The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from knowledge bundles, which could potentially contain adversarial instructions.
Ingestion points: Data is retrieved from local knowledge/ bundles using okf search and okf show commands as described in SKILL.md and references/commands.md.
Boundary markers: The SKILL.md file includes explicit "Safety and authority boundaries" instructions, directing the agent to treat retrieved memory as data rather than instructions and to verify claims against the repository.
Capability inventory: The skill possesses the ability to read and write to the local filesystem, execute shell commands via the Bash tool, and modify Git repositories (subject to user approval).
Sanitization: The instructions require the agent to redact or pseudonymize sensitive facts and credentials before performing any approved writes to the memory.
[COMMAND_EXECUTION]: The provided scripts/preflight.py utility executes shell commands to verify the state of a repository.
Evidence: The script uses subprocess.run to call git rev-parse. It implements strong security measures including a restricted environment (stripping sensitive environment variables) and strict path validation that rejects symbolic links to prevent directory traversal or unauthorized access.
[EXTERNAL_DOWNLOADS]: The skill documentation references external sources for tool installation and updates.
Evidence: references/commands.md and references/upstream.md mention fetching releases and assets from the official project repository on GitHub. These operations are documented as requiring explicit user consent and are directed toward well-known service providers.

okf-agent-memory