The Agent Skills Directory

[SAFE]: The skill manages session state by reading and writing local Markdown files within the project's .agents/ directory using standard file operations.
[SAFE]: The instructions contain explicit safeguards against data exposure, strictly forbidding the storage of secrets, credentials, or legal client names in the anchor artifacts.
[SAFE]: A 'Context Hierarchy & Trust' protocol is implemented to classify external or browser-sourced content as untrusted, directing the agent to prioritize local rules and source code over potentially malicious history.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting session history and tool outputs, which constitutes a potential attack surface for indirect prompt injection.
Ingestion points: Reads active session state and tool outputs to generate anchors in SKILL.md (Step 1) and agents/openai.yaml.
Boundary markers: The skill uses markdown formatting for structure but does not implement strict delimiters or escaping for untrusted data processed into the anchor.
Capability inventory: The agent is restricted to view_file and write_to_file. It lacks capabilities for network communication, subprocess execution, or privilege escalation.
Sanitization: Sanitization is handled via model instructions that mandate verification of untrusted content against source code before action.

context-anchor