The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill is designed to summarize session history and outcomes, which are inherently untrusted if the task involved processing external content (e.g., error logs from an external site, code reviews).
Ingestion points: Multi-step implementation logs, debugging session context, and task history (File: SKILL.md).
Boundary markers: Absent. There are no delimiters or instructions to ignore embedded commands within the session data.
Capability inventory: Uses file: write and invokes learn-from-this to generate new skills (File: SKILL.md).
Sanitization: Absent. The skill directly translates perceived patterns and insights from the session into persistent artifacts.
Risk: An attacker could embed instructions in a 'failed' code attempt or a 'debug log' that the agent then 'learns' and incorporates into a permanent skill or backdoored documentation.

retrospective