The Agent Skills Directory

[COMMAND_EXECUTION]: Utilizes eval to execute the output of local binaries like gstack-slug and gstack-paths. This pattern is risky as any unexpected output from these binaries would be executed directly by the shell.
[COMMAND_EXECUTION]: Runs dynamic Javascript code using bun -e. This bypasses standard script visibility and increases the complexity of the execution environment.
[COMMAND_EXECUTION]: When the environment variable SPAWNED_SESSION is set to true, the skill suppresses interactive user prompts and automatically proceeds with file deletions (git rm) and configuration changes, reducing human oversight for destructive actions.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface by processing natural language data from learnings.jsonl to perform contradiction checks and export documentation.
Ingestion points: Data enters the context from the learnings.jsonl file via the prune and export functionalities.
Boundary markers: None; the agent processes the file content as raw insights without explicit markers or 'ignore' instructions.
Capability inventory: File system access, shell execution, and git management tools.
Sanitization: No validation or filtering is applied to the content of the learnings file before it is interpreted by the agent.

learn