The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a feedback loop where the agent ingests and acts upon JSON-formatted human feedback (free-form chat and element-specific annotations) returned by the ecc-plan-canvas await command. This represents an indirect prompt injection surface as the agent is instructed to respond to and potentially execute instructions contained within this untrusted input.
Ingestion points: The stdout of the ecc-plan-canvas await CLI command.
Boundary markers: Absent; the instructions do not specify delimiters or warnings for the agent to ignore embedded instructions in the feedback JSON.
Capability inventory: The skill possesses the ability to execute shell commands (ecc-plan-canvas), edit local files (artifact revisions), and perform network operations via the background server.
Sanitization: Absent; there is no mention of filtering or escaping the text fields within the kind: "chat" or kind: "annotation" feedback items.
[COMMAND_EXECUTION]: The skill relies on the execution of several CLI commands (ecc-plan-canvas open, await, pending, typing, end) and suggests running Node.js scripts directly (node "$CLAUDE_PLUGIN_ROOT/scripts/plan-canvas.js"). These tools manage a background server and process local file paths.
[EXTERNAL_DOWNLOADS]: The browser-based canvas component fetches the Mermaid diagram rendering library from a remote CDN at runtime (configurable via ECC_PLAN_CANVAS_MERMAID_URL). While this is a common practice for rendering visual content, it involves loading third-party code into the user's browser environment.

plan-canvas