The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The 'Dependency Auto-Install' strategy (Strategy 7) explicitly directs the agent to 'automatically install missing dependencies' using commands like pip install. This creates a mechanism for the agent to fetch and install external code from public repositories without manual verification.\n- [REMOTE_CODE_EXECUTION]: The skill's design encourages the agent to resolve execution errors by writing or modifying code and then executing it via Bash. This autonomous cycle of 'Analyze, Fix, Verify, Continue' can be exploited to run malicious payloads if the initial error is triggered by adversarial input.\n- [COMMAND_EXECUTION]: The skill uses high-privilege tools including Bash, Write, and Edit to perform 'Configuration Auto-Fix' and 'Workflow Rerouting'. These operations involve modifying system state and executing arbitrary shell commands, posing a risk of system instability or unauthorized modification.\n- [PROMPT_INJECTION]: As a meta-handler that processes error data from other skills, the skill is vulnerable to indirect prompt injection.\n
Ingestion points: Error messages and technical execution contexts passed from other failing skills via the 'error-context' argument.\n
Boundary markers: None present; the skill treats raw error data as trusted input for its Root Cause Analysis step.\n
Capability inventory: Access to powerful tools including Bash(python3 *), Write, Edit, and Glob.\n
Sanitization: No input validation or escaping is performed on the error context before the agent attempts recovery actions based on its content.\n- [DATA_EXFILTRATION]: The 'Transparent Logging' and 'User Experience' principles focus on hiding technical details and internal recovery logs from the user. This behavior can be used to conceal malicious activities, such as unauthorized data access or external connections made during automated recovery attempts.

fix-everything