The Agent Skills Directory

[DATA_EXFILTRATION]: The skill references a sensitive file path at ~/.sentry/mcp.json, which is used to cache device-code authentication tokens. This behavior is documented as part of the authentication flow for the Sentry MCP server tools.
[COMMAND_EXECUTION]: The instructions require the execution of various shell commands using pnpm and curl to build, test, and verify tool behavior. These commands target local workspace scripts and the local development server (http://localhost:5173/).
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to pass arbitrary prompts to a CLI tool that interacts with production Sentry data. This creates an ingestion surface where untrusted or malicious instructions in the data could influence the agent's behavior during the QA process.
Ingestion points: The CLI tool accepts user-provided prompts that are executed against production API endpoints (SKILL.md).
Boundary markers: None explicitly mentioned in the instructions to separate prompt input from tool execution logic.
Capability inventory: The CLI can execute Sentry tools (execute_sentry_tool), list tools, and perform authentication actions.
Sanitization: Not explicitly defined in the skill instructions; relies on the underlying CLI and LLM guardrails.

mcp-qa