`agents-cli eval generate` fails with ADK 2.x — two SDK incompatibilities

[Ivan2125]

What happened?

After upgrading from agents-cli 0.1.3 → 0.3.0 (via agents-cli scaffold upgrade), running agents-cli eval generate against a project that uses VertexAiSearchTool fails with two distinct errors, both caused by incompatibilities between the Vertex AI GenAI Eval SDK (vertexai._genai) introduced in 0.3.0 and ADK 2.x.

Steps to Reproduce

1. Start from an agents-cli 0.3.0 project using google-adk>=2.0.0 and a VertexAiSearchTool in any agent.
2. Run:

   agents-cli eval generate --dataset tests/eval/datasets/basic-dataset.json

What did you expect to happen?

Expected Behavior

All eval cases run successfully and a traces artifact is written.

---

Actual Behavior — two errors, 0/N cases succeed

Error 1 — NameError: name 'Optional' is not defined

AgentConfig._get_tool_declarations_from_agent() calls typing.get_type_hints() on every tool. For VertexAiSearchTool (a native ADK tool that exposes _get_declaration() directly) this fails because its underlying class uses from __future__ import annotations with unresolved forward references.

File "...vertexai/_genai/types/evals.py", line 100, in _get_tool_declarations_from_agent
    genai_types.FunctionDeclaration.from_callable_with_api_option(
File "...google/genai/types.py", line 4422, in from_callable_with_api_option
    annotation_under_future = typing.get_type_hints(callable)
NameError: name 'Optional' is not defined

Root cause: The Eval SDK tries to introspect native ADK tools (which have their own _get_declaration()) as plain Python callables. It should check for _get_declaration and skip get_type_hints in that case.

---

Error 2 — Invalid app name 'local agent run'

After patching Error 1, every case still fails because _evals_common.py hardcodes app_name = "local agent run" (with spaces) and passes it to ADK's Runner. ADK 2.x introduced strict name validation (^[a-zA-Z][a-zA-Z0-9_-]*$) that rejects spaces.

Value error, Invalid app name 'local agent run': must start with a letter
and can only consist of letters, digits, underscores, and hyphens.

Relevant lines in vertexai/_genai/_evals_common.py:

# line 2337
app_name = session_inputs.app_name or "local agent run"
# line 2340
app_name = "local agent run"

Root cause: The default app name in the Eval SDK pre-dates ADK 2.x's name constraints. Changing "local agent run" → "local_agent_run" would fix it.

Client information

Client Information

Run agents-cli info to output debugging information about your client.

CLI version: 0.3.0
OS info: Windows-11-10.0.26200-SP0
google-adk: 2.1.0
vertexai: 1.153.1

Command Output / Logs

No response

Anything else we need to know?

• Project was upgraded with agents-cli scaffold upgrade (0.1.3 → 0.3.0). Upgrade itself ran cleanly.
• Both errors affect every eval case — 0/N succeed.
• Workaround for Error 1: inject a static _get_declaration on the VertexAiSearchTool instance to skip introspection.
• Workaround for Error 2: monkey-patch validate_app_name in app/_setup.py to sanitize spaces.
• Both workarounds are fragile; upstream fixes in either the Eval SDK or ADK are the clean solution.