English | 한국어 | 简体中文 | 日本語 | Español | Português (BR) | Français | Русский | Deutsch

Re:ZERO Loop

Re:ZERO Loop is an agent workflow inspired by Return by Death from Re:Zero − Starting Life in Another World.

It lets Subaru implement work, sends the result to seven independent witches for review, preserves failure memory, and retries from HEAD when the route fails.

Table of Contents

• Installation
  • Pi
  • Claude Code
  • Codex
• Usage
• Workflow
• Skills
• Language and Names
• Concepts
  • Natsuki Subaru
  • Return by Death
  • Seven Witches
  • Rem
• License

Installation

Pi

pi install npm:rezero

Git install:

pi install git:github.com/epsilondelta-ai/rezero

Local development:

pi install /path/to/rezero

Claude Code

/plugin marketplace add epsilondelta-ai/rezero
/plugin install rezero@rezero-marketplace

Codex

codex plugin marketplace add epsilondelta-ai/rezero

Then open /plugins, install rezero, and start a new session.

Usage

Initialize witch evaluation tools in a target repository:

/rezero init

Run the loop:

/rezero <task>

Example:

/rezero Add user profile editing with validation and tests

Workflow

0. Init

   • /rezero init detects the repository stack.
   • It configures fitting witch evaluation tools, keeps .rezero/memory/ ignored, verifies available tools, and writes .rezero/tools.md with <!-- rezero-init: v0.1.0 -->.
   • Non-init /rezero requests auto-run init first when .rezero/tools.md, the init marker, or .rezero/memory/ ignore state is missing.
   • Tools that require accounts or external services are recorded as setup notes instead of being faked.

1. Orchestrate

   • /rezero loads rezero.
   • Large requests are split by rezero-plan into ordered tasks with done criteria.
   • Independent tasks may run in parallel through subagents or team agents.

2. Implement

   • Subaru runs one sequential task or one parallel task group from current HEAD.
   • Parallel groups are merged first, then verified as one combined result.

3. Evaluate

   • rezero-witches calls all seven witches in parallel.
   • Witches use fresh context, not Subaru's context, to avoid confirmation bias.
   • Chat shows one verdict table:

| witch | verdict | reason | evidence |
|---|---|---|---|
| Echidna | pass/warning/fail | <short reason> | <command/output/file> |

4. Return by Death
   • Any fail kills the route.
   • Subaru writes minimal failure memory to .rezero/memory/subaru-deaths.md:

## Death <number>

- Fail: <witch + reason>
- Evidence: <minimal test/review/error/defect>
- Next route: <specific change>

• Then resets:

git reset --hard HEAD
git clean -fd

• .rezero/memory/ is ignored, so failure memory survives.

5. Pass

   • Only pass and warning verdicts pass.
   • Warnings are stored by rezero-rem in .rezero/memory/rem.md.
   • Accepted route is committed.
   • .rezero/memory/subaru-deaths.md is deleted after commit.

6. Rem

   • Rem warnings are normal Re:ZERO attempts.
   • They must be implemented, verified, evaluated by witches, and committed with no fail.
   • When all warnings are resolved and accepted, .rezero/memory/rem.md is deleted.

Skills

• rezero-init — detects the stack and sets up witch evaluation tools for /rezero init.
• rezero — /rezero entrypoint; coordinates the full loop.
• rezero-plan — splits large requests into small ordered tasks.
• rezero-subaru — Subaru's single-task implementation loop.
• rezero-witches — dispatches seven fresh-context witch reviews and prints the verdict table.
• rezero-rem — stores, resolves, and deletes warning memory.

Language and Names

Re:ZERO answers in the user's language when supported. Character names in witch verdicts and parallel implementer labels follow the same language. Unsupported languages fall back to English.

Language	Witches	Parallel implementers
English / Spanish / Portuguese / German	Echidna, Typhon, Minerva, Daphne, Carmilla, Sekhmet, Satella	Beatrice, Emilia, Ram, Garfiel, Julius
Korean	에키드나, 티폰, 미네르바, 다프네, 카밀라, 세크메트, 사테라	베아트리스, 에밀리아, 람, 가필, 율리우스
Japanese	エキドナ, テュフォン, ミネルヴァ, ダフネ, カーミラ, セクメト, サテラ	ベアトリス, エミリア, ラム, ガーフィール, ユリウス
Simplified Chinese	艾姬多娜, 堤丰, 弥涅耳瓦, 达芙妮, 卡密拉, 塞赫麦特, 莎缇拉	碧翠丝, 艾米莉娅, 拉姆, 加菲尔, 尤里乌斯
French	Echidna, Typhon, Minerva, Daphné, Carmilla, Sekhmet, Satella	Béatrice, Émilia, Ram, Garfiel, Julius
Russian	Ехидна, Тифон, Минерва, Дафна, Кармилла, Сехмет, Сателла	Беатрис, Эмилия, Рам, Гарфиэль, Юлиус

Concepts

Natsuki Subaru

Subaru is the implementer.

• Starts from current HEAD.
• Implements and verifies the task.
• Calls the witches after work is complete.
• If the route fails, keeps only the memory needed to avoid the same failure.

Return by Death

Return by Death resets the working tree but preserves failure memory.

git reset --hard HEAD
git clean -fd

The code dies. The lesson survives.

Seven Witches

The witches are independent reviewers. They do not inherit Subaru's reasoning, plan, self-assessment, or prior failed route unless required as evidence.

Witch	Focus	Example tools
Echidna (Greed)	Completeness, edge cases, coverage, quality gates	SonarQube/SonarCloud, coverage, Stryker
Typhon (Pride)	Contracts, specs, public interfaces, user intent	typecheck, linter, Spectral, Pact
Minerva (Wrath)	User harm, regressions, runtime failures	full tests, Playwright, Lighthouse CI, k6
Daphne (Gluttony)	Dependency/resource appetite	OSV-Scanner, Knip, source-map-explorer, hyperfine
Carmilla (Lust)	Deceptive UI/docs/names/proof	Playwright screenshots, axe, lychee
Sekhmet (Sloth)	Maintainability, dead code, duplication, complexity	SonarQube/SonarCloud, Knip, jscpd
Satella (Envy)	Integration, security, policy, project consistency	CodeQL, Gitleaks, Trivy, full CI

Verdicts:

• pass — accepted.
• warning — accepted, recorded by Rem.
• fail — rejected; triggers Return by Death.

Rem

Rem is warning memory.

• Warnings that pass review are stored in .rezero/memory/rem.md.
• They remain until fixed, re-evaluated by witches, accepted, and committed.
• When no warnings remain, rem.md is deleted.

License

This project is distributed under the MIT License.