ui-canvas
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from the
ecc-ui-canvas awaitcommand, which returns JSON containing achattext field andpendingEditsdescriptions. The instructions direct the agent to act on this information (e.g., 'Make the requested change in the mockup file') without establishing boundary markers or performing sanitization. This allows external data from the UI tool to influence agent behavior. - Ingestion points: JSON output from
ecc-ui-canvas awaitinSKILL.md. - Boundary markers: Absent; instructions imply direct trust in the feedback content.
- Capability inventory: File system read/write access (e.g.,
.claude/previews/,src/components/), shell command execution (ecc-ui-canvas). - Sanitization: Absent.
- [COMMAND_EXECUTION]: The skill uses the
ecc-ui-canvascommand-line utility for multiple operations, including opening sessions, polling for feedback, and signaling agent state. This requires the presence of a local binary that the agent interacts with continuously during the UI editing session. - [EXTERNAL_DOWNLOADS]: The skill depends on the
ecc-universalpackage and a specific script path ($CLAUDE_PLUGIN_ROOT/scripts/ui-canvas.js) for its core functionality, introducing an external dependency into the agent's runtime environment.
Audit Metadata