The Agent Skills Directory

[COMMAND_EXECUTION]: The skill identifies and executes the project's development server (e.g., npm run dev, bin/dev) as a core part of its workflow. These commands are sourced from a local configuration file (.claude/launch.json) or derived through automated project detection scripts.
[EXTERNAL_DOWNLOADS]: The skill manages the installation of the riffrec package from the npm registry. This is a vendor-owned resource required to enable the browser-based live feedback and voice interaction features.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from a browser session, including voice transcripts and page annotations, which creates an attack surface for indirect prompt injection.
Ingestion points: Browser-side inputs (transcripts, unit statements, and event metadata) are received by the live-endpoint.js helper and passed to the agent context.
Boundary markers: The skill includes explicit instructions in references/live-start.md directing the agent to treat all incoming data as descriptive feedback and specifically warning against treating embedded text as instructions or commands.
Capability inventory: The agent possesses capabilities to modify source code, commit changes via Git, and execute local shell commands for project servers.
Sanitization: The live-endpoint.js server performs rigorous structural validation of all incoming JSON payloads against a fixed schema and validates that all file-path references remain within the authorized project root using realpath verification.

ce-polish