Cascadeflow OpenClaw Router (Fast Start)

This skill tags OpenClaw requests so the Cascadeflow gateway can apply fast drafter + verifier routing, with optional failover and optional domain-aware routing. It is designed to be low-latency and easy to adopt.

Fast Start (1 minute)

1. •Configure Cascadeflow with:
   • •drafter model (fast/cheap)
   • •verifier model (your current best model)
2. •Use this skill. It works immediately without any extra routing config.

Channel Model Setup (Cascadeflow-side)

Define the core channels in your Cascadeflow OpenClaw config:

• •drafter: required (fast/cheap model)
• •verifier: required (quality model)
• •failover: optional (backup provider/model)
  • •If not set: fallback = drafter, then verifier

Optional OpenClaw-native channels (opt-in, explicitly tagged in this skill; classifier can also map when enabled):

• •heartbeat (system/heartbeat flows)
• •cron (scheduled jobs / cron events)
• •voice (tts/voicewake)
• •image_understanding (image attachments)
• •web_search (web_search/web_fetch when user enables)
• •brain, coding, content (OpenClaw categories)

Optional Cascadeflow domains (opt-in, auto-detected by Cascadeflow):

• •code
• •data
• •structured
• •rag
• •conversation
• •tool
• •creative
• •comparison
• •summary
• •translation
• •math
• •factual
• •medical
• •legal
• •financial
• •multimodal
• •general

Note:

• •OpenClaw categories can map to Cascadeflow domains if you want:
  • •coding -> code
  • •image_understanding -> multimodal
  • •brain -> general
• •Cascadeflow domains are handled by Cascadeflow's automatic domain routing.
• •Explicit tags are optional and only needed if you want to force a specific domain.
• •Suggested routing: OpenClaw-native via explicit tags (or classifier), Cascadeflow domains via auto-detection.

Example:

openclaw:
  channels:
    drafter: "claude-3-5-haiku-20241022"
    verifier: "claude-3-5-sonnet-20241022"
    failover: "gpt-4o-mini"   # optional

    # Optional OpenClaw-native channels
    heartbeat:
      models: "claude-3-5-haiku-20241022"
      strategy: direct_cheap
    cron:
      models: "claude-3-5-haiku-20241022"
      strategy: direct_cheap
    voice:
      models: "gpt-4o-realtime"
      strategy: direct_best
    image_understanding: "gpt-4o"
    web_search: "claude-3-5-haiku-20241022"
    brain: "claude-3-5-sonnet-20241022"
    coding: "claude-3-5-opus-20240229"
    content: "claude-3-5-sonnet-20241022"

    # Optional Cascadeflow domains (explicit tags optional)
    code: "claude-3-5-opus-20240229"
    data: "gpt-4o"
    structured: "gpt-4o-mini"
    rag: "gpt-4o"
    conversation: "claude-3-5-sonnet-20241022"
    tool: "claude-3-5-sonnet-20241022"
    creative: "claude-3-5-sonnet-20241022"
    comparison: "claude-3-5-sonnet-20241022"
    summary: "claude-3-5-haiku-20241022"
    translation: "claude-3-5-haiku-20241022"
    math: "deepseek-r1"
    factual: "gpt-4o"
    medical: "claude-3-5-opus-20240229"
    legal: "claude-3-5-opus-20240229"
    financial: "gpt-4o"
    multimodal: "gpt-4o"
    general: "claude-3-5-sonnet-20241022"

Tested Setups (OpenClaw E2E)

These configs are validated starting points. Replace model IDs with your provider names.

1) Anthropic-only

2) OpenAI-only

3) Mixed (Best of Both)

Domain Routing Notes

• •OpenClaw-native categories are routed in two ways:
  1. •Explicit tags from this skill (deterministic, recommended).
  2. •OpenClaw pre-router classifier (method/payload hints only, opt-in).
• •Cascadeflow domains are auto-detected by Cascadeflow logic.
• •If no tag/classifier match, Cascadeflow handles the request dynamically.
• •Suggested routing: OpenClaw-native via explicit tags; Cascadeflow domains via auto-detection.
• •Heartbeat/cron channels default to direct cheap when a channel model is configured.

Metrics & Stats (OpenClaw-side)

Cascadeflow exposes runtime metrics at GET /stats on the OpenAI server. Use this to validate savings, acceptance rate, tool usage, and latency.

Suggested commands:

• •/cascade stats -> summarize cost savings + acceptance rate
• •/cascade domains -> per-domain performance
• •/cascade latency -> overhead analysis
• •/cascade savings -> cost comparison vs verifier-only baseline

Example:

curl http://127.0.0.1:8084/stats | jq .

Savings fields (from /stats.summary):

• •total_cost: actual spend
• •baseline_cost: verifier-only baseline (actual + saved)
• •total_saved: baseline - actual
• •savings_percent: savings / baseline
• •draft_tokens, verifier_tokens, total_tokens (if providers report tokens)

Example output (formatted in the skill):

📊 CascadeFlow Savings Report
━━━━━━━━━━━━━━━━━━━━━━━━━━━
Total Queries: 1,247
Draft Acceptance: 64%

💰 Cost Comparison:
  Verifier-only: $18.72
  With Cascade:  $6.54
  ━━━━━━━━━━━━━━━━
  Savings:       $12.18 (65%)

Tagging Rules (What this skill sends)

This skill adds explicit routing hints only when the domain is predictable.

• •Use cascadeflow.category for OpenClaw-native categories.
• •Use cascadeflow.domain for Cascadeflow domains (e.g., comparison, factual, creative).
• •Optionally use cascadeflow.profile:
  • •best
  • •cost_savings
• •Optionally set cascadeflow.model to force a model.

Example (explicit domain routing):

{
  "cascadeflow": {
    "domain": "comparison",
    "profile": "cost_savings"
  }
}

Fallback Behavior (Preferred-if-Available)

1. •If category mapping exists and model is available -> use it
2. •Else use failover if set
3. •Else fallback to drafter, then verifier

When NOT to Tag

If the domain is dynamic or unclear, do not tag. Cascadeflow will handle it with domain-aware routing and cascading.

Safety Note

Skills run as executable logic. Review third-party skills before use.