MCP resources are listed and read without the toolset's header_provider headers or user credential #7822

🔴 Required Information

Describe the Bug:
With McpToolset(use_mcp_resources=True), MCP resources are listed and read without the headers the toolset sends for tool calls. McpToolset._build_headers resolves header_provider and the per-user credential (credential_key) from the request context. LoadMcpResourceTool calls list_resources() and read_resource(name) without passing any context, so on any server that authenticates requests, resources fail while tools from the same toolset work.

The failure is only logged as a warning ("Failed to list MCP resources …" / "Failed to read MCP resource …"). The model is never told the resources exist, and load_mcp_resource calls return nothing.

Steps to Reproduce:

1. Run a streamable-HTTP MCP server that exposes a resource and returns 401 without Authorization: Bearer user-token.
2. Give an LlmAgent an McpToolset(connection_params=StreamableHTTPConnectionParams(url=...), header_provider=lambda ctx: {"Authorization": "Bearer user-token"}, use_mcp_resources=True).
3. Run a turn in which the model calls load_mcp_resource(resource_names=["handbook"]).

Expected Behavior:
The resource list goes into the instructions, and the resource content is attached after the call, using the same headers as the toolset's tools.

Observed Behavior:

Failed to list MCP resources: Failed to create MCP session: ... Server returned an error response
Failed to list MCP resources: Failed to create MCP session: ... Server returned an error response
Failed to read MCP resource 'handbook': Failed to create MCP session: ... Server returned an error response
resource list offered to the model: False
resource content attached: False
server-side 401s: 3

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at 58332a2)
• Desktop OS: macOS
• Python Version (python -V): 3.11.16

Model Information:

• Are you using LiteLLM: No
• Which model is being used: a scripted BaseLlm (the bug is in the tool, not the model)

🟡 Optional Information

Additional Context:
src/google/adk/tools/load_mcp_resource_tool.py, _append_resources_to_llm_request:

resource_names = await self._mcp_toolset.list_resources()
...
contents = await self._mcp_toolset.read_resource(resource_name)

Both methods accept readonly_context. McpToolset._build_headers only calls header_provider when one is passed, and _get_auth_headers uses it to look up credential_key. get_tools already passes it, which is why tools work.

Minimal Reproduction Code:

# Server: MCPServer with @server.resource("docs://handbook", name="handbook"),
# served by streamable_http_app() behind middleware that returns 401 without
# "Authorization: Bearer user-token".

toolset = McpToolset(
    connection_params=StreamableHTTPConnectionParams(url="http://127.0.0.1:8799/mcp"),
    header_provider=lambda ctx: {"Authorization": "Bearer user-token"},
    use_mcp_resources=True,
)
agent = LlmAgent(name="assistant", model=scripted_model, tools=[toolset])
# scripted_model calls load_mcp_resource(resource_names=["handbook"]) and then answers.

How often has this issue occurred?:

• Always (100%)