The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The procedure explicitly directs the agent to identify and snapshot sensitive data such as credentials and environment variables during the "LOCK SOURCE STATE" and "PACKAGE CONTEXT" phases.
[DATA_EXFILTRATION]: The instruction to "save env state" and "note credentials" creates a risk of exposing sensitive secrets by including them in the agent's output deliverables, such as the context-switch checklist or state inventory.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture current session state from untrusted sources, which introduces a surface for indirect injection if the captured data contains malicious instructions.
Ingestion points: Environment variables, open files, running processes, and terminal history (defined in SKILL.md).
Boundary markers: The skill lacks delimiters or instructions to ignore embedded commands within the captured state data.
Capability inventory: The skill generates commands for environment management and configuration tools including git, kubectl, and aws configure (defined in SKILL.md).
Sanitization: There are no measures described to sanitize or filter the captured state data before it is processed or output.

dimension-door