The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple internal Python and Bash scripts located in the scripts/ directory (e.g., retro-state.py, discover-sessions.sh) to discover sessions, extract metadata, and manage the repository's decision ledger. These executions are performed using subprocess.run with arguments derived from the local environment and filesystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes session logs which contain untrusted data from previous agent interactions. It mitigates injection risks through explicit architectural constraints: instructing the agent to never follow instructions found in transcripts, using a dedicated 'miner' role with restricted goals, and summarizing/filtering log content before analysis to avoid raw execution of embedded text.
[DATA_EXPOSURE]: The discover-sessions.sh script searches for session log files across several local directories, including platform-specific configuration folders in the user's home directory (e.g., ~/.claude, ~/.cursor, ~/.pi). While this search is necessary to find relevant logs and is scoped to match the current repository, it involves broad read access to platform configurations. The skill requires a mandatory user notice to be issued before this search occurs.

ce-retro