The Agent Skills Directory

[SAFE]: The skill implements robust security invariants, such as preventing silent background mutations and requiring explicit user consent for any file system changes or tool registrations. Standard practices for secret management are followed by utilizing environment variables and .env files rather than hardcoding credentials.
[INDIRECT_PROMPT_INJECTION]: The skill includes an ingestion surface for untrusted data, specifically task titles and descriptions, which are processed by an external decision engine (TypeSafe SDK) to profile task complexity and reasoning needs. While this is an attack surface for indirect prompt injection, it does not lead to executable consequences as the output is only used for selecting model tiers and effort levels.
Ingestion points: The task object passed to agent_config_recommend in scripts/agent_config.py flows into abstract_profiler.py.
Boundary markers: None; task descriptions are passed directly to the profiling logic.
Capability inventory: The skill is limited to logic and returning structured JSON metadata. Shell execution (subprocess) is strictly confined to the unit test suite and not used in production logic.
Sanitization: No explicit sanitization or filtering of task description strings is implemented before passing them to the decision client.

agent-config