claude-relais

claude-relais is a Claude Code orchestration skill built on a simple split:

• Think (Claude): high-reasoning orchestration and task design (default: Opus opus-4.6)
• Build (Cursor): fast implementation with Cursor headless agents
• Judge (Guardrails): strict PLAN -> BUILD -> JUDGE contracts with stop/block safety checks

Heavy reasoning models are excellent for planning, but slower and more limit-intensive for direct implementation. claude-relais keeps Claude focused on thinking and uses Cursor for execution throughput.

For many users, this split delivers high sustained output at roughly the cost of two standard plans (around ~$40/month combined), depending on provider pricing, plan limits, and usage patterns.

How it works

1. ROUTER: Claude routes input:

• explicit envoi ... commands are executed directly; otherwise it checks relais/ state.

2. ONBOARD OR CONTINUE:

• fresh repo or missing roadmap -> guided onboarding (mode + PRD + roadmap); existing state -> next-step options.

3. PLAN: orchestrator defines one bounded task and scope.
4. BUILD: task is dispatched to Cursor headless agent (mandatory).
5. VERIFY/UPDATE: Claude verifies against git truth, updates roadmap/state, and loops.

Mode stop conditions:

• task: stop after one successful task
• milestone: stop when active milestone is done
• autonomous: continue until blocked/limit/signal

Requirements

• Claude Code CLI installed and authenticated (claude whoami)
• Git
• Bash
• Cursor CLI installed and authenticated (cursor agent whoami)

Install (one command)

git clone https://github.com/clementrog/claude-relais.git && cd claude-relais && ./scripts/install.sh

What this does:

• installs skill files into your Claude skills directory
• writes config.local.json with default model/builder
• runs preflight checks (Claude + Cursor required)
• does not require installing legacy envoi npm packages

Verify installation

./scripts/smoke.sh

You should see:

• CLAUDE_RELAIS_PRECHECK:PASS:summary:Preflight passed.
• CLAUDE_RELAIS_SMOKE:PASS:summary:Skill install and preflight checks passed.

First run in Claude Code

1. Restart Claude Code if it was already open.
2. Open your target repo.
3. Ask Claude to run with claude-relais orchestration for your task.
4. Keep task scope explicit and bounded per tick.

Configuration

Defaults:

• orchestrator model: opus-4.6
• builder mode: cursor (enforced; non-cursor builder tasks are blocked)
• destructive command gates on by default: deny_prefixes includes rm, sudo, git reset --hard, git checkout --, git clean -fd, git clean -fdx, mkfs, dd
• require_explicit_for_destructive: true
• preflight blocks tracked symlinks that resolve outside repo root

Permission note:

• do not auto-enable global --dangerously-skip-permissions at skill load
• use envoi autonomy --set fast for low-friction mode with guardrails

Override at install time:

./scripts/install.sh --model opus-4.6 --builder cursor

Custom install location:

./scripts/install.sh --dest ~/.claude/skills

Troubleshooting

• Missing Claude CLI/auth: run login flow, then ./scripts/preflight.sh
• Cursor auth warning: run cursor agent login then retry preflight
• Skill not picked up: ensure <skills-dir>/claude-relais/SKILL.md exists, then restart Claude Code
• Duplicate /claude-relais entries or odd behavior:

rm ~/.claude/commands/claude-relais.md

• Reinstall cleanly:

./scripts/uninstall.sh --yes
./scripts/install.sh --force

Reference docs

• SKILL.md
• references/how-it-works.md
• references/configuration.md
• references/troubleshooting.md