The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/find_transcript.py executes git commands (e.g., git blame, git show, git rev-parse) using the subprocess.run method. The script uses argument lists and avoids shell=True, which is a safe practice that prevents shell injection attacks.
[DATA_EXPOSURE]: The skill accesses the ~/.claude/projects/ directory to read .jsonl transcript files. This behavior is necessary for its stated purpose of recovering reasoning from previous sessions. The data is processed locally to extract reasoning excerpts and is not exfiltrated to any external domains.
[PROMPT_INJECTION]: The skill processes untrusted historical data (past transcripts) which could theoretically contain instructions intended to influence the agent (Indirect Prompt Injection). The skill mitigates this risk through instructions in SKILL.md that direct the agent to treat transcripts as evidence rather than ground truth and through logic in the script that filters out non-reasoning noise.

recall-reasoning