Imperial Orchestrator

Use this skill when the user wants OpenClaw to coordinate many models intelligently instead of relying on a single default model.

This skill implements a pragmatic version of a 三省六部 / multi-role court pattern:

• - 中枢总管: receives the request, classifies it, selects an execution mode, and keeps the session alive
• 内阁规划: breaks down larger tasks into sub-tasks
• 六部执行: coding, ops, security, legal, writing, finance, project management
• 都察院审核: optional review pass when quality or safety matters
• 生存模式: if providers fail, keep working with the best remaining model or a local fallback

Goals

1. 1. Discover all available models from local config and optional health snapshot files.
2. Match tasks to the best role and model set.
3. Avoid 401 retry storms by opening an auth circuit breaker.
4. Prefer cross-provider fallback over same-provider fallback.
5. Keep the main session alive by switching to degraded mode if necessary.

Important reality checks

• - OpenClaw loads skills from bundled skills, ~/.openclaw/skills, or <workspace>/skills. Workspace skills have the highest precedence.
• Skills are AgentSkills-compatible folders centered on SKILL.md with YAML frontmatter.
• OpenClaw supports agent-specific workspaces and community usage shows coding-agent delegation, but model overrides and sub-agent routing can still be inconsistent depending on provider/version.
• Therefore this skill is intentionally built as a hybrid of policy + scripts instead of assuming every runtime route works natively.

What this skill contains

• - config/model_registry.yaml: model capability tags and routing preferences
• INLINECODE4: role definitions with forbidden actions and model strategies
• INLINECODE5: deep system prompts for each role (identity, rules, red lines)
• INLINECODE6: deterministic routing hints
• INLINECODE7: auth breaker, retry, degrade rules
• INLINECODE8: standardized test tasks per category (coding/writing/reasoning/ops/security/finance)
• INLINECODE9: main planner/router/selector (now with benchmark-aware scoring)
• INLINECODE10: discover models from openclaw.json
• INLINECODE11: probe each model with real API calls to verify availability
• INLINECODE12: run the same task against all models, score and rank by category
• INLINECODE13: print machine-readable routing state
• INLINECODE14: unified CLI entry point
• INLINECODE15: sample config patches and install docs

三省六部角色体系

每个角色配备深度 system prompt（定义在 config/agent_prompts.yaml），包含身份认同、职责边界、行为准则、协作意识和生死线五个维度。

中枢

三省

六部

急递铺

Operating rules

1) Treat 401 as toxic

If a model fails with an auth error:

• - mark the model as INLINECODE17
• cool down the whole auth chain / provider profile
• do not keep retrying sibling models on the same auth chain first
• prefer a different provider or a local fallback

2) Main router must stay light

Never assign the largest prompt, heaviest workspace, or the most fragile provider to the router-chief.

3) Cross-provider fallback first

Fallback order should be:

1. 1. same role, different provider
2. same role, local fallback
3. secondary role with adjacent duties
4. emergency-scribe

4) Degrade instead of dying

If the best specialist models are unavailable:

• - still answer with architecture, checklists, pseudocode, or a reduced plan
• do not return All models failed if any model remains viable

Suggested workflow

统一入口（推荐）

CODEBLOCK0

手动操作

CODEBLOCK1

Output contract

The router returns JSON including:

• - mode: direct / planthenexecute / multi_agent / degraded
• INLINECODE20 + lead_title (角色官衔)
• INLINECODE22 (深度角色提示词，可直接注入 agent)
• INLINECODE23 + INLINECODE24
• INLINECODE25 (角色禁止行为列表)
• INLINECODE26
• INLINECODE27
• INLINECODE28
• INLINECODE29 (含 benchmark 分数信息)

Integration notes

This skill does not hard-code one provider API. Instead it prepares routing decisions and failure state so you can:

• - feed the selected model into your OpenClaw agent config
• spawn the right sub-agent
• call your own shell wrapper around provider endpoints
• keep a machine-readable health snapshot for external automation

Install

Workspace-only

Copy this folder to:

CODEBLOCK2

Global shared install

Copy this folder to:

CODEBLOCK3

Safety

• - Never send secrets into prompts.
• Keep auth probe payloads minimal.
• Treat provider and token health as separate state from model quality.
• Do not assume a model is safe to route merely because it appears in config.