compaction-resilience-guard
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill documentation provides command-line examples for running the local Python script (
guard.py) to check state, simulate fallbacks, or view reports. These commands are standard for tool interaction and do not involve shell injection or privileged operations. - [REMOTE_CODE_EXECUTION]: Analysis of
guard.pyconfirms there are no network operations, remote downloads, or calls toeval()/exec()on external data. The summarization strategies are implemented using local string manipulation (L1, L2, L3 deterministic truncation). - [DATA_EXFILTRATION]: The skill only interacts with local state files within the specified
~/.openclawdirectory. No network capabilities or sensitive file access patterns (like.sshor.aws) are present. - [CREDENTIALS_UNSAFE]: No hardcoded API keys, tokens, or private secrets were found in the scripts, metadata, or example state files.
- [INDIRECT_PROMPT_INJECTION]: While the skill processes user-supplied text for simulation (
--simulate), it performs purely programmatic analysis (Shannon entropy, character count, line splitting) and does not interpolate the data into a prompt for another LLM call within this skill's code. Sanitization is effectively handled by treating the input as raw data for local heuristics.
Audit Metadata