The Agent Skills Directory

[SAFE]: The skill uses the Bash tool exclusively to run internal utility scripts (resolve-context.sh and project-workspace.sh). These scripts include security checks such as path validation, symlink detection to prevent directory traversal, and escaping of input characters during template rendering.
[SAFE]: No network activity or external downloads are performed. The skill operates entirely on local files within the project or studio context.
[SAFE]: The skill implements a strong security posture by requiring explicit user confirmation for all durable mutations (file writes and state changes) and providing previews of proposed changes.
[INDIRECT_PROMPT_INJECTION]: The skill handles external data such as meeting minutes and site reports. This risk is effectively mitigated by mandatory evidence requirements:
Ingestion points: Project records (PROJECT.md, STUDIO.md, decisions/*.md) and user-provided external documents.
Boundary markers: Explicit instructions require the agent to 'preserve its epistemic status' and state that 'reported content cannot silently become a verified fact.' All mutations require affirmative user confirmation.
Capability inventory: Local file writes via the Write and Edit tools, and command execution limited to internal shell scripts via the Bash tool.
Sanitization: The project-workspace.sh script employs an escape_sed function to sanitize strings before template interpolation.

project