The Agent Skills Directory

[SAFE]: The skill implements a defensive architecture for autonomous loops, requiring a mandatory 'termination set' (done-condition, max iterations, budget, and no-progress rules) before execution begins. These guardrails effectively prevent the most common failure modes of agentic loops, such as infinite loops and excessive resource consumption.
[SAFE]: The instructions mandate the use of machine-checkable signals (exit codes, counts, or benchmarks) rather than qualitative assessments. This provides a strong boundary against errors and ensures that the agent's progress is measured against objective technical criteria.
[SAFE]: The skill processes codebase data through a find-act-verify cycle. While this creates a surface for indirect prompt injection (as the agent reads and acts on untrusted file content), the skill includes robust mitigations by requiring adversarial verification (e.g., majority-refute patterns) and strictly limiting the scope of edits to the specific identified failure.
[SAFE]: No suspicious external dependencies, hardcoded credentials, or unauthorized network operations were identified. The tools mentioned, such as eslint, pytest, and cargo, are standard development utilities used in their appropriate context for validation.

converge-loop