unlazy

unlazy is a completion-discipline skill for substantial agent work. You write an acceptance ledger first. The agent executes reviewed checks, reverifies returned work, and may report only what the evidence supports. The repo Leonxlnx/unlazy is JavaScript, created 2026-08-09, MIT. On 2026-09-17 GitHub listed 3,410 stars and 236 forks. The source targets 2.1.0 and is not a tagged GitHub release; pin a commit if you need an immutable install. 3,410 stars is a heat signal, not traffic.

Install with the skills CLI: npx skills add Leonxlnx/unlazy. Manual paths: ~/.claude/skills/unlazy or ~/.codex/skills/unlazy. Invoke /unlazy in Claude Code or $unlazy in Codex. Compare Agent Skills for a broader pack, Cachebeat if you only needed a cache keepalive, or Claude Code for the host.

Key Features

• Gate contract: GATES.md lists runnable checks. A gate passes only on exit 0 plus an EXPECT: match on combined output. Evidence is a versioned SHA-256 of the check definition, not a prose checkbox.
• Depth Tree: split a task N layers deep and give every leaf the full time budget of the whole task. Method notes live in references/method.md.
• Checker: scripts/gate-check.mjs (Node 16+, no third-party runtime packages). --status never executes. Normal mode can run a check after you --approve the exact oracle.
• Scoped pipelines: .unlazy/<scope>/ with PLAN, parent gates, and leaf ledgers. Leaves claim disjoint OWNS: paths. --jobs 1-64 is opt-in; checks are sequential by default.
• Optional Claude Code Stop hook: blocks exit while gates or dispatch waves are unmet. Install only with consent.

Limitation: CHECK: lines are shell. Approval is consent, not a sandbox. The checker cannot prove that an English title and a script mean the same thing. Research citations motivate the design; they do not prove a fixed quality gain. Earlier README ratios are called historical and not reproducible from this repo.

Use Cases

• Refactors and migrations where "looks done" is not enough.
• Parallel subagents that must not overwrite each other's files.
• People who should skip it: tiny one-file edits. The ledger overhead is the point, and it is cost.

Pricing

Getting Started

1. npx skills add Leonxlnx/unlazy (add -g for user-level).
2. Ask /unlazy tree 5 refactor the payment module and verify every migration path.
3. Copy templates/gates-leaf.md to GATES.md, fill placeholders, then node .../gate-check.mjs --status GATES.md.
4. Read every CHECK: line. --approve only after that. --reverify before you trust a parent.

First-party start: the README and SKILL.md.

Frequently Asked Questions

Does this make the model less lazy by itself?

No. It adds structure and runnable gates. Papers cited in the README describe laziness and underthinking; they are not a benchmark of unlazy.

Will --status run my tests?

No. --status is non-executing. After approval, a normal run can execute the oracle.

Is 2.1.0 released?

The README says the current source targets 2.1.0 and is not identified as a tagged release. Pin a commit.

Alternatives

• Agent Skills: a wider engineering skill pack, not a gate checker.
• Cachebeat: keeps Claude Code's prompt cache warm; does not verify work.
• Claude Code: the usual host for /unlazy.

Tips

1. Put success-only markers in check scripts. Do not copy the expected number into EXPECT:.
2. Keep approval storage outside the repo (~/.unlazy/approved).
3. Use separate worktrees when OWNS: paths would collide.

Conclusion

unlazy is a strict, open skill for agents that stop too early. Install it, write the ledger, and refuse "done" without evidence. Skip it for small edits, and never treat approval as a sandbox.