The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands and a bundled Python script to perform its audit. Specifically, it executes bash -lc to locate state directories and runs python3 context-clean-up/scripts/context_cleanup_audit.py to process logs. These operations are restricted to the local environment and aligned with the skill's stated purpose.
[DATA_EXFILTRATION]: The auditing script reads session history logs from ~/.openclaw. It extracts message previews and metadata, writing them to a JSON report within the workspace (e.g., memory/context-cleanup-audit.json). While this involves processing sensitive session data, there are no network operations or external transmission patterns detected; the data remains local to the user's environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data in the form of previous session transcripts.
Ingestion points: Reads .jsonl files from ~/.openclaw/agents/main/sessions/ via scripts/context_cleanup_audit.py.
Boundary markers: The generated JSON report does not use specific boundary markers or 'ignore' instructions for the extracted text previews.
Capability inventory: The skill has exec, write, edit, and process capabilities.
Sanitization: The audit script performs basic string cleanup (whitespace removal) and truncation for previews but does not sanitize the content for potential instructions. Given the tool's purpose and lack of automated execution of transcript content, the risk is negligible.

context-clean-up