feat(HEL-850): disable thinking for translate/polish via reasoning_effort none

LM Studio's OpenAI-compat endpoint silently ignores chat_template_kwargs
(probed live); reasoning_effort "none" verifiably removes reasoning tokens
on both streaming and blocking paths. Pass it through the existing
LLMClient extra_body channel for translate/polish first and correction
requests only; all other callers keep their default policy.

Co-authored-by: multica-agent <github@multica.ai>

calliope-backend/src/calliope/routers/prompt_tools.py

@@ -19,6 +19,10 @@ from calliope.config import settings
 router = APIRouter()
 TOTAL_TIMEOUT = 120
 CHUNK_TIMEOUT = 45
+# Verified against the assigned translate/polish LM Studio service: sending
+# reasoning_effort "none" removes reasoning tokens on both the streaming and
+# blocking paths, while chat_template_kwargs was silently ignored there.
+NO_THINKING_BODY = {"reasoning_effort": "none"}
 # Ordered, lossless syntax tokens. Ordinary words and bare numbers stay editable.
 # Tags/dialogue/parameters are opaque before looking for brackets inside them.
 _TECH = re.compile(
@@ -115,8 +119,14 @@ async def _disconnected(request: Request) -> None:
         await asyncio.sleep(0.05)
 
 
-async def _run(request: Request, client: LLMClient, messages: list[dict], temp: float) -> str:
-    upstream = asyncio.create_task(client.chat(messages, temperature=temp))
+async def _run(
+    request: Request,
+    client: LLMClient,
+    messages: list[dict],
+    temp: float,
+    extra_body: dict | None = None,
+) -> str:
+    upstream = asyncio.create_task(client.chat(messages, temperature=temp, extra_body=extra_body))
     disconnect = asyncio.create_task(_disconnected(request))
     try:
         done, _ = await asyncio.wait({upstream, disconnect}, return_when=asyncio.FIRST_COMPLETED)
@@ -192,7 +202,13 @@ async def transform(body: Transform, request: Request) -> dict:
                 timeout=CHUNK_TIMEOUT,
             )
             result = await asyncio.wait_for(
-                _run(request, client, messages, 0.2 if direction else 0.6),
+                _run(
+                    request,
+                    client,
+                    messages,
+                    0.2 if direction else 0.6,
+                    NO_THINKING_BODY,
+                ),
                 max(0.001, TOTAL_TIMEOUT - (time.monotonic() - started)),
             )
             if not result.strip():

calliope-backend/tests/test_prompt_tools.py

@@ -502,3 +502,48 @@ def test_polish_can_rephrase_within_group(client, provider, text, reply):
     response = call(client, text=text, action="polish")
     assert response.status_code == 200, response.text
     assert response.json()["result"] == reply.strip()
+
+
+def test_thinking_off_on_first_and_correction_requests(client, provider):
+    provider.replies = [
+        "broken",
+        lambda payload: payload["messages"][1]["content"].replace("霓虹灯", "neon lights"),
+    ]
+    response = call(client, text="(霓虹灯:1.2)")
+    assert response.status_code == 200, response.text
+    assert len(provider.calls) == 2
+    assert [call_["reasoning_effort"] for call_ in provider.calls] == ["none", "none"]
+    assert [call_["temperature"] for call_ in provider.calls] == [0.2, 0.2]
+
+
+def test_polish_thinking_off_on_first_and_correction_requests(client, provider):
+    provider.replies = ["broken", lambda payload: payload["messages"][1]["content"]]
+
+    response = client.post(
+        "/api/prompt-tools/transform", json={"action": "polish", "text": "a cat (neon:1.2)"}
+    )
+    assert response.status_code == 200, response.text
+    assert len(provider.calls) == 2
+    assert [call_["reasoning_effort"] for call_ in provider.calls] == ["none", "none"]
+    assert [call_["temperature"] for call_ in provider.calls] == [0.6, 0.6]
+
+
+def test_blocking_fallback_preserves_thinking_off(client, provider):
+    provider.blocking = True
+    provider.reply = lambda payload: payload["messages"][1]["content"].replace("猫", "cat")
+    assert call(client).status_code == 200
+    assert len(provider.calls) == 2 and not provider.calls[1].get("stream")
+    assert [call_["reasoning_effort"] for call_ in provider.calls] == ["none", "none"]
+
+
+@pytest.mark.asyncio
+async def test_plain_llm_call_not_forced_thinking_off(provider):
+    from calliope.agent.llm import LLMClient
+
+    client = LLMClient(base_url=f"http://127.0.0.1:{provider.server_port}/v1", model="translate")
+    try:
+        assert await client.chat([{"role": "user", "content": "hi"}], 0.7) == "a cat"
+    finally:
+        await client.close()
+    assert provider.calls
+    assert all("reasoning_effort" not in call_ for call_ in provider.calls)