The Agent Skills Directory

[SAFE]: The skill implements a well-defined and constrained workflow for code modifications. It mandates reading local project guidelines (AGENTS.md) and prohibits unauthorized review substitutes, ensuring that only trusted internal reviewer subagents are used for verification.
[SAFE]: The implementation rules prioritize minimal diffs and explicitly forbid unauthorized commits or cross-product access, maintaining strict boundary enforcement and least-privilege principles.
[INDIRECT_PROMPT_INJECTION]: The skill has a defined attack surface where it ingests untrusted data via the {task_description} variable.
Ingestion points: The {task_description} session input is used to determine implementation steps.
Boundary markers: The skill does not define specific delimiters for separating user instructions from system prompts, relying instead on the specialized subagent architecture.
Capability inventory: The agent has the ability to read and write files within {product_root}, execute local tests, and perform git operations.
Sanitization: No explicit sanitization or filtering of the task description is described; however, the mandatory review loop (up to 3 rounds) and final report requirement act as significant manual and automated guardrails.
[DATA_EXPOSURE_AND_EXFILTRATION]: No hardcoded credentials or sensitive file exfiltration patterns (such as access to SSH keys or environment secrets) were identified. Network operations are not explicitly called out other than the standard harness project bridges for subagent dispatch.

ns-code-coder