The Agent Skills Directory

[SAFE]: The skill uses a standard platform mechanism (a 'Stop' hook) to execute a local script (evaluate-session.sh) when a session ends. This script is lightweight and only performs basic file existence checks and line counting on the session transcript.
[INDIRECT_PROMPT_INJECTION]: The skill inherently processes session transcripts, which constitute untrusted data. However, this is the primary and intended purpose of the skill (learning from session history).
Ingestion points: The evaluate-session.sh script reads the transcript file path provided by the agent platform via stdin.
Boundary markers: None identified in the script logic, as the script only counts messages to trigger an agent-led evaluation.
Capability inventory: The script is limited to directory creation (mkdir), file existence checks (-f), and pattern counting via grep. It does not perform network operations or execute arbitrary code from the transcript.
Sanitization: The script uses shell quoting for the $transcript_path variable to prevent command injection from the JSON input field.

continuous-learning