[zakahan]

Agent(skills_mode="aio_sandbox", skills=[...]) now uses the existing SkillsToolset to load Skill instructions and execute bash/read/write/edit inside a session-bound AgentKit sandbox. Empty skills use the sandbox default catalog; multiple directories are passed to GET /v1/skills as repeated path parameters.

The ADK app/user/session identity determines the binding, which is reused across turns. Construction does not read host Skill files, commands do not inherit host environment variables, and failed/expired bindings never fall back to local execution. Remote tools handle timeout, cancellation and shell cleanup without automatically retrying mutations. Existing local and skills_sandbox behavior is preserved; AIO callers using cloud SkillSpace IDs must switch to sandbox directories.

Validation:

• 153 related Python tests passed after rebasing onto upstream/main (23014771).
• Full pre-commit passed, including Ruff and secret scanning.
• Authorized live acceptance with doubao-seed-evolving verified the sandbox Tool ID, host environment isolation, all five tools, and sandbox reuse across turns.
• Documentation type checks and build passed.
• Earlier full regression on base c662cc8d stopped with 3409 passed, 8 skipped and 6 failures: three migration-delivery tests lacked zip, and three Harness resource tests lacked llama_index. Full regression is not claimed green.

Includes paired English/Chinese usage documentation and a concise PRD. No dependency changes.