Playbook + CLI · Claude Code × Codex
One brain, two agents
Use the same second brain with Claude Code and Codex: the right agent and model for each task, manuals and skills always in sync, and usage tracked. Pure bash, no dependencies.
Use the same second brain with Claude Code and Codex: the right agent and model for each task, manuals and skills always in sync, and usage tracked. Pure bash, no dependencies.
Your second brain (the astra-2cerebro kit’s Markdown folder) can be read by Claude Code and Codex. Each one enters through a different door: CLAUDE.md e .claude/skills on one side, AGENTS.md e .agents/skills on the other side. cerebro-duplo keeps the interfaces the same and decides which one to use.
duplo.sh verificar checks that AGENTS.md = CLAUDE.md, that each skill is mirrored byte for byte, and that each one has the openai.yaml from Codex. sync fixes it. The skill duplo-verificar reports and proposes changes without making them on its own.
The skill qual-agente reads the request, chooses the agent, model, and effort level using a table (MCP configured, repo open, heavy reasoning, low-cost routine), and returns the command. Profiles .env store the model and effort level for each task type.
uso registrar logs each session in registro-uso.md; uso resumo totals by agent and profile. No made-up prices: minutes and sessions here, money in the provider’s pricing table.
A short cycle, repeated in each session. The model and effort flags are the actual ones from both CLIs (read from --help): --model/--effort in Claude Code, -m/-c model_reasoning_effort in Codex.
Specific MCP → the agent that has it. Code → where the repo is open. Audit, architecture, large wiki → raciocinio. Find, summarize, draft → rotina.
duplo.sh claude raciocinio prints and executes cd cérebro && claude --model fable --effort high. --so-mostrar only prints.
Did you edit the manual or a skill? sync rewrites the mirror from the canonical version (AGENTS.md, .claude/skills). Never deletes; warns.
One line per session: agent, profile, model, effort, minutes, task. On Thursday, uso resumo shows where the quota went.
One astra-2cerebro brain and bash. Agents are only needed when starting a session.
Folder created by the astra-2cerebro kit, with AGENTS.md, CLAUDE.md, .claude/skills/ e .agents/skills/.
# if you don't have one yet git clone https://github.com/inematds/astra-2cerebro.git # follow the kit's INSTALAR.md
Linux, macOS, Git Bash, or WSL. diff, awk, cp. No Node, Python, or packages.
bash --version # 4+ diff --version awk --version
Optional. verificar, sync e uso work without either one.
claude --version # tested: 2.1.263 codex --version # tested: 0.153.3
Real commands. Replace ~/meu-cerebro through your brain’s folder.
Clones the repo and copies the two skills to .claude/skills e .agents/skills, e o duplo.sh (with perfis/ e modelos/) to scripts/. Runs the verificar at the end.
git clone https://github.com/inematds/cerebro-duplo.git cd cerebro-duplo bash instalar.sh ~/meu-cerebro
Identical manual, mirrored skills, openai.yaml present, valid manual routes. Exit code 1 if there’s a PROBLEM; a broken route is just a WARNING (the /iniciar creates multiple files afterward).
cd ~/meu-cerebro bash scripts/duplo.sh verificar # OK: AGENTS.md and CLAUDE.md are identical # OK: 10 identical skills on both sides # == result: 0 issue(s), 6 warning(s)
The model names included in the repo are for this machine. Edit scripts/perfis/*.env. Trusted sources: claude --help (aliases fable, opus, sonnet; effort low…max) and your ~/.codex/config.toml.
bash scripts/duplo.sh perfis # code claude: claude-opus-5-5/high codex: gpt-6-astra/high # standard claude: sonnet/medium codex: gpt-6-astra/medium # reasoning claude: claude-opus-5-5/high codex: gpt-6-astra/high # routine claude: sonnet/low codex: gpt-6-astra/low nano scripts/perfis/padrao.env
Inside any session. The skill responds in four lines: agent · profile · model/effort, two reasons, the command, and the log entry for the end.
# in Claude Code /qual-agente preciso auditar o cérebro, faz um mês que não rodo # in Codex $qual-agente preciso auditar o cérebro, faz um mês que não rodo # typical response: # codex · reasoning · gpt-6-astra/high # Auditing requires heavy reasoning: a cheap model inflates the score. # Codex because Claude Code has already used more of its quota this week. # bash scripts/duplo.sh codex raciocinio # when closing: bash scripts/duplo.sh uso registrar codex raciocinio 25 "auditoria mensal"
The script prints the command before running it. With --so-mostrar only prints (useful for checking the flags without using quota).
bash scripts/duplo.sh claude raciocinio --so-mostrar # command: cd ~/meu-cerebro && claude --model fable --effort high --name duplo:raciocinio bash scripts/duplo.sh codex rotina --so-mostrar # command: codex -C ~/meu-cerebro -m gpt-6-astra -c model_reasoning_effort=\"low\" bash scripts/duplo.sh codex raciocinio # actually opens
Always edit the canonical version (AGENTS.md, .claude/skills/...). O sync rewrites the other side. --fonte claude reverses the direction in the manual when the good edit is in CLAUDE.md.
bash scripts/duplo.sh sync # AGENTS.md → CLAUDE.md; .claude/skills → .agents/skills bash scripts/duplo.sh sync --fonte claude # CLAUDE.md → AGENTS.md bash scripts/duplo.sh verificar # 0 issue(s)
One line per session in registro-uso.md (created from the template the first time). The summary is what drives the quota rule: when the week is tight, step down one level or switch agents.
bash scripts/duplo.sh uso registrar codex raciocinio 25 "auditoria mensal" bash scripts/duplo.sh uso registrar claude rotina 8 "resumo reunião 06/09" bash scripts/duplo.sh uso resumo # | Agent | Sessions | Minutes | # | claude | 1 | 8 | # | codex | 1 | 25 | # total: 2 sessions, 33 minutes
The skill duplo-verificar runs the verificar and turns the output into a short report with one fix per item. It doesn’t fix anything on its own: it ends by asking if it can apply the changes.
/duplo-verificar # Claude Code $duplo-verificar # Codex
The complete version, with explanations and quota rules, is at docs/roteamento.md.
| Task | Agent | Profile |
|---|---|---|
| Uses a specific MCP (calendar, email, browser) | what has the MCP configured | padrao |
| Long code, refactoring, connection script | where the repo is already open | codigo |
| Audit, architecture, difficult decision, large wiki | either one (more quota available) | raciocinio |
| Find a document, summarize a meeting, draft an email | either one | rotina |
| Interview, discovery conversation | what you’d rather discuss | padrao |
| Scheduled routine, without supervision | what already has the scheduling | rotina |
Captured on this machine against test/fixture, a small brain for the kit.
One more line in the CLAUDE.md and a skill edited only in the mirror.
$ duplo.sh --dir cerebro verificar PROBLEMA: AGENTS.md e CLAUDE.md divergem (diff abaixo) +- linha só no CLAUDE.md PROBLEMA: skill 'rotina' difere entre .claude/skills e .agents/skills: Files .claude/skills/rotina/SKILL.md and .agents/skills/rotina/SKILL.md differ == resultado: 2 problema(s), 6 aviso(s) $ duplo.sh --dir cerebro sync manual: CLAUDE.md gravado a partir de AGENTS.md skills: 1 sincronizada(s) .claude/skills -> .agents/skills $ duplo.sh --dir cerebro verificar OK: AGENTS.md e CLAUDE.md são idênticos OK: 8 skills idênticas nos dois lados == resultado: 0 problema(s), 6 aviso(s)
bash test/testes.sh: 38 assertions covering verify, sync, usage, profiles, and installer.
PASS: verificar passa na fixture íntegra PASS: verificar falha quando AGENTS.md e CLAUDE.md divergem PASS: sync conserta o manual a partir de AGENTS.md PASS: verificar acusa skill sem cópia no espelho PASS: sync não apaga skill só no espelho PASS: uso resumo soma por agente (claude 2 sessões, 45 min) PASS: claude --so-mostrar imprime cd + claude --model --effort PASS: codex --so-mostrar imprime codex -C -m -c model_reasoning_effort PASS: duplo.sh instalado verifica o cérebro com as 10 skills ... == resultado: 38 passaram, 0 falharam
Version 1.0.0 completes the verify → route → record cycle. Everything else is optional and depends on actual usage.
duplo.sh with check, sync, claude/codex by profile, usage; skills qual-agente e duplo-verificar; 4 profiles; installer; 38 tests; 6 documents.verificar like pre-commit in the versioned brain (today it’s a snippet in docs/paridade.md).DUPLO_<AGENTE>_* in the profiles and cmd_<agente> in the script for any runtime that can read AGENTS.md (the path is in docs/migracao.md)./rotina from the kit that runs uso resumo every Friday and records the reading in rotinas/registro.md.