chat template would not put the model
   in its trained regime (the Qwen3-VL reranker recipe is the example).
diff --git a/docs/adr/0004-recipes.md b/docs/adr/0004-recipes.md
index 8861210..ea0013d 100644
--- a/docs/adr/0004-recipes.md
+++ b/docs/adr/0004-recipes.md
@@ -32,6 +32,7 @@ hash is the recipe's identity in every answer and every scorecard.
 | Section | Holds |
 |---|---|
 | `model` | Identifier as the backend knows it, revision when known, the adapter kind, and for endpoints the dialect the probe detected. |
+| `endpoint` | For endpoint-backed kinds: base URL, the environment variable holding an API key, the dialect, the timeout, and `extra_body`, server-specific fields merged into the top level of every request (never replacing a field the readout sets; omitted from the hash when empty, so older recipes keep theirs). |
 | `template` | For endpoints: how state, instructions and options are rendered into messages or a raw prompt; where the question goes; the answer-slot suffix; template arguments such as thinking off; whether assistant prefill is used. For encoders: the sequence layout and marker scheme. |
 | `answers` | The answer tokens for `noul`, the identifier scheme for `choice`, the level tokens for `score`, each with the token ids the probe verified. |
 | `readout` | The rung of the ladder to use or `auto`; the permutation policy and call count for `choice`. |
diff --git a/jevify/adapters/endpoint/client.py b/jevify/adapters/endpoint/client.py
index 39927f2..f6a0b3d 100644
--- a/jevify/adapters/endpoint/client.py
+++ b/jevify/adapters/endpoint/client.py
@@ -5,6 +5,7 @@
 import asyncio
 import random
 import re
+from collections.abc import Mapping
 from typing import Any
 
 import httpx
@@ -28,6 +29,7 @@ def __init__(
         timeout_s: float = 120.0,
         http: httpx.AsyncClient | None = None,
         max_retries: int = 3,
+        extra_body: Mapping[str, Any] | None = None,
     ) -> None:
         stripped = base_url.rstrip("/")
         self.root = stripped[: -len("/v1")] if stripped.endswith("/v1") else stripped
@@ -35,6 +37,7 @@ def __init__(
         self.api_key = api_key
         self.timeout_s = timeout_s
         self.max_retries = max_retries
+        self.extra_body = dict(extra_body or {})
         self._http = http or httpx.AsyncClient(timeout=httpx.Timeout(timeout_s))
 
     def _headers(self) -> dict[str, str]:
@@ -44,6 +47,8 @@ def _headers(self) -> dict[str, str]:
         return headers
 
     async def _request(self, method: str, url: str, body: dict[str, Any] | None = None) -> Any:
+        if body is not None and self.extra_body:
+            body = _merge_extra(body, self.extra_body)
         attempt = 0
         while True:
             try:
@@ -76,6 +81,16 @@ async def aclose(self) -> None:
         await self._http.aclose()
 
 
+def _merge_extra(body: dict[str, Any], extra: dict[str, Any]) -> dict[str, Any]:
+    """Add the recipe's extra fields; never let one replace a field the readout set."""
+    clash = sorted(set(body) & set(extra))
+    if clash:
+        raise BackendError(
+            f"endpoint.extra_body may not set {', '.join(clash)}: jevify sets it for the readout"
+        )
+    return {**body, **extra}
+
+
 def _retry_delay(response: httpx.Response, attempt: int) -> float:
     header = response.headers.get("retry-after-ms") or response.headers.get("retry-after")
     if header:
diff --git a/jevify/compose.py b/jevify/compose.py
index cdb53dd..bb065f4 100644
--- a/jevify/compose.py
+++ b/jevify/compose.py
@@ -37,6 +37,7 @@ def build_backend(recipe: Recipe, *, http: Any | None = None) -> Backend:
             recipe.endpoint.base_url,
             api_key=api_key,
             timeout_s=recipe.endpoint.timeout_s,
+            extra_body=recipe.endpoint.extra_body,
             http=http,
         )
         return EndpointBackend(client, recipe)
@@ -51,6 +52,7 @@ def build_backend(recipe: Recipe, *, http: Any | None = None) -> Backend:
             recipe.endpoint.base_url,
             api_key=api_key,
             timeout_s=recipe.endpoint.timeout_s,
+            extra_body=recipe.endpoint.extra_body,
             http=http,
         )
         return RerankBackend(client, recipe)
@@ -66,6 +68,7 @@ def build_backend(recipe: Recipe, *, http: Any | None = None) -> Backend:
                 recipe.endpoint.base_url,
                 api_key=api_key,
                 timeout_s=recipe.endpoint.timeout_s,
+                extra_body=recipe.endpoint.extra_body,
                 http=http,
             )
             from jevify.adapters.embedding.adapter import EndpointEmbedder
diff --git a/jevify/recipes/schema.py b/jevify/recipes/schema.py
index 1c8a087..0775216 100644
--- a/jevify/recipes/schema.py
+++ b/jevify/recipes/schema.py
@@ -32,6 +32,10 @@ class EndpointSpec(_Strict):
     api_key_env: str | None = None
     dialect: Dialect = "auto"
     timeout_s: float = Field(default=120.0, gt=0)
+    # Server-specific request fields sent at the top level of every request, e.g.
+    # Ollama's `reasoning_effort: "none"`, the only way its /v1 route turns thinking
+    # off. Left out of the dump when empty so recipes that predate it keep their hash.
+    extra_body: dict[str, Any] = Field(default_factory=dict, exclude_if=lambda v: not v)
 
 
 class QuestionTemplates(_Strict):
diff --git a/tests/fakes/fake_openai_server.py b/tests/fakes/fake_openai_server.py
index 1ac31a5..a7e409b 100644
--- a/tests/fakes/fake_openai_server.py
+++ b/tests/fakes/fake_openai_server.py
@@ -34,6 +34,9 @@ class Behaviour:
     supports_tokenize: bool = True
     thinking_kwarg: str | None = "thinking"
     thinking_default_on: bool = True
+    # a top-level request field that turns thinking off, as Ollama's /v1 route takes
+    # `reasoning_effort: "none"` and ignores template arguments
+    thinking_off_field: tuple[str, Any] | None = None
     # how the template spells "thinking off": nothing, or a closed empty block (Qwen3.5)
     thinking_off_style: str = "omit"
     cache_tail_tokens: int = 0  # trailing prompt tokens never reported cached (llama.cpp)
@@ -227,6 +230,9 @@ def _rendered_chat(self, body: dict[str, Any]) -> str:
         thinking_on = self.behaviour.thinking_default_on
         if self.behaviour.thinking_kwarg and kwargs.get(self.behaviour.thinking_kwarg) is False:
             thinking_on = False
+        off_field = self.behaviour.thinking_off_field
+        if off_field and body.get(off_field[0]) == off_field[1]:
+            thinking_on = False
         last = body["messages"][-1]
         if last["role"] == "assistant" and body.get("continue_final_message"):
             if self.behaviour.prefill == "eos_appended":
diff --git a/tests/test_endpoint_adapter.py b/tests/test_endpoint_adapter.py
index d6649ba..bd943a2 100644
--- a/tests/test_endpoint_adapter.py
+++ b/tests/test_endpoint_adapter.py
@@ -8,6 +8,7 @@
 import httpx
 import pytest
 
+from jevify.adapters.endpoint.readout import MissingLabelsError
 from jevify.compose import build_backend
 from jevify.domain.distribution import Distribution
 from jevify.domain.questions import ChoiceQuestion, NoulQuestion, Option, State
@@ -189,3 +190,53 @@ def test_per_option_choice_composes_yes_no_readouts(fixtures, fake_server_factor
     assert d.as_mapping()["billing"] == pytest.approx(math.exp(4) / (1 + math.exp(4)))
     assert answer.off_menu_mass is None
     assert answer.rung == "top_k"
+
+
+def _with_endpoint(recipe, **update):
+    return recipe.model_copy(update={"endpoint": recipe.endpoint.model_copy(update=update)})
+
+
+OLLAMA_LIKE = Behaviour(
+    scores={" yes": 0.0, " no": -1.0},
+    dialect="generic",
+    thinking_kwarg=None,
+    thinking_off_field=("reasoning_effort", "none"),
+)
+
+
+def test_generic_dialect_cannot_turn_thinking_off_without_extra_body(fixtures, fake_server_factory):
+    # The failure this field exists for: template arguments never reach a generic
+    # server, so a thinking model's one-token readout is its first thinking token.
+    server, http = fake_server_factory(OLLAMA_LIKE)
+    recipe = load_recipe(fixtures / "recipes" / "fake-vllm.yaml")
+    backend = build_backend(_with_endpoint(recipe, dialect="generic"), http=http)
+
+    with pytest.raises(MissingLabelsError):
+        run(backend.evaluate(run(backend.warm(STATE)), [NoulQuestion(id="q", instructions="x")]))
+
+
+def test_extra_body_reaches_every_request_on_the_generic_dialect(fixtures, fake_server_factory):
+    server, http = fake_server_factory(OLLAMA_LIKE)
+    recipe = load_recipe(fixtures / "recipes" / "fake-vllm.yaml")
+    recipe = _with_endpoint(recipe, dialect="generic", extra_body={"reasoning_effort": "none"})
+    backend = build_backend(recipe, http=http)
+
+    [answer] = run(
+        backend.evaluate(run(backend.warm(STATE)), [NoulQuestion(id="q", instructions="x")])
+    )
+
+    assert Distribution.from_logprobs(answer.logprobs).labels == ("false", "true")
+    assert len(server.requests) == 2  # warm and question
+    assert all(req["reasoning_effort"] == "none" for req in server.requests)
+    assert all("chat_template_kwargs" not in req for req in server.requests)
+
+
+def test_extra_body_cannot_override_a_readout_field(fixtures, fake_server_factory):
+    server, http = fake_server_factory(Behaviour(scores={" yes": 0.0, " no": -1.0}))
+    recipe = load_recipe(fixtures / "recipes" / "fake-vllm.yaml")
+    recipe = _with_endpoint(recipe, extra_body={"max_tokens": 64})
+    backend = build_backend(recipe, http=http)
+
+    with pytest.raises(BackendError, match="max_tokens"):
+        run(backend.warm(STATE))
+    assert server.requests == []  # refused before anything was sent
diff --git a/tests/test_recipes.py b/tests/test_recipes.py
index 79a20be..3aa13b0 100644
--- a/tests/test_recipes.py
+++ b/tests/test_recipes.py
@@ -181,3 +181,28 @@ def test_render_identifier_choice_is_a_single_part() -> None:
     assert rendered.composition == "single"
     assert len(rendered.parts) == 1
     assert rendered.parts[0].labels == ("x", "y")
+
+
+def test_extra_body_is_part_of_the_recipe_and_its_hash(tmp_path: Path) -> None:
+    doc = yaml.safe_load((FIXTURES / "fake-vllm.yaml").read_text())
+    plain = tmp_path / "plain.yaml"
+    plain.write_text(yaml.safe_dump(doc))
+    doc["endpoint"]["extra_body"] = {"reasoning_effort": "none"}
+    extra = tmp_path / "extra.yaml"
+    extra.write_text(yaml.safe_dump(doc))
+
+    assert load_recipe(plain).endpoint.extra_body == {}
+    assert load_recipe(extra).endpoint.extra_body == {"reasoning_effort": "none"}
+    assert recipe_hash(load_recipe(plain)) != recipe_hash(load_recipe(extra))
+
+
+def test_a_recipe_without_extra_body_keeps_the_hash_it_had_before_the_field_existed() -> None:
+    # Scorecards cite recipe hashes; a new optional field must not re-hash old recipes.
+    # Literals recorded on the parent commit, before `endpoint.extra_body` was added.
+    shipped = Path(__file__).parent.parent / "recipes" / "gemma-4-e4b-it.llamacpp.yaml"
+    assert recipe_hash(load_recipe(FIXTURES / "fake-vllm.yaml")) == (
+        "134663112a13f27e6869011075c4eb14dabd9bb8fa8321a883731f42db0bde0e"
+    )
+    assert recipe_hash(load_recipe(shipped)) == (
+        "22181ad02a27f7c7958d40132340cef25e7e9b0c1d32a068b330745a16054bad"
+    )