Getting increase latency with MCPToolset after version 1.16.0

[hcadioli]

Describe the bug
After upgrading to ADK v1.16, requests that stream tool outputs show a noticeable latency increase. Tracing reveals that McpToolset.get_tools() is invoked multiple times during a single streamed interaction, repeatedly reconstructing the same tool list. This results in unnecessary network/IPC work and object creation, compounding latency.

To Reproduce
Steps to reproduce the behavior:

1. Install 'google-adk==1.16.*' (or build from main including v1.16 changes).
2. Configure an MCP server with a stable set of tools (no dynamic changes at runtime).
3. Run a streaming generation that calls MCP tools (e.g., tool-calling with partial deltas).
4. Observe logs/trace: get_tools() is called several times per request; end-to-end latency increases vs. pre-1.16.

Expected behavior

• get_tools() should only resolve/build the tool list once per McpToolset instance (or per session), and reuse it when the set of available tools is unchanged.
• Streaming should not incur repeated, identical tool discovery costs.

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. macOS, Linux, Windows]
• Python version(python -V):
• ADK version(pip show google-adk):

Model Information:

• Are you using LiteLLM: Yes/No
• Which model is being used(e.g. gemini-2.5-pro)

Additional context
Add any other context about the problem here.

[hcadioli]

possible fix?
#3236

[puritatemcordis]

can attest to the significant increased latency with ADK v1.16.0. We were previously using 1.14.1 and wanted to upgrade to 1.16 for the compaction + tool reflection plugin; however, after upgrading and moving it to production, we noticed all agent responses were taking exponentially longer to respond causing us to revert back to 1.14

[thiagosalvatore]

I can confirm it here. It is critical and it is impacting production. The stream is incredibly slow and it is overloading our MCP servers as it is calling it for every token being streamed

[puritatemcordis]

@hangfei @DeanChensj @Jacksunwei : could we get some visibility on this issue? There are multiple folks impacted preventing us from upgrading to the latest version of ADK

[Jacksunwei]

Root Cause

In v1.16.0, _maybe_add_grounding_metadata() in base_llm_flow.py calls agent.canonical_tools() 3-4 times per LLM response to check for google_search_agent. This causes McpToolset.get_tools() to be invoked 4-5x per response (1x during preprocessing + 3-4x for grounding checks) instead of just 1x.

Why Caching is Wrong

• get_tools(readonly_context) is designed for per-user permissions
• Different users should see different tools based on their authorization
• Naive caching at the toolset level breaks this security model

Recommended Fix

Cache tools per invocation context instead of per toolset:

1. Store fetched tools in invocation_context (same request = same user = safe to cache)
2. Reuse cached tools in _maybe_add_grounding_metadata()

Location: Fix should be in base_llm_flow.py, not mcp_toolset.py

Temporary Workaround

If your MCP toolset is static (same tools for all users, no dynamic tools), you can subclass McpToolset and cache get_tools() as a temporary workaround until the proper fix is released.

[Jacksunwei]

@xuanyang15 , could you take a look at this one?

[hcadioli]

hey @Jacksunwei thank you so much for the guidance here.

@xuanyang15, does this change look correct?
#3299