The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to collect environment information using env | grep -E "NODE|NPM|PATH" and subsequently post it to GitHub issues using the gh CLI tool. This pattern creates a risk where sensitive build tokens (e.g., NPM_TOKEN, NODE_AUTH_TOKEN) or internal environment paths could be inadvertently exfiltrated and publicly exposed if they match the filter patterns.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process error logs and stack traces from external commands (pnpm test, custom scripts), which is a common attack vector for indirect prompt injection.\n
Ingestion points: The skill captures output from shell commands and local log files like error-log.txt directly into the agent's context.\n
Boundary markers: The instructions use markdown code blocks for logs but lack explicit system-level instructions for the agent to disregard instructions that might be malicious or embedded within the error data.\n
Capability inventory: The skill utilizes powerful and potentially destructive commands including git reset --hard (file system modification) and gh issue comment (external communication).\n
Sanitization: No sanitization or filtering of the captured error logs is performed before they are processed or stored in long-term memory via mcp__memory__add_observations.

error-recovery