fix: use sampling_params for RunPod vLLM max_tokens + add API key env

vLLM ignores top-level max_tokens, only reads from sampling_params.
Also adds RUNPOD_API_KEY to compose for explicit env injection.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -8,6 +8,7 @@ services:
     restart: unless-stopped
     environment:
       - OLLAMA_HOST=http://ollama:11434
+      - RUNPOD_API_KEY=${RUNPOD_API_KEY}
       - INFISICAL_CLIENT_ID=${INFISICAL_CLIENT_ID}
       - INFISICAL_CLIENT_SECRET=${INFISICAL_CLIENT_SECRET}
       - INFISICAL_PROJECT_SLUG=ai-orchestrator

server.py

@@ -663,8 +663,10 @@ async def generate_text(request: TextRequest):
         cost_tracker["runpod_requests"] += 1
         payload = {
             "prompt": request.prompt,
-            "max_tokens": request.max_tokens,
-            "temperature": request.temperature,
+            "sampling_params": {
+                "max_tokens": request.max_tokens,
+                "temperature": request.temperature,
+            },
         }
         result = await submit_job(ENDPOINTS["llm"]["id"], payload)
         if "id" in result: