upskill eval -m flag is swallowed by fast-agent's ignore_unknown_args=True

[nagarjunakanamarlapudi]

Bug Description

When running upskill eval ./skills/my-skill -m qwen3:4b -v, the -m (model) flag is not parsed correctly by the eval command. Instead, the model name (e.g., qwen3:4b) is passed as a chat prompt to the default fast-agent, resulting in a chat response instead of an evaluation.

Environment

• upskill version: 0.2.1
• • fast-agent-mcp version: 0.4.41+
• • • Python: 3.14
• • • • OS: macOS

Steps to Reproduce

1. Create a skill with test cases in skill_meta.json
2. 2. Run: upskill eval ./skills/my-skill -m qwen3:4b -v
3. 3. Observe that the model name is treated as a chat prompt

Expected Behavior

The -m flag should configure the model used for evaluation.

Actual Behavior

The model name is passed to the default agent as a chat message, resulting in output like:

|> default
qwen3:4b

|< default gpt-5-mini
Qwen3 is the latest iteration in the Qwen series...

Root Cause Analysis

In cli.py, the _fast_agent_context() function creates a FastAgent with ignore_unknown_args=True:

@asynccontextmanager
async def _fast_agent_context() -> AsyncIterator[object]:
    fast = FastAgent(
        "upskill",
        ignore_unknown_args=True,  # <-- This causes the issue
    )

This setting causes fast-agent to consume unknown CLI arguments (like -m qwen3:4b) instead of letting Click parse them for the eval command.

Workaround

Run upskill eval without the -m flag and configure the model via fastagent.config.yaml:

default_model: sonnet

Then run:

upskill eval ./skills/my-skill -v

Suggested Fix

The --model/-m and --base-url flags need to be properly parsed before fast-agent initialization, or fast-agent-mcp's argument parsing should be disabled for these specific CLI commands.