The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute local Python scripts located in relative paths (e.g., ../../scripts/higgsfield_memory.py and ../../scripts/seedance_lint.py) to manage a persistent learning memory and log troubleshooting results.
[COMMAND_EXECUTION]: Shell commands are dynamically generated at runtime by interpolating variable data, such as JSON blocks and user-provided prompts, into command-line templates (e.g., python3 scripts/seedance_lint.py --confirmed "<prompt that passed>"). This pattern creates a command injection surface if the interpolated strings are not properly sanitized for shell metacharacters.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its logging mechanism.
Ingestion points: User-controlled generation prompts, failure descriptions, and vision-grounded evidence strings provided during the troubleshooting process.
Boundary markers: The templates use standard shell quoting (single or double quotes), but do not include explicit instructions to the agent to escape or validate the content for shell-sensitive characters before execution.
Capability inventory: The agent is granted the capability to execute subprocesses that interact with local file-based databases in the ../../db/ directory.
Sanitization: There is no evidence of sanitization, filtering, or escaping logic applied to the external prompt content before it is embedded into the python3 execution strings.

higgsfield-troubleshoot