The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by capturing user requests and agent responses into a log file, which is subsequently read and processed during 'Review mode'. This creates a surface for indirect prompt injection, as malicious instructions embedded in a user's original request could be stored and later parsed by the agent. However, the instructions for the review phase are strictly scoped to clustering, scoring, and formatting, which significantly mitigates the risk of the agent executing logged content.
Ingestion points: Conversation history and user requests are processed in CAPTURE.md; the magic-fetch.jsonl file is read in REVIEW.md.
Boundary markers: The skill uses JSONL formatting which provides structural boundaries between entries, though it does not explicitly use security delimiters for the data fields.
Capability inventory: The skill utilizes file system tools (Write, Edit, Read) to manage the log file.
Sanitization: Data is stored in structured JSON, preventing raw text execution during the logging phase.
[COMMAND_EXECUTION]: The skill utilizes the environment's file-writing tools to append entries to a local log file. This is the intended primary purpose of the skill and is performed using standard agent tools rather than direct shell execution. The default file path is limited to the local working directory.

magic-fetch