The Agent Skills Directory

[SAFE]: The skill implements a '3x3 lens' thinking framework (fractal thinking) using various external engines (codex, claude, grok, qwen, kimi, pi). All detected behaviors are aligned with the stated purpose of coordinating multiple AI outputs.
[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen in several scripts (fractal.py, fractal_viz.py, engine_legs.py, etc.) to invoke AI engine CLIs (claude, codex, grok, pi, kimi). These are the primary functions of the skill and are handled using safe patterns, such as piping input via stdin or temporary files to avoid command injection via shell arguments.
[EXTERNAL_DOWNLOADS]: The skill references several external resources, including GitHub repositories (github.com/alexio777/fractal-delphi) and API endpoints (api.deepseek.com, api.deepseek.com/v1, token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic). These are consistent with the vendor's own infrastructure and established AI service providers.
[PROMPT_INJECTION]: The skill includes a dedicated safety module (codex_guard.py) that implements a 'pre-flight guard' to screen every prompt for offensive intent or cyber-abuse patterns before sending them to external engines. This is a security best practice for skills that process user-supplied questions.
[CREDENTIALS_UNSAFE]: The skill interacts with local authentication for various CLIs (e.g., ~/.codex/auth.json, ~/.kimi-code/credentials, ALIBABA_CODING_PLAN_API_KEY). It does not exfiltrate these credentials; rather, it instructs the local CLIs to use existing authenticated sessions to avoid spending API tokens. This is a legitimate cost-saving measure for the user.
[DATA_EXFILTRATION]: While the skill communicates with external AI services, this is the core intended functionality. The data sent is limited to the user's question and technical context. There are no patterns suggesting the exfiltration of sensitive files like SSH keys or environment secrets.
[DYNAMIC_EXECUTION]: The skill uses importlib.import_module in engine_legs.py for lazy loading of backend modules (claude_backend, codex_backend, etc.). This is a standard Python design pattern used here to allow for backend mocking during testing and does not constitute a security risk.

fractal