MS-Agent Capability Gateway

The skill exposes a wide range of tools via MCP (deep research, code generation, video generation, file editing), but provides no permission controls, user confirmation mechanisms, or data-flow transparency. Documentation claims code generation runs in a Docker sandbox, but the enforcement and isolation are unverified. File editing tools can modify arbitrary paths without path restrictions or confirmation. Dependencies (e.g., akshare, pyright) are not security-audited. External effects are significant (e.g., generating videos, writing files), and no rollback mechanism is provided. Publisher is unverified, but source attribution is traceable (Apache-2.0, public repo). Score is below the anchor of 12, hence deductions.

SKILL.md and reference files provide detailed self-consistent instructions, including async patterns, parameters, and SOPs, but no executable or reproducible test evidence is provided. Although CI workflows and test files exist, the test file (e.g., test_partial_round.py) covers framework logic, not the skill's key paths (research, codegen, video). Thus, key path reproducibility is unverified. Error handling is documented, but actual failure cases are missing. Score is below the anchor of 10.

The skill clearly defines target audience (users needing deep research, code generation, etc.) and scenarios, and provides a quick decision guide for trigger conditions. However, non-fit boundaries are not explicitly stated, e.g., would it falsely trigger for simple summaries? Environment fit: dependencies on LLM APIs (e.g., ModelScope, OpenAI, Google) may have limited reachability from mainland China, and the docs do not address mainland China network access. Hence, deduct from the anchor of 15.

Documentation is well-structured with progressive disclosure (SKILL.md overview, reference files for details). Install and dependency notes are provided (e.g., pip install ms-agent), plus configuration examples for nanobot, OpenClaw, and Hermes. Parameters and tool naming are stable. Known limitations are disclosed (e.g., dependency on search API availability). License (Apache-2.0) and version (1.0.0) are given, but there is no changelog or maintenance responsibility statement. Publisher identity is unverified, but license and repo are public. Score is below the anchor of 15.

The skill describes capabilities (research, codegen, etc.), but provides no actual output examples or validation results. Claimed outputs like reports and videos are said to be directly usable, but there is no evidence they are verified. After completing the core task, outputs may require heavy rework, e.g., report quality depends on LLM. Thus, score is far below the anchor of 7.

Only author-provided documentation and examples, no third-party execution evidence. CI workflows and tests exist, but do not cover the skill's key paths. Claims cannot be independently reproduced. Score is far below the anchor of 5.