The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data such as session transcripts and repository traces to build evaluation charts (see references/session-derived-atlas.md). While this constitutes an indirect prompt injection surface, the framework includes robust mitigations:
Ingestion points: Session data is retrieved via the seq utility.
Boundary markers: Explicit 'Visibility boundaries' and 'actor-readable inventory' protocols are defined to separate actor inputs from evaluator-only data.
Capability inventory: The skill utilizes shell commands for environment setup and executes emulated 'harnesses' within isolated workspaces.
Sanitization: Strict guidelines require sanitizing shareable reports and prohibit committing raw sessions or credentials.
[COMMAND_EXECUTION]: The skill instructions include the use of CLI utilities like seq, cas, and git for session discovery and environment validation. These tools are used for legitimate setup tasks and adhere to the vendor's intended functionality for environment emulation.
[DYNAMIC_EXECUTION]: The skill manages the execution of 'executable worlds' and agent 'harnesses' defined in content-addressed contracts. The framework emphasizes safety through the use of isolated worktrees, deterministic reset recipes, and restricted effect policies (e.g., denying network and side effects by default) to ensure that emulated runs do not affect the host system or live environments.

emulator