The Agent Skills Directory

Dynamic Instruction Discovery: The skill directs the agent to locate and load additional SKILL.md files from a parent directory (e.g., ../../.opencode/skills/*/SKILL.md). This behavior allows the agent's operational logic to be dynamically extended at runtime based on the file system structure. While intended for cross-repo development, this mechanism relies on the integrity of the discovered files.
Information Boundary Enforcement: The skill contains detailed instructions to prevent the leakage of proprietary or internal information from a parent project into the public-facing workerd repository. This is a defensive measure designed to maintain data privacy during cross-boundary tasks.
Indirect Prompt Injection Surface: By reading external markdown files into its context, the agent is exposed to potential instructions embedded in those files.
Ingestion points: ../../.opencode/skills/*/SKILL.md and ../../AGENTS.md (SKILL.md)
Boundary markers: The skill includes an 'Information Boundary' section that warns against leaking internal data, though it does not provide specific delimiters for the ingested content itself.
Capability inventory: The agent uses a Read tool to ingest the content of discovered files into its reasoning context.
Sanitization: No explicit sanitization or validation of the ingested markdown content is mentioned beyond the agent's general reasoning instructions.

parent-project-skills