The Agent Skills Directory

[SAFE]: The skill is designed to improve agent performance by analyzing local interaction history. It includes a dedicated utility script (transcript_common.py) that implements comprehensive redaction for emails, API keys, EVM addresses, and generic secrets using regular expressions to ensure sensitive data is not processed by the model or exported.
[COMMAND_EXECUTION]: The skill uses bundled Python scripts (transcript-miner.py and transcript-inspect.py) to process local JSONL transcript files. These operations are restricted to the local filesystem (targeting ~/.claude and ~/.codex directories) and do not involve network connectivity or external downloads.
[DATA_EXPOSURE]: Access to local transcript data is the primary function of the skill. Security risks are mitigated by explicit instructions in SKILL.md forbidding the exposure of credentials and requiring an external-disclosure review before any transcript content is written to public or shared repository artifacts.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests and analyzes historical chat transcripts (which are untrusted data), there is an inherent surface for indirect prompt injection. This risk is addressed through instructional guardrails in the 'Evidence Contract' and 'Scope and Authority' sections, which direct the agent to treat transcripts as auditable evidence rather than authoritative instructions.

agents-introspection