The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses internal configuration and conversation history files located in the ~/.claude/ directory, specifically reading state.json files for job metadata and .jsonl transcript files for conversation content. This data is used to provide summaries to the user.
[COMMAND_EXECUTION]: The skill employs several system commands for job management: kill is used to terminate background process PIDs, rm -rf is used to delete job directories in the jobs sub-folder, and printf is used to append color-coding metadata to transcript files.
[COMMAND_EXECUTION]: A local Python script is executed via a heredoc (`python3
<<'EOF'`) to parse job state files and generate formatted summaries of active background tasks.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it reads and summarizes external data sources such as job outputs and conversation transcripts.
Ingestion points: Files located at ~/.claude/jobs/*/state.json and ~/.claude/projects/*/*.jsonl.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands during the summarization process.
Capability inventory: The agent has access to file deletion (rm), process termination (kill), file modification (printf), and local script execution (python3).
Sanitization: There is no explicit instruction to sanitize or escape data before processing.

claude-agents