unlazy is a skill for AI agents that enforces completion discipline through a structured gate-based verification system.
The tool addresses the problem of AI model laziness and underthinking by introducing the Depth Tree method, which splits tasks into multiple layers and allocates the full time budget to each leaf node, multiplying effort with depth. Rather than allowing agents to complete work prematurely, unlazy requires writing an acceptance ledger first, executing reviewed checks against that ledger, reverifying returned work, and reporting only what evidence supports. Gates are runnable checks that pass only when their process exits zero and output matches an expected fingerprint, creating an unambiguous record of what was actually verified.
The tool suits teams building substantial AI-agent workflows where premature completion or underthinking poses real risks. It works with supported agents through the skills CLI or manual installation into agent directories. The core checker requires Node 16 or newer but has no third-party runtime dependencies. Developers should understand that normal mode prints resolved commands without executing them on first run, allowing inspection before approval; once an oracle is approved, normal mode can execute. The gate contract is strict: both stdout and stderr must fit within a 1 MiB limit, and evidence includes a versioned SHA-256 digest of the gate definition itself, detecting structural drift. Manual gates with human evidence remain compatible alongside runnable gates. The parser rejects malformed ledgers, duplicate gate ids, and incomplete definitions, preserving line endings when updating.
Development activity shows sustained focus on the core verification mechanism, with recent work refining the gate contract semantics and evidence binding. The project maintains a detailed changelog and pins installations to exact commits for immutability. Documentation emphasizes the research basis grounding the approach in model behavior, and the skill description is designed to trigger naturally within agent systems.
