The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a custom Bash script assets/orca-eval-scan.sh to locate and parse project-specific session logs.\n- [DATA_EXFILTRATION]: The skill accesses sensitive data within ~/.claude/projects/, specifically reading .jsonl files that store complete conversation transcripts. These logs may contain sensitive information, hardcoded secrets, or private code shared during previous sessions.\n- [DYNAMIC_EXECUTION]: The script assets/orca-eval-scan.sh dynamically executes a multi-line Python block using python3 -c to perform JSON parsing and data distillation of the log files.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from past conversation logs, creating a surface for indirect injection attacks.\n
Ingestion points: The assets/orca-eval-scan.sh script reads and extracts text from all .jsonl files in the local project history directory.\n
Boundary markers: The extracted content is structured under headers and truncated (e.g., to 200 characters for prompts), but there are no explicit safety markers to prevent the agent from obeying instructions found within the logs.\n
Capability inventory: The skill can write files to the llmwiki/ directory and propose modifications to project configuration files like CLAUDE.md or the creation of new skills.\n
Sanitization: The processing script performs basic character escaping and newline removal for formatting purposes but does not implement security-focused filtering of the log content.

orca-eval