unlazy

A skill that makes an agent write down what finished means before it starts. Each gate in the ledger is a shell command plus the output it must produce, and a run counts as a pass only when the process exits 0 and the expectation matches the combined output, so a claim of completion has to survive a re-run. A checker script — Node 16 or newer, no third-party runtime packages — can inspect a ledger without executing anything, execute it after an explicit approval, or re-run every runnable gate including the ones already marked complete, and an optional Stop hook refuses to let a Claude Code session end while gates are unmet or launch waves are unfinished. Above the single gate sits the Depth Tree method it is built on: split the task N layers deep and give every leaf the full time budget of the whole task, so the effort spent rises with the depth of the tree instead of being divided by it, and orchestrated runs claim disjoint file ownership before independent leaves start together.

Who built itA developer in Munich writing as Leonxlnx, with an account opened on 2025-07-03, 47 public repositories, 2,167 followers and the bio “I build cool stuff”. He wrote 24 of this repository’s 49 commits; the remaining 25 are spread across nine other accounts plus two commits with no linked account. His other repositories include taste-skill, at 91,596 stars, and agentic-ai-prompt-research, at 2,550.

1. 01

   An anti-laziness skill with 3,782 stars in its first month

   The repository was created on 2026-08-09T23:39:12Z and its oldest commit is stamped fifteen seconds earlier — 2026-08-09T23:38:57Z, “unlazy: anti-laziness skill built on the Depth Tree method” — which is what a local history pushed into a fresh repository looks like. The published history is 49 commits over twenty-five days: 48 in August and one in September, the newest dated 2026-09-03T09:31:08Z, “fix: bind gate evidence and harden Windows identity”. By 2026-10-01 it held 3,782 stars, 281 forks, 12 watchers and nine open items, eight of them pull requests, the newest of those dated 2026-09-22. Ten accounts appear in the contributor list; the owner wrote 24 of the 49 commits, 47 commits carry a linked account, and twenty carry co-author trailers, every one of them a Claude model — twelve Opus 5, five Fable 5, two Opus 4.8 with a million-token context and one more Opus 5 with the same context. This record files the project as maintained rather than active: four weeks without a push is a pause rather than an ending, nothing is archived, and the queue kept moving after the commits stopped — eight pull requests are still open, dated from 2026-09-11 to 2026-09-22, every one of them newer than the last commit on main.

2. 02

   The Depth Tree, and a research claim written out in full

   The repository states its core in one line: the Depth Tree method, which splits a task N layers deep and gives every leaf the full time budget of the whole task, so effort multiplies with depth — grounded, it says, in 2025–2026 research on model laziness, underthinking and premature completion. Its research section refuses the strong reading of that grounding: research supports the failure modes motivating explicit structure, it says, and does not prove that unlazy produces a fixed improvement; it then argues from numbers. SlopCodeBench, where the best agent passed 14.8% of checkpoints; METR’s Time Horizon 1.1, reporting a 196.5-day overall P50 doubling-time fit against 130.8 days after 2023, with a warning not to quote the shorter figure as the all-years estimate; and s1’s budget forcing, which appends Wait when a model tries to stop, while denying that one token always improves work. Below that is a sources list ordered newest first, from 2025-02-18 to 2026-08-06 plus one undated page. Every entry carries a title, a link and a date, two of them a venue, and none an author name, so the citations are specific but not full references. The README also withdraws earlier evidence of its own: previous revisions cited a six-run internal comparison whose raw artifacts are not in the repository, to be read as historical design input rather than a benchmark guarantee.

3. 03

   What the agent has to write down before it is allowed to stop

   Installation is one command through the skills CLI, npx skills add Leonxlnx/unlazy, or a clone into ~/.claude/skills/unlazy or ~/.codex/skills/unlazy; it answers to /unlazy, to $unlazy in Codex, or to a natural-language trigger. The README’s example is /unlazy tree 5 refactor the payment module and verify every migration path; that number is the depth. What the agent owes is a ledger: each gate is a checkbox line with a title, a CHECK: shell command, an EXPECT: string, an optional CWD: and an EVIDENCE: line. The checker has three states: --status, the only mode that never executes; a normal run, which on an unapproved oracle prints the resolved command, expectation, working directory, shell and inherited PATH instead of running it, and which the README warns is not a permanent dry run; and --approve, after the user has read every command and script it calls. --reverify re-runs every runnable gate, including completed ones. A runnable gate passes only when the process exits 0 and the expectation matches the combined output, both capped at 1 MiB; a larger matcher string is never truncated into success. The parser rejects ledgers with no gates, duplicate ids, incomplete runnable gates, invalid expectations, or an abandonment with no reason or an unknown gate id. Abandonment is a terminal handoff rather than a pass: exit 1 with HANDOFF REQUIRED.

4. 04

   The hook, the shell, and the machine where nothing could be read

   The optional piece is a Claude Code Stop hook, installed only with consent through install-hooks.mjs: it scans the session’s resolved ledger and dispatch state and returns Claude Code’s top-level decision: "block" while gates are unmet or waves are incomplete, without running any check itself. Its session-keyed guard releases after six consecutive blocks without semantic progress, and metadata-only edits do not reset it — a fix from a pull request where the earlier version hashed ledger bytes, so any edit rearmed the counter and trapped an agent that could not finish. The checker takes its shell from --shell, then UNLAZY_SHELL, then the platform default, and checks inherit the launch environment including PATH, so the same checker started from Git Bash sees tools that one started from PowerShell does not; --shell changes the interpreter, it does not install grep or tail. On Windows a reader found every file read failing closed: on his machine fstat().dev never equals lstat().dev, so the checker could read no ledger and a clean clone scored 5 of 32 tests. The fix compares bigint descriptor identities through a second non-creating handle; the snapshot-based guarantee in SECURITY.md stays. An earlier Windows round removed orphaned processes: killing the shell left a check’s children running, so a timeout now calls taskkill /pid <PID> /f /t.

5. 05

   Gates that could not fail, and the argument nobody read

   The most useful rounds came from people running the skill on real work. One report showed that a gate whose CHECK: never observes what its title claims still returns PASS with evidence and exit 0, because the leaf check, --reverify and the Stop hook all consult the same oracle; a second reporter reached the same class from a different direction, trialling the skill on a sixteen-week course readiness audit with gates authored by another model under another harness and no hook installed, where the gate that worried him guarded the task’s one hard constraint — read only, no writes. The maintainer fixed the machine-detectable false greens and merged the authoring guidance, while saying plainly that the checker still cannot prove that prose and command mean the same thing and that the boundary is now explicit rather than hidden. A separate bug is worth recording because it was reported more than once: indexOf("--timeout") returns -1 when the option is absent, so the next index is 0, the filter drops the first positional argument, and naming a single ledger file either widened the run to every ledger in the directory or reported success without reading the file at all. It was fixed with explicit named-file targeting and a regression test, and one of the two reports was a pull request its own author closed as a duplicate of an earlier one after checking the open queue.

6. 06

   Fifteen items of community work, and no release to put them in

   Three reports went after the hierarchy rather than one gate: an abandoned required gate could print ALL MET and hand a parent completion credit; approval could stay current after the called script changed; and a plan could omit one of two requested deliverables while every declared gate passed. The first and third were fixed — abandonment can no longer promote a parent, and plans now inventory independently omittable requirements and reread the request before dispatch — while the second was closed as a documented boundary rather than a defect, because approval binds the declared command and environment, not transitive script bytes. There are no releases and no tags at all. The README says the current source targets 2.1.0, is not a tagged release, and tells readers to pin an exact commit; what would be release notes lives in a 16,636-byte CHANGELOG.md that also records contributor history and pull-request links. That unreleased 2.1.0 is where the community work landed, in fifteen listed items that integrate the useful parts of the pull requests while repairing their edge cases; the test count moved with it, from 137 at commit 265fbd5d to 188, with the self-check at 15 of 15. Continuous integration runs across Ubuntu, Windows and macOS on Node 16, 20 and 24 — 9 of 9 jobs in August, 10 of 10 in September with a pinned Node 22.14.0 and libuv 1.49.2 Windows job.