feat(scouts): real triage LLM call via scout-triage-system prompt

app/scouts/engine.py

@@ -28,6 +28,8 @@ from datetime import datetime, timedelta, timezone
 from sqlalchemy import select
 from sqlalchemy.exc import IntegrityError
 
+from app.core.langfuse_client import extract_usage, get_langfuse, get_prompt_or_fallback
+from app.core.llm import get_llm
 from app.db import async_session
 from app.models import CloudScoutConfig, ScoutTriageQueue
 from app.scouts.connectors.base import ItemContent, ItemRef, TriageVerdict
@@ -188,5 +190,81 @@ class ScoutEngine:
             await session.commit()
 
     async def _triage_llm(self, scout: CloudScoutConfig, content: ItemContent) -> TriageVerdict:
-        """Stub — real implementation in Task 24."""
-        raise NotImplementedError("Real triage LLM call lands in Task 24")
+        """Call the scout-triage-system Langfuse prompt to classify an item as relevant or spam.
+
+        Uses gpt-4o-mini with JSON mode. Wraps the LLM call in a Langfuse generation
+        observation when Langfuse is configured.
+        """
+        import json  # noqa: PLC0415
+
+        from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+
+        _TRIAGE_FALLBACK = (
+            "You are a triage classifier for an executive-assistant scout that watches a "
+            "{source_type} feed.\n"
+            'The scout\'s purpose is: "{scout_purpose}".\n\n'
+            "Given one item, decide whether it is RELEVANT (worth surfacing to the user as a "
+            "potential task / event / note / project) or SPAM (advertising, mass marketing, "
+            "phishing, bulk notifications with no actionable content).\n\n"
+            "Item:\n"
+            "  - Subject: {item_subject}\n"
+            "  - From:    {item_sender}\n"
+            "  - Body (truncated): {item_body_truncated_2k}\n\n"
+            'Return JSON only, matching this schema:\n'
+            '  {{"verdict": "relevant" | "spam", "reason": <short string>, "confidence": <0..1>}}\n\n'
+            "Be conservative on \"spam\" — if a message could plausibly be a personal/work "
+            "email, mark it relevant."
+        )
+
+        template, prompt_obj = get_prompt_or_fallback("scout-triage-system", _TRIAGE_FALLBACK)
+
+        body_trunc = (content.body_text or "")[:2000]
+        variables = dict(
+            source_type=scout.provider,
+            scout_purpose=scout.prompt_template or "",
+            item_subject=content.metadata.subject or "",
+            item_sender=content.metadata.sender or "",
+            item_body_truncated_2k=body_trunc,
+        )
+
+        if prompt_obj is not None:
+            try:
+                system_text = prompt_obj.compile(**variables)
+                if isinstance(system_text, list):
+                    system_text = "\n".join(
+                        m.get("content", "") for m in system_text if isinstance(m, dict)
+                    )
+            except Exception as exc:
+                logger.warning("scout triage: compile failed: %s", exc)
+                system_text = template.replace("{{source_type}}", variables["source_type"]) \
+                    .replace("{{scout_purpose}}", variables["scout_purpose"]) \
+                    .replace("{{item_subject}}", variables["item_subject"]) \
+                    .replace("{{item_sender}}", variables["item_sender"]) \
+                    .replace("{{item_body_truncated_2k}}", variables["item_body_truncated_2k"])
+        else:
+            system_text = template.format(**variables)
+
+        llm = get_llm(model="gpt-4o-mini", temperature=0)
+        llm_json = llm.bind(response_format={"type": "json_object"})  # type: ignore[attr-defined]
+
+        messages = [
+            SystemMessage(content=system_text),
+            HumanMessage(content="Classify this item."),
+        ]
+
+        lf = get_langfuse()
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="scout-triage",
+                model="gpt-4o-mini",
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm_json.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm_json.ainvoke(messages)
+
+        data = json.loads(response.content)
+        return TriageVerdict(**data)