Disable model reasoning for NPC turns; give structured output headroom

services/convsim-core/convsim_core/runtime/sidecar.py

@@ -265,12 +265,22 @@ async def start(
         log_fh.write(header)
         log_fh.flush()
 
+        # NPC turns need fast, in-character JSON — not visible chain-of-thought.
+        # Reasoning-family models (Qwen3, DeepSeek-R1 distils) otherwise spend
+        # hundreds of tokens thinking, blowing the turn token budget (truncated
+        # JSON → fallback utterance) and multiplying latency. LLAMA_ARG_REASONING
+        # is honoured by current llama-server builds and silently ignored by
+        # older ones, unlike the equivalent CLI flag which would fail startup.
+        env = dict(os.environ)
+        env.setdefault("LLAMA_ARG_REASONING", "off")
+
         try:
             self._process = await asyncio.create_subprocess_exec(
                 *cmd,
                 stdout=log_fh,
                 stderr=subprocess.STDOUT,
                 creationflags=CREATE_NO_WINDOW,
+                env=env,
             )
         except OSError as exc:
             self._state = SidecarState.CRASHED

services/convsim-core/convsim_core/services/turn_pipeline.py

@@ -463,6 +463,12 @@ async def process_turn(
         messages=messages,
         json_schema=NPC_TURN_OUTPUT_SCHEMA,
         scripted_turn_index=turn_number,
+        # Structured turn output (utterance + deltas + rubric observations)
+        # regularly lands in the 300-600 token range; the 512 default left no
+        # headroom and a truncated JSON object degrades to the fallback
+        # utterance. Reasoning is disabled at the engine level, so the extra
+        # budget costs nothing when unused.
+        max_tokens=1024,
     )
     logger.debug(
         "Calling runtime %s for session %s turn %d (estimated %d tokens, truncated=%s)",

services/convsim-core/tests/test_sidecar.py

@@ -1037,3 +1037,50 @@ async def test_structured_output_via_fake_server(tmp_path):
 
     finally:
         await sidecar.stop()
+
+
+@pytest.mark.asyncio
+async def test_start_disables_model_reasoning_via_env(tmp_path, monkeypatch):
+    """The spawned llama-server must run with LLAMA_ARG_REASONING=off.
+
+    Reasoning-family models (Qwen3, R1 distils) otherwise burn the turn token
+    budget on chain-of-thought: the structured turn JSON gets truncated, every
+    turn degrades to the fallback utterance, and latency multiplies. The env
+    var (unlike the CLI flag) is ignored by older llama-server binaries, so
+    user-installed engines keep starting.
+    """
+    from convsim_core.runtime.sidecar import LlamaCppSidecar
+
+    captured: dict = {}
+
+    async def _fake_exec(*args, **kwargs):
+        captured["env"] = kwargs.get("env")
+        raise OSError("stop after capture")
+
+    monkeypatch.setattr("asyncio.create_subprocess_exec", _fake_exec)
+    sidecar = LlamaCppSidecar(log_dir=str(tmp_path))
+    with pytest.raises(RuntimeError):
+        await sidecar.start("/tmp/model.gguf", executable="/tmp/llama-server")
+
+    assert captured["env"] is not None
+    assert captured["env"].get("LLAMA_ARG_REASONING") == "off"
+
+
+@pytest.mark.asyncio
+async def test_start_respects_existing_reasoning_env(tmp_path, monkeypatch):
+    """An explicit user override of LLAMA_ARG_REASONING wins over the default."""
+    from convsim_core.runtime.sidecar import LlamaCppSidecar
+
+    captured: dict = {}
+
+    async def _fake_exec(*args, **kwargs):
+        captured["env"] = kwargs.get("env")
+        raise OSError("stop after capture")
+
+    monkeypatch.setenv("LLAMA_ARG_REASONING", "on")
+    monkeypatch.setattr("asyncio.create_subprocess_exec", _fake_exec)
+    sidecar = LlamaCppSidecar(log_dir=str(tmp_path))
+    with pytest.raises(RuntimeError):
+        await sidecar.start("/tmp/model.gguf", executable="/tmp/llama-server")
+
+    assert captured["env"].get("LLAMA_ARG_REASONING") == "on"