convez.ai
AgenteAutomaçãoAvançado
h1902y

Zuzuu

h1902y/zuzuu

Memória + conhecimento + guardrails evoluindo para agentes de código

Ver no GitHub
BomPluginScore · de 0 a 100

Agent que enriquece agentes de codificação (Claude Code, Gemini, OpenCode) com memória evoluindo, conhecimento, ações e guardrails, usando rastreamento OpenTelemetry host-agnóstico e sem dependências. Captura sessões, aplica gates de segurança e evolui faculdades automaticamente a partir de traces reais.

Instalação

$npx convez-ai h1902y/zuzuu

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
1
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 2 meses
Versão
—
Issues abertas
40
Contribuidores
2

Casos de uso

  • Enriquecer agentes de codificação existentes com memória e contexto persistente
  • Implementar gates de segurança e guardrails automatizados em fluxos de agentes
  • Rastrear e evoluir comportamentos de agentes a partir de sessões reais com OpenTelemetry

Veredito

Ferramenta sofisticada e bem-engineered para quem executa agentes de codificação em produção e precisa de observabilidade, segurança e aprendizado contínuo. Fase early (Observe consolidado, Serve robusto, Evolve testado mas não yet validado em corpus real), mas arquitetura sólida e zero-deps é impressionante.

README

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

zuzuu

ci npm node license

Give the coding agent you already run an evolving Memory, Knowledge, Actions, and Guardrails — grown from how you actually work.

Your host agent — Claude Code, Codex, Gemini CLI, OpenCode — supplies the brain (the reasoning loop + the model). zuzuu wraps the host you already pay for: it serves faculties to it, observes every session as an OpenTelemetry trace, and (the end-game) evolves the faculties from those traces — human-gated, across versioned generations. We never run a competing agent loop and never drive the host headlessly.

Install npm i -g @zuzuucodes/cli — the command is zz (or zuzuu). Published with provenance; releases auto-publish from main via GitHub OIDC.

Status (honest): early build, moving fast. Observe works (5 real hosts, verified). Serve delivers the faculty home (zuzuu init), a session digest to every host, an enforced guardrails gate on all 5, and five faculties sharing one proposal/review spine. Evolve is now wired and tested — trace miners → a mechanical eval lens → human-gated zuzuu review → versioned generations (mint / rollback / drift-check) — but not yet proven on a real graduation corpus (the loop runs + passes hermetic tests; it hasn't yet improved an agent from real sessions end-to-end). Full design: docs/DESIGN.md.

What works today

npm install -g @zuzuucodes/cli   # zero dependencies — installs the `zuzuu` command (short alias: `zz`)

# no coding agent yet? one command gives you a fully faculty-equipped one:
zuzuu code        # scaffold the faculty home, install + wire OpenCode, launch it (capture + gate + grounding)

zuzuu web         # prefer a visual app? the workbench: Home → review ceremony → embedded-terminal agent sessions

# already run Claude Code / Gemini / Codex / OpenCode / pi? wrap the one you have:
zuzuu init        # scaffold your project's agent home (.zuzuu/) — git-style, hidden like .git
zuzuu explain     # the 5 faculties + how graduation works (you're always in the loop)
zuzuu inbox       # what's pending your approval · zuzuu review to approve/reject
zuzuu capture     # turn your latest agent session into an OpenTelemetry trace
zuzuu trace --last
zuzuu enable [--host gemini-cli|codex|opencode|pi]   # live capture + the guardrails gate
zuzuu doctor      # health + lost-session reconciliation

zuzuu code is the bundled-host path (Stage 2): it detects OpenCode (installs it on first run, with your OK — never an npm dependency, the zero-dep policy holds), wires the zuzuu plugin, and launches the real opencode — we configure + launch, never fork or drive it.

Claude CodeGemini CLICodexOpenCodepi
post-hoc capture✅ rich✅ thin✅ rich✅ rich✅ rich
live capture✅ hooks✅ hooks✅ hooks¹✅ plugin✅ extension
guardrails gate✅ PreToolUse✅ BeforeTool✅ PreToolUse¹✅ tool.execute.before✅ tool_call

¹ Codex is interactive-only — codex exec (headless) fires no hooks (verified, v0.138.0), so live capture + gate work when you run Codex interactively; headless Codex still gets post-hoc zuzuu capture.

All five verified against real sessions — never fixtures; every host's live capture + gate was wired from real captured hook payloads and dogfooded end-to-end (docs/LOG.md exp-11 Gemini/Codex, exp-12 OpenCode/pi). Gate semantics are host-honest: deny hard-blocks everywhere; ask maps to a native prompt on Claude, defers to the host elsewhere.

Prerequisites: Node ≥ 22 — that's it. You need at least one supported agent you've already used, so a session exists to capture. (Hacking on zuzuu itself? git clone https://github.com/h1902y/zuzuu && cd zuzuu && npm link.)

zuzuu init behaves like git init: empty dir → scaffolds the agent home + AGENTS.md/CLAUDE.md; existing project → adds the home and injects a small delimiter-marked block into your existing instruction files (your text is never touched); already initialized → restores missing pieces only. The home is hidden like .git and self-explaining — a .zuzuu/ dir you can read and version in git (the only visible footprint is the managed block + three .gitignore lines; transparency lives in zuzuu status / explain / digest): .zuzuu/README.md (the explainer) · knowledge/ (verified facts) · memory/ (curated episodes) · actions/ (runbooks) · instructions/ (steering) · guardrails/ (enforced rules), plus generations/ (your checkpoints). Machine internals are dot-prefixed + git-ignored (.zuzuu/.traces/, .zuzuu/.live/). (Pre-2026-06-12 homes at agent/ migrate automatically on zuzuu init, or via zuzuu migrate --home.)

Live capture (zuzuu enable) is invisible by design: a minimal lifecycle hook set (Claude Code, Gemini CLI, Codex), a bus plugin (OpenCode), or an extension (pi) — each wrapped so it always exits 0 / fails open — it can never break your agent. The same hook carries the guardrails gate, applied in each host's own idiom (Claude/Codex hookSpecificOutput, Gemini {decision:"deny"}, OpenCode throws from tool.execute.before, pi returns {block:true} from tool_call). Most hosts emit no clean end-signal when a terminal is killed, so zuzuu doctor reconciles lost sessions afterward from the transcript still on disk (nothing lost).

Where your data lives: transcripts are read read-only; output is git-native in your repo — .zuzuu/sessions.json (small tracked index, each session linked to a commit) + .zuzuu/.traces/*.otlp.jsonl (local, git-ignored). Nothing is uploaded; no raw tool input/output on the trace (byte sizes only).

Verify / troubleshoot: npm test (hermetic) · npm run playground (⏭️ skip = that host isn't on your machine, not a failure) · zuzuu doctor (env + session health). "No host detected" → use a supported agent once in the repo, then retry.

The idea in one diagram

   the host agent (yours)          zuzuu
  ┌─────────────────────┐     ┌──────────────────────────────┐
  │ Cognition · Model · │ ◄── │ SERVE   faculties:           │
  │ Workspace           │     │   knowledge · memory ·       │
  │  (we never drive)   │     │   actions · instructions ·   │
  │                     │     │   guardrails (enforced)      │
  └──────────┬──────────┘     ├──────────────────────────────┤
             │ sessions       │ OBSERVE traces (OTel,        │
             └──────────────► │         git-native)          │
                              ├──────────────────────────────┤
                              │ EVOLVE  eval → propose →     │
                              │         human gate → new     │
                              │         generation  [design] │
                              └──────────────────────────────┘

Five faculties, each mapping onto a cognitive system — Knowledge (semantic: what's true), Memory (episodic: what happened), Actions (procedural: how to do things), Instructions (directive: who the agent is), Guardrails (protective: what it must not do — enforced on tool calls, fail-open). They improve across versioned generations, proposals mined from traces, always human-approved. That loop is the product; everything here is a step toward it.

Repo map

PathWhat
zuzuu/ + bin/zuzuu.mjsthe CLI — capture pipeline (zuzuu/capture/), live lifecycle, faculty home (product surface)
web/the visual workbench — nested project (daemon + React SPA), ships bundled inside the npm package
tests/hermetic unit + regression (npm test) + real-data smoke playgrounds (npm run playground)
docs/DESIGN.md (the canon) + LOG.md (the build journal) + inspiration/ (the research shelf)

How this is built (the method)

Prove on real data, record in the journal. Every capability is verified against real sessions/wire data (never invented fixtures) before it counts; the record lives in docs/LOG.md (append-only). Live experimentation happens in disposable project directories outside the repo, keeping the codebase clean. Built in public — day-by-day on X (@h1902y).

License & status

Personal project, early and changing daily. Issues/ideas welcome.

Ver o original no GitHub

Bom para

Relacionados