AKS-Lab-GitHubCopilot agent-framework.instructions.md

Microsoft Agent Framework + GitHub Copilot SDK conventions

Imports

from agent_framework import AgentResponse, FunctionTool, tool
from agent_framework.github import GitHubCopilotAgent, GitHubCopilotOptions
from copilot.generated.rpc import PermissionDecisionApproveOnce
from copilot.session import PermissionRequestResult

from src.shared.copilot import build_copilot_client

For MCP servers (FastMCP, not agents):

from mcp.server.fastmcp import FastMCP

Model

Every agent in this repo uses GitHubCopilotAgent with GitHubCopilotOptions(model="gpt-5.5"). Enforced — do not introduce AzureOpenAIChatClient, OpenAIChatClient, GitHubCopilotChatClient, or ChatAgent(client=...). If a future lab needs a different model, add it via config but keep gpt-5.5 as the default.

Building an agent

agent.py exposes exactly one factory returning a _RunnableAgent adapter (so make_app only needs .run(message)):

def _approve_all(_req: object, _ctx: dict[str, str]) -> PermissionRequestResult:
    return PermissionDecisionApproveOnce()


async def build_agent(settings: Settings) -> _RunnableAgent:
    agent = GitHubCopilotAgent(
        instructions=SYSTEM_PROMPT,
        client=build_copilot_client(),
        name="inventory",
        description="ZavaShop inventory specialist (...).",
        tools=list(TOOLS),               # usually [] — MCP is wired below
        default_options=GitHubCopilotOptions(
            model=settings.copilot_model,
            timeout=settings.copilot_timeout_seconds,
            on_permission_request=_approve_all,
            mcp_servers={
                "inventory": {
                    "type": "http",
                    "url": settings.inventory_mcp_url,
                    "tools": ["*"],
                    "timeout": int(settings.copilot_timeout_seconds * 1000),
                },
            },
        ),
    )
    return _RunnableAgent(agent=agent)

client=build_copilot_client() is mandatory so the lab GITHUB_TOKEN is passed explicitly to the Copilot SDK. on_permission_request=_approve_all is also mandatory. Without it the SDK denies every tool call and the LLM silently reports "all specialists rejected the request". mcp_servers[*].timeout is milliseconds, unlike GitHubCopilotOptions.timeout, which is seconds.