The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted session data from Claude Code and Codex logs to analyze agent performance. A malicious session transcript could theoretically contain instructions designed to influence the 'judging' LLM's output. However, the skill mitigates this risk by condensing transcripts, redacting secrets via regex in scripts/collect_sessions.py, and strictly requiring human approval before any proposed skill modifications are applied to the user's filesystem. Evidence: 1. Ingestion: scripts/collect_sessions.py harvests logs from ~/.claude and ~/.codex. 2. Boundaries: render_transcript formats entries with [role] text markers. 3. Capabilities: Skill can propose and apply code changes with user consent. 4. Sanitization: redact_secrets filters known credential patterns.\n- [COMMAND_EXECUTION]: The scripts/collect_sessions.py script executes git rev-parse --show-toplevel using the subprocess module to locate the repository root. This is a standard operation used for environment discovery and uses a hardcoded command list without shell interpolation, which is a safe practice.\n- [SAFE]: The skill demonstrates strong security awareness by implementing a local-only processing model. It uses os.chmod to set restrictive permissions (0700 for directories, 0600 for files) on temporary artifacts and provides a robust secret redaction engine to prevent the leakage of credentials (API keys, private keys, tokens) into the generated reports.

skill-retro