The Agent Skills Directory

[DATA_EXPOSURE]: The skill accesses the local T3 conversation database (statev2.sqlite) to analyze interaction patterns. The scripts/collect.mjs script includes a robust redact() function with comprehensive regex patterns designed to remove private keys (RSA/OpenSSH), API tokens (OpenAI, GitHub, AWS, Slack), JWTs, and other credentials before the data is processed or stored. Extracted data is saved locally to ~/.local/state/relay-audit/ and is not transmitted to remote servers automatically.\n- [COMMAND_EXECUTION]: The skill executes git and gh (GitHub CLI) commands via execFileSync to inspect repository history and manage issues. These operations are scoped to the user's local projects and are executed as part of the audit process, with issue creation requiring explicit user decisions.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes historical conversation logs and repository code, creating a surface for indirect prompt injection.\n
Ingestion points: Conversation messages from the T3 database and repository files accessed via git (e.g., AGENTS.md, lint configs).\n
Boundary markers: The skill contains instructions to summarize audited threads rather than quoting them verbatim in public trackers and utilizes redaction to neutralize potential malicious instructions or secrets.\n
Capability inventory: Writing to the local state directory, executing git and gh CLI tools, and calling the Fast Ship API.\n
Sanitization: Active regex-based redaction of sensitive identifiers, secrets, and tokens within the data collection script.

relay-audit