The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes internal helper scripts to manage its review workflow. It uses scripts/peer-job-runner.py to launch detached supervisor processes and scripts/cross-model-doc-review.sh to coordinate with various model provider CLIs. These scripts use subprocess.Popen, os.fork, and os.setsid to manage the process lifecycle and execute system commands.
[EXTERNAL_DOWNLOADS]: The scripts/packs-resolve.py script resolves 'Compound Packs' by performing git clone on repositories defined in the repository's configuration files (config.yaml and config.local.yaml). This allows the agent to download and incorporate external content from user-specified Git sources.
[DATA_EXFILTRATION]: As part of its 'Cross-Model Judgment Pass', the skill transmits the full content of the document being reviewed to third-party AI service providers, including OpenAI, Anthropic, xAI, Google, and Cursor. While this is a documented feature for independent verification, it involves transmitting potentially sensitive planning data to external servers.
[DYNAMIC_EXECUTION]: The skill executes code at runtime using npx to load and run specific versions of the acpx package. Additionally, scripts/cross-model-doc-review.sh dynamically constructs command-line strings to invoke various external provider CLIs (e.g., codex, claude, grok).
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided documentation which is fed into persona-based subagents. If these documents contain adversarial instructions, they could potentially influence the behavior of the reviewers. The skill acknowledges this surface but considers it a low-risk vector for planning documents.
[PRIVILEGE_ESCALATION]: The skill instructions direct the agent to request 'require_escalated' permissions in certain environments (like Codex) to enable network communication for cross-model validation with external AI providers.

ce-doc-review