SUPERLOOPY
GitHub112

Evidence-first loop harness for Codex and Claude Code

Done means proven.

Superloopy turns one prompt into a bounded loop. It writes the criteria first, proves each one with a real command, and won't call the work done until the receipts exist.

Install
codex plugin marketplace add https://github.com/beefiker/superloopy
codex plugin add superloopy@beefiker

Run in a terminal, restart Codex, approve the hooks.

Then type loopy in front of any task.

See a real loopStar on GitHub

01 · How it works

One prompt, one bounded loop. Every step leaves a file you can open later.

  1. 01

    Criteria

    loop begin turns the brief into a goal with concrete pass criteria before any code changes.

  2. 02

    Proof

    loop prove -- <cmd> runs the check and stores its full output as evidence, pass or fail.

  3. 03

    Gate

    loop check · loop finish confirms every criterion has a valid artifact. Finish re-runs the proof before it completes.

  4. 04

    Receipt

    A report, a gate file and every capture stay under .superloopy/evidence, next to the code.

$ superloopy loop begin --brief "Fix cart total" --mode light
superloopy began: G001 Fix cart total

$ superloopy loop prove -- npm test
superloopy prove: G001/C001 -> fail (.superloopy/evidence/G001-C001-capture.txt)
# fix: drop the stray “+ 1”
$ superloopy loop prove -- npm test
superloopy prove: G001/C001 -> pass (.superloopy/evidence/G001-C001-capture.txt)
$ superloopy loop prove -- node --test total.test.js
superloopy prove: G001/C002 -> pass (.superloopy/evidence/G001-C002-capture.txt)

$ superloopy loop check
superloopy check: ok
Evidence summary: 2 artifact-backed criteria, 0 unresolved, 0 invalid

$ superloopy loop finish
superloopy finish: complete
Real output from Superloopy 0.22.3 fixing a one-line cart bug.

02 · The crew

Six roles, one job each. Type loopy team to delegate in parallel. A worker without a receipt doesn't count.

See the crew in the backend guide (Korean)

03 · Skills

Lanes for the work you do. The proof rule is the same in every lane.

  1. 01superloopy-loopThe evidence loop: criteria, proof, gate.loopy <task>
  2. 02superloopy-backendContracts first, sibling conventions, guarded migrations.loopy <backend task>
  3. 03superloopy-frontendScreen-based UI, with visual QA as evidence.$superloopy:superloopy-frontend
  4. 04superloopy-cloneRebuild a page from a URL with specs and screenshot diffs.loopy clone <url>
  5. 05superloopy-slidesHTML decks and PPTX, checked visually.loopy slides <topic>
  6. 06superloopy-researchResearch that cites what it actually read.loopy <research question>
  7. 07superloopy-doctorDiagnose install, hooks and host wiring.superloopy doctor
  8. 08humanize-koreanKorean prose that sounds human, facts untouched.$superloopy:humanize-korean
  9. 09say-it-straightDirect, concise prose without losing meaning.$superloopy:say-it-straight
  10. 10i-have-adhdAction-first output, only when you ask for it.$superloopy:i-have-adhd

Invoke any skill with $superloopy:<skill> (Codex) · /superloopy:<skill> (Claude Code)

04 · FAQ

Questions, answered with receipts.

How is this different from a careful prompt?

A prompt asks for care. Superloopy checks for it: criteria are written before the work starts, each one needs a command-backed artifact, and finish re-runs the proof before it completes.

Where does the evidence live?

In your project, under .superloopy/evidence: command captures, the gate file and a markdown report. Keep the folder in .gitignore, or attach the report to a pull request.

What happens when the loop can't finish?

It says so. The report names the blocker and keeps the evidence found so far. With the Stop hook on, a stalled loop is marked blocked instead of done.

Does it work with Claude Code?

Yes. Codex and Claude Code both get the loop, the crew and the skills. Each host installs it once.

Agents move fast. Superloopy makes done mean proven.

Install SuperloopyRead the docs