The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script tools/extract_prompts.py to extract and normalize user prompts from agent transcripts stored on the local filesystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from past AI session transcripts, which could contain malicious instructions or prompt injection attempts designed for the original agent.
Ingestion points: Conversation logs are read from ~/.claude/projects/, ~/.codex/sessions/, and ~/.gemini/tmp/ via the adapter system.
Boundary markers: Transcript data is parsed and structured into JSON format before being batch-processed by analysis subagents.
Capability inventory: The skill has read access to local transcript directories, write access to a history file (~/.roast-me-history.json), and executes local Python scripts.
Sanitization: The extraction script truncates prompt text to 1500 characters and filters for meaningful issues before report generation.
[PERSISTENCE]: The skill maintains a local history of audit results by appending data to ~/.roast-me-history.json. This behavior is intended for trend tracking over time and does not include executable code.
[DATA_EXFILTRATION]: The skill accesses sensitive directories containing AI assistant conversation histories, specifically ~/.claude/projects/, ~/.codex/sessions/, and ~/.gemini/tmp/. While this access is necessary for the skill's auditing function, these files contain highly sensitive interaction logs.

roast-me