The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest human feedback and artifact content, which can then influence the agent's actions through a structured JSON response.
Ingestion points: Untrusted data enters the agent context via the output of the ecc-plan-canvas await command, which provides chat messages, element annotations, and artifact snippets (SKILL.md).
Boundary markers: The data is structured as JSON; however, there are no specific prompt instructions for the agent to treat the text content as untrusted or to use explicit delimiters when processing the natural language fields.
Capability inventory: The agent can execute shell commands, read and modify local files, and transition between planning and implementation phases based on the feedback verdict (SKILL.md).
Sanitization: No explicit sanitization, filtering, or validation of the human-provided feedback or reviewed artifact content is specified.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute the ecc-plan-canvas CLI tool (or a specific Node.js script path) to manage the review server and poll for feedback batches.
[EXTERNAL_DOWNLOADS]: The skill utilizes a pinned CDN to load Mermaid.js in the browser for rendering diagrams within the review interface. This is a client-side resource for the human user.

plan-canvas