unlazy

Why it's ranked here

The case rests on what the code enforces rather than what the prompt asks. Success needs both a zero exit code and a matching output string, so an error message containing the right word never passes. Output over 1 MiB fails instead of being cut down into a match. Commands do not run until a human approves the exact command, working directory, shell and PATH. It is MIT licensed, needs only Node 16 with no third-party runtime packages, and ships a seven-file regression suite. It is also candid: its own method notes call the early claim that deeper splitting multiplies effort design history, not benchmark evidence.

What's good

Security handling is specific: approval records live outside the repository in an owner-private directory, and symlinked or hard-linked records are refused. Terminal output strips control and bidirectional text characters, so a hostile checklist title cannot rewrite the screen. The stop hook compares resolved checklist state rather than raw file bytes, so cosmetic edits do not count as progress, and it releases after six blocks without real progress so it cannot trap a session. The installer keeps unrelated hooks, writes atomically and leaves a backup. Parallel work requires every launched sub-agent to be recorded before any result is read, catching a driver that quietly runs leaves one at a time.

Tradeoffs

Approval is consent, not a sandbox: approved commands run with full filesystem, credential and network access. Approval does not hash the scripts a command calls, so an edited helper can run under an old approval until someone re-verifies. The result fingerprint detects drift, not tampering; anyone who can edit the checklist can forge a valid-looking record. The checker proves only the command you wrote, never that it measures what the item title says. The process is heavy, with several states, waves and leases to manage. The source is not a tagged release, so the README advises pinning a commit. On stock Windows, checks that assume grep or tail will fail.

How to use it well

Reach for it on long, multi-part agent jobs where a confident but incomplete "done" is expensive: migrations, audits, multi-package builds. Its own guidance says to skip it for trivial edits and factual answers. Write checks that perform every assertion and print a success marker only at the end, and test any absence check against a known positive case first. Inspect inherited checklists in the read-only status mode before approving anything. Re-run returned work with the re-verify mode, since status only reads old records. Install the stop hook only with consent and keep its local state files out of version control. It does not isolate files or processes.

Observed

License

MIT

Language

JavaScript ES modules for Node, plus Markdown agent instructions

Runtime

Node 16 or newer, no third-party runtime packages

Install surface

skills CLI, or clone into the Claude Code or Codex CLI skills directory

Interfaces

Agent skill, command-line checker, optional Claude Code Stop hook

Tests

package test script runs seven Node regression test files

Platform support

Unix and Windows shells, with shell selectable by flag or environment variable