feat: forward model config env vars to sandbox in execute_skillsvars to sandbox in execute_skills

[tinawenqiao]

PR: feat: forward model config env vars to sandbox in execute_skills

Problem

When execute_skills invokes the remote AIO sandbox, it only passes these environment variables:

env_vars = {
    "TOS_SKILLS_DIR": ...,
    "SKILL_SPACE_ID": ...,
    "TOOL_USER_SESSION_ID": ...,
    "PYTHONPATH": ...,
}

The sandbox then falls back to the default model baked into its Docker image (e.g., doubao-seed-2-0-lite), completely ignoring any model configuration set in the local .env or agentkit.yaml.

This means users who configure a custom model (e.g., OpenAI, Anthropic, or a third-party proxy) via MODEL_AGENT_NAME / MODEL_AGENT_PROVIDER / MODEL_AGENT_API_BASE / MODEL_AGENT_API_KEY cannot use that model inside the sandbox.

Scenario

User configures .env:

MODEL_AGENT_NAME=gpt-4o-mini
MODEL_AGENT_PROVIDER=openai
MODEL_AGENT_API_BASE=https://proxy.example.com/v1
MODEL_AGENT_API_KEY=sk-xxxxx

Local Agent works fine with gpt-4o-mini. But when execute_skills invokes the sandbox, the sandbox's internal agent still uses doubao-seed-2-0-lite because these env vars are not forwarded.

Solution

Forward model-related environment variables to the sandbox via env_vars, so the sandbox's internal agent respects the same model configuration as the local agent.

The sandbox's VeADKConfig (in config.py) already reads these env vars at startup — we just need to pass them through.

Changes

In veadk/tools/builtin_tools/execute_skills.py, after the env_vars dict construction, inject model configuration:

# 注入模型配置，让沙箱使用与本地 Agent 相同的模型。
# VeADK 的 VeADKConfig 在沙箱启动时会读取这些环境变量。
for model_key in [
    "MODEL_AGENT_NAME",
    "MODEL_AGENT_PROVIDER",
    "MODEL_AGENT_API_BASE",
    "MODEL_AGENT_API_KEY",
]:
    model_val = os.getenv(model_key, "")
    if model_val:
        env_vars[model_key] = model_val

Backward Compatibility

• If the env vars are not set locally, nothing changes (empty strings are skipped).
• The sandbox's default model config still applies as a fallback.
• The if key not in env check in the sandbox wrapper code ensures existing env vars in the sandbox image are not overwritten.