Fix SemIfEngine.generate: sample from the low-level llama.cpp backend

semif_agent/engine.py

@@ -95,23 +95,54 @@ class SemIfEngine:
     def generate(
         self,
         messages: list[dict],
-        temperature: float = 0.2,
+        temperature: float = 0.0,
         max_tokens: int = 256,
     ) -> str:
         """Drive the pinned decision model in the normal way: text generation.
 
-        SemIf scoring reads option logits directly; this instead uses the
-        underlying llama.cpp chat-completion endpoint on the same loaded model,
-        e.g. for skill-tree authoring. Each call resets the KV cache by
-        default, so interleaving scoring and generation on one model is safe.
+        SemIf scoring reads option logits directly; this instead autoregressively
+        samples from the same llama.cpp context, e.g. for skill-tree authoring.
+        Generation decodes the chat template through the backend's low-level
+        context (there is no high-level chat-completion object on the CPU
+        backend), stopping at the tokenizer's eos token. The KV cache is cleared
+        at the start, so interleaving scoring and generation on one model is safe.
         """
         model, tokenizer, metadata = self._ensure_loaded()
         try:
-            reply = model.create_chat_completion(
-                messages=messages,
-                temperature=temperature,
-                max_tokens=max_tokens,
+            import numpy
+
+            engine = model.engine
+            prompt_ids = tokenizer.apply_chat_template(
+                messages,
+                tokenize=True,
+                add_generation_prompt=True,
+                enable_thinking=False,
             )
+            engine.clear()
+            logits = engine._decode(prompt_ids, 0, 0, True)
+            generated: list[int] = []
+            rng = numpy.random.default_rng()
+            for position in range(max_tokens):
+                token = _sample_token(logits, temperature, rng)
+                if token == tokenizer.eos_token_id:
+                    break
+                generated.append(token)
+                logits = engine._decode([token], len(prompt_ids) + position, 0, True)
+            return tokenizer.decode(generated).strip()
+        except EngineUnavailable:
+            raise
         except Exception as exc:
             raise EngineUnavailable(f"generation failed: {exc}") from exc
-        return reply["choices"][0]["message"]["content"].strip()
+
+
+def _sample_token(logits, temperature: float, rng) -> int:
+    """Pick the next token from next-position logits: greedy or temperature."""
+    import numpy
+
+    if temperature <= 0.0:
+        return int(numpy.argmax(logits))
+    scaled = numpy.asarray(logits, dtype=numpy.float64) / max(temperature, 1e-6)
+    scaled = scaled - scaled.max()
+    probabilities = numpy.exp(scaled)
+    probabilities /= probabilities.sum()
+    return int(rng.choice(probabilities.size, p=probabilities))

tests/test_skills.py

@@ -84,6 +84,12 @@ def test_generate_category_without_engine_raises():
         generate_category(engine, Request("anything"), {})
 
 
+def test_generate_without_engine_raises():
+    engine = SemIfEngine(EngineConfig())
+    with pytest.raises(EngineUnavailable):
+        engine.generate([{"role": "user", "content": "hi"}])
+
+
 def test_build_tree_includes_registry_stubs(tmp_path):
     registry = CategoryRegistry(str(tmp_path / "categories.json"))
     registry.register("delivery", "Track and manage package deliveries.")