fix(test): correct integration test for load_tools with agent injection

[cagataycali]

Description

Fixes a failing integration test test_load_tools_without_agent in the MCP client test suite.

Problem

The test expected that not explicitly passing the agent parameter would result in no agent being available:

result = agent.tool.mcp_client(
    action="load_tools",
    connection_id="test_connection"
    # Note: agent parameter is not provided  <-- Wrong assumption!
)
assert "agent instance is required" in result["content"][0]["text"]  # ❌ Fails

Actual Error: "Connection 'test_connection' not found"

Root Cause

When using agent.tool.mcp_client(...) (direct tool calling), the SDK automatically injects the agent into the tool invocation via invocation_state. This is expected behavior.

From sdk-python/src/strands/tools/executors/_executor.py:137-140:

invocation_state.update(
    {
        "agent": agent,  # <-- SDK always injects this
        "model": agent.model,
        ...
    }
)

Solution

Changed the test to verify connection validation instead (the actual error path when using the direct tool calling convention):

def test_load_tools_nonexistent_connection(self, agent):
    """Test load_tools action with nonexistent connection."""
    result = agent.tool.mcp_client(
        action="load_tools",
        connection_id="nonexistent_connection"
    )
    assert "Connection 'nonexistent_connection' not found" in result["content"][0]["text"]

Note

The unit test (tests/test_mcp_client.py:616) still properly covers the "no agent" case by calling mcp_client() directly (not through agent.tool.x).

Related Issues

• Investigation tracked in cagataycali/strands-coder#94

Type of Change

• Bug fix (non-breaking change which fixes an issue)

Testing

• Verified fix locally - test now passes
• All existing tests pass

Checklist

• I have read the CONTRIBUTING document
• I have added tests that prove my fix is effective
• My changes generate no new warnings

---

Automated by strands-coder 🦆

[cagataycali]

✅ Test Fix Ready for Review

This PR fixes an integration test for load_tools with agent injection and all CI checks are passing!

Status Summary

• ✅ CI: All checks SUCCESS
• ✅ Mergeable: No conflicts
• ✅ Test-only change

What This Fixes

Corrects the integration test for load_tools functionality when agents are injected, ensuring test accuracy.

Impact

• Test correctness improvement
• No production code changes
• Zero risk

Ready for maintainer review! 🦆

[mkmeral]

The failing test: test_load_tools_without_agent expected error "agent instance is required" but got "Connection 'test_connection' not found"

Why it ever passed: Test was written July 28, 2025 when SDK didn't auto-inject agent into invocation_state.

When it broke: November 24, 2025 - SDK PR #1213 (commit aaf9715) added automatic agent injection. Released in SDK v1.19.0 on Dec 3, 2025.

The breaking change was in sdk-python, not tools:

# src/strands/tools/executors/_executor.py
invocation_state.update({
+   "agent": agent,  # <-- Added in PR #1213
    "model": agent.model,
    ...
})