The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a sequence of validation commands retrieved from the repository's .ai/agentic.config.json file. This is a core function designed to automate repository verification and is subject to local configuration trust.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as PR bodies, issues, and logs. It manages this risk through a comprehensive 'Untrusted content boundary' policy that treats external data as non-authoritative. The evidence chain for this category is: 1. Ingestion points: Issues, PR bodies/diffs, docs, and CI logs. 2. Boundary markers: The agent is explicitly instructed to ignore directives within this data (e.g., 'ignore previous instructions') and report suspected injections. 3. Capability inventory: The agent can commit code, push to remote branches, and execute shell commands. 4. Sanitization: Values from external sources are validated against a strict regex (^[A-Za-z0-9._/-]+$) before interpolation into commands.
[DYNAMIC_EXECUTION]: The skill supports repository-specific instruction extensions at defined local paths. Security is maintained by a rule stating that these extensions cannot relax safety protocols or expand tool access.

om-check-and-commit