feat(ML-3a): POST /emma + EMMA_BACKEND flagg (vertex nå → local VM ved trigger)

main.py

@@ -17,6 +17,7 @@ CG4-credits: GET /billing/credits — kreditt-saldo, burn rate og tom-dato.
              Krever GOOGLE_CREDIT_TOTAL_USD i Cloud Run env for runway-beregning.
 TERMINAL: POST /terminal/exec — whitelisted kommandoer: health, billing, build, logs, help. ✅
 VM: GET /vm/ssh-key — henter public key fra Compute Engine metadata. ✅
+ML-3a: POST /emma — intern agent (Emma Vauger). EMMA_BACKEND=vertex (nå) | local (VM-trigger). ✅
 """
 
 import os
@@ -64,6 +65,21 @@ PROJECT_ID = os.environ.get("GOOGLE_CLOUD_PROJECT", "propane-will-491900-m5")
 # Guard+-tiers som får tilgang til SMS-varsler
 SMS_ALLOWED_TIERS = {"guard", "shield", "enterprise"}
 
+# ── EMMA: backend-flagg ───────────────────────────────────────────────────────
+# EMMA_BACKEND=vertex  → Vertex AI (nå, bruker eksisterende kreditter)
+# EMMA_BACKEND=local   → lokal Gemma 4 på g2-standard-4 Spot VM (trigger: første kunde)
+# EMMA_LOCAL_URL       → http://<VM-IP>:8000/generate  (kun brukt når EMMA_BACKEND=local)
+EMMA_BACKEND   = os.environ.get("EMMA_BACKEND", "vertex")
+EMMA_LOCAL_URL = os.environ.get("EMMA_LOCAL_URL", "http://localhost:8000/generate")
+EMMA_MODEL     = os.environ.get("EMMA_MODEL", "google/gemma-3-12b-it")  # Vertex model ID
+
+# Last Emma sitt system-prompt fra world.md ved oppstart
+_WORLD_MD_PATH = pathlib.Path(__file__).parent / "docs" / "gemma" / "world.md"
+try:
+    _EMMA_SYSTEM_PROMPT = _WORLD_MD_PATH.read_text(encoding="utf-8")
+except FileNotFoundError:
+    _EMMA_SYSTEM_PROMPT = "Du er Emma Vauger, intern AI-assistent for Vauco AS. Vær analytisk og direkte."
+
 app = FastAPI(
     title="OSVauco OPAX Agent",
     description="Agent for OSVauco-OPAX platform.",
@@ -206,14 +222,6 @@ class TerminalExecRequest(BaseModel):
 
 @app.post("/terminal/exec")
 async def terminal_exec(req: TerminalExecRequest):
-    """
-    Whitelisted terminal-kommandoer:
-      health   → GET /health internt
-      billing  → GET /billing/summary
-      build    → GET /opax/build-status
-      logs     → siste 20 linjer fra Cloud Logging
-      help     → returner kommandoliste
-    """
     cmd = req.cmd.strip().lower().split()[0] if req.cmd.strip() else ""
     if cmd not in TERMINAL_WHITELIST:
         return JSONResponse(
@@ -301,10 +309,6 @@ async def terminal_exec(req: TerminalExecRequest):
 # ── VM: GET /vm/ssh-key ────────────────────────────────────────────────────────
 @app.get("/vm/ssh-key")
 async def vm_ssh_key(instance: str = "osvauco-dev-vm"):
-    """
-    Henter public SSH-nøkkel fra Compute Engine instance metadata.
-    Bruker google.auth ADC — krever compute.instances.get på service account.
-    """
     try:
         creds, project = google.auth.default()
         creds.refresh(google.auth.transport.requests.Request())
@@ -340,7 +344,7 @@ async def vm_ssh_key(instance: str = "osvauco-dev-vm"):
         raise HTTPException(status_code=500, detail=f"vm/ssh-key feilet: {e}")
 
 
-# ── OQ-29: GET /notify/channels — konfigurerte delivery-kanaler ───────────────
+# ── OQ-29: GET /notify/channels ───────────────────────────────────────────────
 @app.get("/notify/channels")
 async def get_notify_channels():
     return {
@@ -352,6 +356,104 @@ async def get_notify_channels():
     }
 
 
+# ── EMMA: ML-3a intern agent ──────────────────────────────────────────────────
+class EmmaRequest(BaseModel):
+    message: str
+    session_id: str = "emma-default"
+    history: Optional[List[Dict[str, str]]] = Field(
+        default=None,
+        description="Valgfri samtalehistorikk: [{role: user|model, content: str}]"
+    )
+
+
+class EmmaChatService:
+    """
+    Håndterer Emma-kall mot enten:
+      - Vertex AI (EMMA_BACKEND=vertex) — nå, bruker Vertex-kreditter
+      - Lokal Gemma 4 VM  (EMMA_BACKEND=local)  — når trigger er nådd
+    Bytt backend: sett EMMA_BACKEND=local + EMMA_LOCAL_URL=http://<VM-IP>:8000/generate
+    """
+
+    async def chat(self, message: str, history: Optional[List[Dict]] = None) -> str:
+        if EMMA_BACKEND == "local":
+            return await self._local(message, history)
+        return await self._vertex(message, history)
+
+    async def _vertex(self, message: str, history: Optional[List[Dict]] = None) -> str:
+        try:
+            import vertexai
+            from vertexai.generative_models import GenerativeModel, Content, Part
+
+            vertexai.init(project=PROJECT_ID, location="us-central1")
+            model = GenerativeModel(
+                EMMA_MODEL,
+                system_instruction=_EMMA_SYSTEM_PROMPT,
+            )
+
+            chat_history = []
+            for turn in (history or []):
+                role = turn.get("role", "user")
+                content = turn.get("content", "")
+                chat_history.append(Content(role=role, parts=[Part.from_text(content)]))
+
+            chat = model.start_chat(history=chat_history)
+            response = chat.send_message(message)
+            return response.text
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Emma (Vertex) feilet: {e}")
+
+    async def _local(self, message: str, history: Optional[List[Dict]] = None) -> str:
+        """
+        Kaller lokal Gemma 4 på VM via HTTP.
+        Forventet format: POST {EMMA_LOCAL_URL} med {prompt, system_prompt}
+        """
+        turns = ""
+        for turn in (history or []):
+            role = "User" if turn.get("role") == "user" else "Emma"
+            turns += f"{role}: {turn.get('content', '')}\n"
+        full_prompt = f"{turns}User: {message}\nEmma:"
+
+        try:
+            async with httpx.AsyncClient(timeout=60.0) as client:
+                resp = await client.post(
+                    EMMA_LOCAL_URL,
+                    json={"prompt": full_prompt, "system_prompt": _EMMA_SYSTEM_PROMPT, "max_tokens": 1024},
+                )
+            resp.raise_for_status()
+            data = resp.json()
+            return data.get("response") or data.get("text") or data.get("generated_text", "")
+        except httpx.TimeoutException:
+            raise HTTPException(status_code=504, detail="Emma (lokal VM) svarte ikke innen 60s")
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Emma (lokal) feilet: {e}")
+
+
+_emma = EmmaChatService()
+
+
+async def _emma_chat(request: Request, req: EmmaRequest):
+    """
+    POST /emma — intern agent, kun for Chris (IAP-beskyttet).
+    Backend: EMMA_BACKEND env-var (vertex | local).
+    """
+    start = time.monotonic()
+    response_text = await _emma.chat(req.message, req.history)
+    duration = round(time.monotonic() - start, 3)
+    store = get_store()
+    store.push("emma", "last_duration_s", duration)
+    store.push("emma", "last_backend", EMMA_BACKEND)
+    store.push("emma", "last_success", True)
+    return JSONResponse({
+        "response": response_text,
+        "backend":  EMMA_BACKEND,
+        "model":    EMMA_MODEL,
+        "duration_s": duration,
+        "session_id": req.session_id,
+    })
+
+app.add_api_route("/emma", endpoint=require_auth(_emma_chat), methods=["POST"])
+
+
 # ── MODELLER ──────────────────────────────────────────────────────────────────
 class RunRequest(BaseModel):
     message: str
@@ -756,13 +858,11 @@ async def authenticated_billing_anomalies(request: Request):
 app.add_api_route("/billing/anomalies", endpoint=require_auth(authenticated_billing_anomalies), methods=["GET"])
 
 
-# ── CG4-credits: GET /billing/credits ────────────────────────────────────────
 async def authenticated_billing_credits(request: Request, days: int = 90):
     try:
         return JSONResponse(BillingAgent().get_credits_status(days=days))
     except Exception as exc:
         return JSONResponse(status_code=500, content={"error": str(exc)})
-
 app.add_api_route("/billing/credits", endpoint=require_auth(authenticated_billing_credits), methods=["GET"])
 
 
@@ -886,7 +986,7 @@ def run_dag(req: DagRequest):
     store = get_store()
     log_dag_execution(dag_id=req.session_id, agent_results=results, total_duration_s=total_dur)
     store.push(AGENT_ID, "last_dag_total_duration_s", total_dur)
-    store.push(AGENT_ID, "last_dag_agent_count", len(agent_ids))
+    store.push(AGENT_ID, "last_dag_agent_count", len(tasks))
     store.push(AGENT_ID, "last_dag_success_count", sum(1 for r in results if r["success"]))
     return {
         "total_duration_s": total_dur,