claude-code-orchestrator

# Claude Code Orchestrator Use `claude-agent` when Claude Code should do implementation, frontend exploration, visual variant generation, or review work in the background while Codex keeps orchestration context small. ## Defaults - Model: `opus` - resolves to Opus 4.7 on first-party Anthropic. - Effort: `max` by default. Prefer `--xhigh` for bulk workbench variant sweeps. - Permission mode: `auto`. - Runtime: tmux session per agent. - State: `~/.claude-agent/jobs`. ## Standard Loop ```bash claude-agent start "Task prompt" --dir /path/to/repo --map --xhigh claude-agent await-turn <jobId> claude-agent capture <jobId> 160 --clean claude-agent send <jobId> "Follow-up instruction" claude-agent await-turn <jobId> claude-agent capture <jobId> 220 --clean ``` Use `claude-agent jobs --json` for compact status and `claude-agent attach <jobId>` when a human needs the live tmux session. ## Workbench Variant Prompt Shape Ask Claude Code to read only the compressed design inputs: ```text Read apps/workbench/AGENT_DESIGN_RULES.md, src/generated/theme-summary.md, and apps/workbench/src/variants/<context>/<brief>/brief.md. Create v1-v5 as distinct React/Tailwind variants that obey the variant contract. Run the workbench design check and fix violations. Report the files changed and the visual direction of each variant. ``` For bulk variants use `--xhigh`. For final convergence or hard review use default `max`. ## Local Service Use the service when another CLI needs to manage agents: ```bash TOKEN="$(claude-agent service token)" curl -sS -H "Authorization: Bearer $TOKEN" \ -H "content-type: application/json" \ -d '{"prompt":"Review this workbench brief","cwd":"/path/to/repo","effort":"xhigh"}' \ http://127.0.0.1:4877/jobs ``` Start it with: ```bash claude-agent service start ``` ## Rules - Keep prompts bounded. Link to maps, briefs, and generated summaries instead of pasting large files. - Use `await-turn`, then `capture --clean`; do not stream huge tmux logs into the main context. - Use `send` to continue the same agent after scoping so it keeps its local context. - Use `kill` only when an agent is stuck or clearly off-track.