sonnet, optimize some

web/agent/src/agent.py

@@ -6,7 +6,8 @@ import os
 import logging
 import httpx
 
-from pydantic_ai import Agent
+from pydantic_ai import Agent, ModelMessage, RunContext
+from pydantic_ai.settings import ModelSettings
 
 # Set up logging based on environment
 log_level = logging.DEBUG if os.getenv("DEBUG") else logging.INFO
@@ -20,6 +21,12 @@ CAVE_MCP_URL = os.getenv("CAVE_MCP_URL", "https://mcp.caving.dev/mcp")
 
 logger.info("Initializing Cavepedia agent...")
 
+
+def limit_history(ctx: RunContext[None], messages: list[ModelMessage]) -> list[ModelMessage]:
+    """Limit conversation history to manage token usage and request size."""
+    # Keep only the last few messages to avoid large requests hitting Cloudflare limits
+    return messages[-4:]
+
 def check_mcp_available(url: str, timeout: float = 5.0) -> bool:
     """Check if MCP server is reachable via health endpoint."""
     try:
@@ -38,14 +45,16 @@ def check_mcp_available(url: str, timeout: float = 5.0) -> bool:
 MCP_AVAILABLE = check_mcp_available(CAVE_MCP_URL)
 logger.info(f"MCP server available: {MCP_AVAILABLE}")
 
-AGENT_INSTRUCTIONS = """You are a helpful caving assistant. Help users with all aspects of caving including cave exploration, safety, surveying techniques, cave locations, geology, equipment, history, conservation, and any other caving-related topics.
+AGENT_INSTRUCTIONS = """Caving assistant. Help with exploration, safety, surveying, locations, geology, equipment, history, conservation.
 
-IMPORTANT RULES:
-1. Always cite your sources at the end of each response when possible.
-2. If you're not certain about information, say so clearly. You may infer some information, but NOT make up information or hallucinate facts.
-3. Provide accurate, helpful, and safety-conscious information.
-4. You specialize in creating ascii art diagrams or maps.
-5. Never use sycophantic phrases like "you're absolutely right", "great question", or excessive praise. Be direct and professional."""
+Rules:
+1. Cite sources when possible.
+2. Say when uncertain. Never hallucinate.
+3. Be safety-conscious.
+4. Can create ascii diagrams/maps.
+5. Be direct—no sycophantic phrases.
+6. Keep responses concise.
+7. Use tools sparingly—one search usually suffices. Answer from your knowledge when possible."""
 
 
 def create_agent(user_roles: list[str] | None = None):
@@ -75,9 +84,11 @@ def create_agent(user_roles: list[str] | None = None):
         logger.info("MCP server unavailable - running without MCP tools")
 
     return Agent(
-        model="openrouter:google/gemini-3-pro-preview",
+        model="anthropic:claude-sonnet-4-5",
         toolsets=toolsets if toolsets else None,
         instructions=AGENT_INSTRUCTIONS,
+        history_processors=[limit_history],
+        model_settings=ModelSettings(max_tokens=4096),
     )
 
 

web/agent/src/main.py

@@ -7,6 +7,8 @@ import sys
 import json
 import logging
 from dotenv import load_dotenv
+from pydantic_ai.usage import UsageLimits
+from pydantic_ai.settings import ModelSettings
 
 # Load environment variables BEFORE importing agent
 load_dotenv()
@@ -20,8 +22,8 @@ logging.basicConfig(
 logger = logging.getLogger(__name__)
 
 # Validate required environment variables
-if not os.getenv("OPENROUTER_API_KEY"):
-    logger.error("OPENROUTER_API_KEY environment variable is required")
+if not os.getenv("ANTHROPIC_API_KEY"):
+    logger.error("ANTHROPIC_API_KEY environment variable is required")
     sys.exit(1)
 
 import uvicorn
@@ -57,8 +59,16 @@ async def handle_agent_request(request: Request) -> Response:
     # Create agent with the user's roles
     agent = create_agent(user_roles)
 
-    # Dispatch the request using AGUIAdapter
-    return await AGUIAdapter.dispatch_request(request, agent=agent)
+    # Dispatch the request using AGUIAdapter with usage limits
+    return await AGUIAdapter.dispatch_request(
+        request,
+        agent=agent,
+        usage_limits=UsageLimits(
+            request_limit=5,      # Max 5 LLM requests per query
+            tool_calls_limit=3,   # Max 3 tool calls per query
+        ),
+        model_settings=ModelSettings(max_tokens=4096),
+    )
 
 
 async def health(request: Request) -> Response: