Copilot instructions

1---
2applyTo: "**/autogen/**/*.py,**/multi_agent/**/*.py"
3---
4 
5# Microsoft AutoGen Development Standards
6 
7## Agent Configuration
8 
9- Always set `max_consecutive_auto_reply` — unlimited auto-reply is a runaway cost risk
10- Define explicit `is_termination_msg` functions — never rely on agents deciding to stop
11- Set temperature to 0.0–0.3 for agents performing deterministic tasks (code, structured data)
12- Use `human_input_mode="NEVER"` in production; use `"ALWAYS"` or `"TERMINATE"` in interactive flows
13 
14## Code Execution
15 
16- Always use Docker for code execution in production: `code_execution_config={"use_docker": True}`
17- Never allow auto code execution without sandboxing
18- Limit working directory to a scoped temp directory
19- Validate all tool function signatures with Python type annotations
20 
21## GroupChat
22 
23- Define a clear `speaker_selection_method` — `"auto"` uses LLM routing; define custom logic for deterministic routing
24- Set `max_round` on all GroupChats — default is unlimited (cost risk)
25- Keep GroupChat agent count to the minimum needed — every extra agent adds routing complexity and cost
26 
27## Security
28 
29- Never hardcode API keys — read from environment variables only
30- Validate all external tool calls before execution
31- Log every agent message exchange with sender and token count for cost tracking
32 
33## LLM Config
34 
35```python
36config_list = autogen.config_list_from_json("OAI_CONFIG_LIST")
37llm_config = {
38    "config_list": config_list,
39    "temperature": 0.1,
40    "max_tokens": 2000,
41    "timeout": 60,
42}
43```
44 
45- Always set `timeout` to prevent blocking on slow model responses
46- Always set `max_tokens` — unbounded generation is a cost risk
47