Claude Codex Coordination
sun9huni/claude-codex-coordination
Coordenação multi-agente via SSOT em arquivos, enforcement hooks e handoffs validados
Ver no GitHubUm harness padronizado e clonável para projetos onde Claude Code, Codex e Cursor compartilham o mesmo diretório. Implementa sincronização via arquivos (status, contracts, handoffs) com hooks de enforcement, slash skills nativos e validação de estado entre agentes.
Instalação
npx convez-ai sun9huni/claude-codex-coordination1 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
Casos de uso
- Projetos com múltiplos agentes IA (Claude, Codex, Cursor) trabalhando simultaneamente
- Workflows com handoffs de tarefas entre diferentes ferramentas de código
- Manutenção de estado único e sincronizado em ambientes colaborativos
Veredito

Solução robusta para equipes que precisam coordenar múltiplos agentes IA em um único workspace, eliminando drift entre agentes via policies-as-files e estado derivado. Requer ~30min de setup manual, mas reduz drasticamente fricção em projetos multi-agente.
README
A tradução para o português deste README ainda está na fila.
claude-codex-coordination
A standardized, cloneable harness for projects where Claude Code, Codex, and/or Cursor share the same working directory.
This is not a Claude Code plugin — it is a file layout + policy + enforcement convention that any of those agents can pick up by reading its entry document. It works because the coordination layer lives in files (CURRENT.md, status, contracts) rather than chat history.
What you get
- One session-start ritual shared by all three agents (CLAUDE.md / AGENTS.md / WORKFLOW.md / takeover-prompt.md cannot drift apart).
- Schema-validated handoff state in per-slice
.agent/status/<slice>.mdbatons (yaml frontmatter the next agent can parse without trusting chat) plus a derived.agent/handoffs/CURRENT.mdindex over them. - Enforcement hooks (
PreToolUse,SessionStart,Stop) that turn prose policies into actual blocks — force pushes, hard resets, rm -rf on shared storage, etc. - Native primitives wired in:
.claude/skills/ships 11 slash skills (4 process + 4 expertise + 3 workflow-chain),.claude/agents/for restricted-tool subagents,.claude/hooks/for enforcement and productivity (optionalpost-edit-format.sh). - SessionStart additionalContext injection: bootstrap state + detected inventory pushed into context at turn 0; no reliance on CLAUDE.md being re-read.
- Spec → Plan → Execute chain (
/brainstorm,/write-plan,/execute-plan): obra/superpowers-style workflow discipline plugged into the existing.agent/contracts/lifecycle. - Slice + routing system so multi-workstream projects don't blur boundaries between work areas.
- Memory policy that prevents the auto-memory private-context leak
(project state belongs in
.agent/, not in any one agent's memory).
What this is not
- Not a Claude Code marketplace plugin. No install step.
- Not a magic generator. You will spend ~30 minutes filling in the slices, contracts, and routing table for your project.
- Not opinionated about your tech stack — this is process scaffolding.
Quick start
# 1. Clone into your project (or an empty workspace dir)
git clone https://github.com/sun9huni/claude-codex-coordination.git my-project
cd my-project
rm -rf .git
git init
# 2. Fill in the placeholders
# - CLAUDE.md / AGENTS.md / WORKFLOW.md: replace <project-name> and
# describe your slices in the routing table
# - ./scripts/init-slice.sh <slice>: scaffolds one
# .agent/status/<slice>.md baton + harness file per slice; fill
# in each baton's goal and first remaining_actions
# - .agent/projects/<slice>-harness.md: deep workflow detail per slice
# 3. (Optional) enable the optional hooks for your stack
# cp .claude/hooks/optional/pre-bash-slurm-gate.sh .claude/hooks/ # if HPC
# cp .claude/hooks/optional/pre-bash-db-gate.sh .claude/hooks/ # if DB
# Then add them to .claude/settings.json hooks block.
# 4. (Optional) port the example agents
# See examples/research-deployment/.claude/agents/
# 5. First handoff to create a baseline (claims the slice baton and
# regenerates the derived CURRENT.md index)
./scripts/handoff.sh claude <slice> # or codex / cursor
# 6. Commit
git add -A && git commit -m "harness: initial setup from claude-codex-coordination"
Then open the workspace in Claude Code (or Codex / Cursor) — the
session-start hook fires, the agent reads CURRENT.md, and you are
ready to work.
Architecture in one paragraph
The workspace root holds three entry documents (CLAUDE.md, AGENTS.md,
WORKFLOW.md) that point at one source of truth: the per-slice
.agent/status/<slice>.md batons, indexed by the derived
.agent/handoffs/CURRENT.md. Each baton's Markdown body is
human-readable; its yaml frontmatter is machine-validated. .claude/
contains Claude's native primitives — hooks that block dangerous Bash,
skills that are real slash commands, subagents that delegate with a
narrow tool surface. .agent/ is the cross-agent state: handoffs,
slice status, contracts, project harness files. When the session ends,
./scripts/handoff.sh <next-agent> <slice> refreshes the baton and the
index (no-slice mode additionally snapshots git state into
.agent/handoffs/state/) so the next agent — human, Claude, Codex, or
Cursor — resumes from files, not chat history.
See docs/design.md for the longer story.
Documentation
- HARNESS_USAGE.md — day-to-day reference
- docs/design.md — why each piece exists
- docs/customization.md — adapting to your project
- docs/concepts/ — deeper dives on slices, handoffs, hooks, memory, multi-agent coexistence
- examples/research-deployment/ — a filled-in example for an HPC/research project
License
MIT — see LICENSE.
Status
v0.1.0 — initial release. The conventions have been used in
production by one research workspace (8 phases of design captured in
docs/design.md). Expect rough edges; please open
issues or PRs.