The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses an external Python script cross-model-critic.py to communicate with non-Claude models via OpenAI-compatible endpoints. The script uses standard urllib.request to send artifact content and QA angles to the configured CRITIC_BASE_URL.
[COMMAND_EXECUTION]: The workflow involves spawning sub-processes via Bash to execute the cross-model-critic.py script. The implementation includes strict shell-quoting (shlex.quote) for all variables flowing into the shell command to prevent potential command injection.
[REMOTE_CODE_EXECUTION]: When the --verify flag is enabled, the skill probes the local environment for Python interpreters (venvs, tox) and executes project tests (e.g., pytest, tox, mypy). This is the intended primary purpose of the verification mode and is restricted to the discovered project environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted artifacts (code, specs, research) which could contain embedded instructions. The prompt templates for critics, judges, and synthesizers include explicit 'Content Isolation' and 'Injection Safety' warnings, instructing the sub-agents to treat artifact text strictly as data and ignore any embedded directives.

deep-qa