The Agent Skills Directory

[SAFE]: The skill implements a comprehensive security model for managing agent state. The apex-state.py script includes a secret redaction mechanism that masks API keys, bearer tokens, and private keys using multiple regex patterns before persisting data to disk. The redaction is applied to user-supplied task descriptions, run IDs, and event messages.
[SAFE]: The framework enforces strict filesystem security by setting private permissions (0700 for directories, 0600 for files) and performing explicit checks to prevent symlink attacks and path traversal beyond the defined repository root during state initialization and artifact recording.
[COMMAND_EXECUTION]: The skill interacts with Git to capture repository baselines and status. These interactions are performed via subprocess.run using fixed argument lists without a shell. The inputs, such as file paths and branch names, are validated and passed through the redaction engine to prevent command injection or accidental data exposure.
[INDIRECT_PROMPT_INJECTION]: The skill explicitly identifies and mitigates the risk of processing untrusted data. The instructions in step-00-init.md mandate that all content from code, documentation, tool output, and external systems be treated as untrusted data that cannot expand the agent's authority. The framework also utilizes a coordinator/worker model where independent workers are assigned bounded packets, and their output is strictly reviewed by the coordinator.

apex