convez.ai
Skill
eternalwaitt

eternal-stack

eternalwaitt/eternal-stack

Ver no GitHub
InicialPluginScore · de 0 a 100

Deterministic hooks, skills, and install profiles for Claude Code

Instalação

$npx convez-ai eternalwaitt/eternal-stack

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 semanas
Versão
v0.13.0
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Eternal Stack

Hooks and skills that make Claude Code actually finish what it starts.

I got tired of AI agents that say "done" when they aren't. That mark tasks complete without verifying. That skip the plan you approved and start improvising halfway through. That drift silently until something breaks.

Eternal Stack is my answer to that. It's a set of hooks, skills, and install profiles for Claude Code that enforce the habits that matter — read before editing, verify before claiming done, finish the plan you agreed to — and make them impossible to quietly skip.

I run this in production every day to build real software. It's opinionated by design.

Current release: VERSION — CHANGELOG


The problems it solves

Every Claude Code power user eventually hits the same wall: the agent is smart but it isn't disciplined. It wants to be helpful so badly that it'll declare victory prematurely, silently switch approaches mid-task, or drift away from what you actually agreed to build.

The thing that keeps happeningWhat Eternal Stack does about it
Agent says "done", nothing is verifiedStop hooks block completion claims without passing a checklist
Skills installed locally drift from sourceskill-contract-check.mjs and install metadata detect drift on every session
Thin plans get shipped to execution without pushbackplan-readiness-check.mjs gates any plan that's missing scope, risks, or rollback steps
Parallel edits spiral out of controlTask packets, wave overlap checks, and write-scope enforcement keep changes bounded
Settings get overwritten on reinstallMerge-in-place hooks, settings audit, and rollback backups protect your config

Quick start

git clone https://github.com/eternalwaitt/eternal-stack.git
cd eternal-stack
./scripts/install.sh --profile core
./scripts/doctor.sh

That's it. The core profile is the right default for almost everyone — it installs the observer stack, repo-owned etrnl-* agents, and verification tests. The full profile adds optional CodeGraph, Beads, and Hindsight companions for heavier setups.

See docs/install.md for profiles, strict mode, rollback, and migration.

Install writes ~/.claude/etrnl/install.json and ~/.codex/etrnl/install.json so each host can detect source drift. Local auto-update from your checkout is on by default; set ETRNL_AUTO_UPDATE=0 to opt out.

~/.claude/scripts/update.sh    # or ./scripts/update.sh from the repo
./scripts/doctor.sh            # optional: ./scripts/doctor.sh --jobs 8
tests/test-hooks.sh

What ships with it

Hooks — the enforcement layer. PreToolUse guards, PostToolBatch observer, prompt routing, compact recovery, stop verification, sycophancy blockers, port guard, and more. tests/test-hooks.sh exercises 85+ fixtures. These run automatically; you don't think about them.

Skills — the orchestration layer. Repo-owned /etrnl-* commands for planning, execution, audits, CI, commits, PRs, and operations. They call each other cleanly and stay namespaced so they don't collide with anything else you've installed. Full inventory in docs/skills.md.

Scripts — deterministic helpers for ledgers, browser QA, workflow health, code-health inventory, deep-audit validation, and release hygiene. Boring in the best way — they do one thing and always do it the same way.

Agents — bounded subagents (etrnl-executor, reviewers, scout, adversary, browser-QA) installed to ~/.claude/agents/. They have narrow scopes and write limits by design.


Profiles

ProfileWhat it includes
coreHooks, skills, agents, doctor, rollback, and ETRNL state. Start here.
fullEverything in core, plus CodeGraph/Beads bootstrap paths and Hindsight canaries

Want the hard blockers — the ones that refuse to let an agent proceed at all? Enable strict mode after you've run doctor, passed hook tests, rehearsed rollback, and done a smoke pass:

ETRNL_ENABLE_STRICT=1 ./scripts/install.sh

Don't skip those steps. Strict mode with untested hooks will interrupt things you didn't want interrupted.


Documentation

DocWhat it covers
AGENTS.mdAgent and contributor rules — how this repo works
CLAUDE.mdThin Claude Code wrapper that imports AGENTS.md
docs/install.mdInstall, update, uninstall, profiles, strict mode
docs/skills.mdetrnl-* orchestration and bundled skill inventory
docs/health-stack.mdCode and documentation health gates
docs/eternal-stack-coverage.mdCapability coverage map
docs/RELEASING.mdMaintainer release workflow
CREDITS.mdVendored and inspirational sources

Safety

Emergency bypass — use it when something is legitimately blocked and you know why:

export CLAUDE_GUARD_DISABLED=1

This repository is public-safe. Private identity, accounts, credentials, transcripts, local memories, plans, and background notes don't belong in tracked files and won't be here.


Acknowledgments

To my mentor and friend @PierreAndreis: thank you. You have been a steady source of inspiration in how you build, how you teach, and how you treat people along the way. Eternal Stack carries more of your influence than any changelog line could capture, and I'm grateful for it.

Credits

Inlined reference modules (Brooks, oRPC, Prisma, SQL), bundled skill attribution, and the projects that inspired specific patterns are documented in CREDITS.md.

Ver o original no GitHub

Bom para

Relacionados