The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses historical session logs located in sensitive user directories (e.g., ~/.claude/projects/, ~/.codex/sessions/, ~/.grok/sessions/, and ~/.gemini/antigravity-cli/). These logs contain interaction history and potentially sensitive development context. The skill mitigates risks by enforcing read-only access (e.g., using sqlite3 -readonly) and requiring mandatory redaction of API keys, tokens, and passwords before displaying information to the user.
[INDIRECT_PROMPT_INJECTION]: The skill mines untrusted data from past session transcripts. This creates an attack surface where instructions embedded in historical logs could attempt to influence the agent's current environment analysis. The skill limits this risk by only producing a ranked proposal list for the user to review and explicitly select, rather than executing commands directly from the mined content.
[COMMAND_EXECUTION]: The skill utilizes local command-line tools such as jq, sqlite3, zstd, awk, and grep to process and filter log data. These operations are restricted to the local environment and are used to extract specific message types (human prompts, tool results) to generate a summary for the user. It also uses awk for round-robin batching of files for parallel processing.

retro