2818 Openclaw Integration Plan 1f5dcc77

OpenClaw Integration Strategy (Plan)

Goals

• Winning DX: fastest possible onboarding for OpenClaw users with minimal config.
• Winning UX: clear savings and latency improvements with zero surprises.
• Lean core: OpenClaw integration remains optional and isolated from default Cascadeflow flow.
• Low overhead: pre-routing decisions must add minimal latency.

Non-Goals (v1)

• No cost caps or budget enforcement at launch.
• No breaking changes to existing Cascadeflow routing APIs.

Design Principles

• One-minute start: set drafter + verifier and it works.
• Progressive disclosure: domain/channel routing is opt-in and additive.
• Deterministic where possible: predictable OpenClaw-native routes pre-routed.
• Dynamic where needed: Cascadeflow domain/complexity routing handles the rest.
• Separation of concerns: OpenClaw integration stays in its own module + skill.

Integration Choice (OpenClaw-side)

Recommended: Cascadeflow provider + optional OpenClaw skill (hybrid).

• Provider handles all OpenClaw LLM calls via Cascadeflow.
• Skill provides explicit tagging and quick UI enablement for domain channels.

Architecture Overview

1) OpenClaw Pre-Router (Opt-in)

Purpose: ultra-low-latency routing for predictable OpenClaw-native events.

• Uses method/payload hints only (no extra model call).
• Applies when OpenClaw integration is enabled.
• Explicit tags from the OpenClaw skill always override classifier decisions.

2) Cascadeflow Rule Engine (Replaces policy)

Purpose: centralized routing rules for tiers, KPIs, and domain-aware routing.

• Replaces policy feature; no legacy policy code retained.
• Centralizes: user tiers, profiles, budgets/limits (future), domain routing, KPI gates.
• Rules are backward compatible with existing config keys.

3) Cascadeflow Core Routing

Purpose: handle dynamic or ambiguous requests.

• Domain detection, tool complexity, cascade verification, and fallback logic.

OpenClaw User Flow (Winning DX)

1. Install Cascadeflow provider in OpenClaw.
2. Set two models: drafter + verifier.
3. Optional: add failover channel.
4. Optional: add OpenClaw-native or Cascadeflow domain channels when ready (via Cascadeflow config).

The default path always works with only drafter + verifier.

Channel Model Setup (Cascadeflow-side)

Required

• drafter (fast/cheap)
• verifier (quality)

Optional

• failover (backup provider/model)
  • If not set: fallback = drafter, then verifier

Optional OpenClaw-native channels (opt-in)

• heartbeat, cron, voice, image_understanding, web_search, brain, coding, content

Optional Cascadeflow domains (opt-in)

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

Routing Rules (Source of Truth)

OpenClaw-native routes

• Explicit tags: from OpenClaw skill.
• Classifier: pre-router uses method/payload hints (no model call).
• If no match: cascadeflow domain/complexity routing.
• Channel mapping: OpenClaw categories map to channel names (optional) so users can route trivial system events (e.g., heartbeat/cron) to a smaller model than the drafter.
• Per-channel strategy: allow strategy per channel; default heartbeat/cron to direct_cheap when a channel model is configured.

Cascadeflow domains

• Auto-detected by Cascadeflow domain routing.
• Explicit tags are optional and only needed to force a specific domain.

Rule Engine v1

Inputs

• Domain + confidence
• Complexity + tool call profile
• User tier / profile / KPI flags
• OpenClaw tags (if present)

Outputs

• routing_strategy: direct, cascade, or cascade_if_low_confidence
• preferred_channel: optional channel override
• reason: decision trace for debugging

Migration Targets

• Move user tiers + profile selection into Rule Engine.
• Move domain routing decision logic into Rule Engine.
• Keep actual domain detection and model selection in core routing.

Savings/Accuracy Targets (Acceptance Criteria)

For any OpenClaw use case with routing enabled:

• Savings target: >= 60%
• Accuracy target: >= 95%

If unmet:

• Investigate, engineer, and propose a validated solution before launch.

OpenClaw Skill (First-touch DX)

Purpose: fastest setup, explicit tags, and optional domain/channel routing.

• Provide guidance for quickstart and optional domain routes.
• Provide explicit tags for OpenClaw-native routes.
• Explain how to enable the OpenClaw pre-router classifier.
• Keep skill doc lean and aligned with OpenClaw docs format.

DX Risks & Mitigations

• Too many surfaces (tiers/profiles/rules):
  • Mitigation: a single quickstart path and progressive disclosure.
• Ambiguous domain tags:
  • Mitigation: explicit tag guidance + classifier overrides.

Implementation Plan

Phase 0 — Discovery

• Confirm OpenClaw protocol schema and event taxonomy.
• Validate where skill tagging hooks exist.
• Confirm OpenClaw “browser” behavior (content fetch vs. analysis).

Discovery findings (2026-02-04)

• Protocol source of truth: TypeBox schemas live in src/gateway/protocol/schema.ts; AJV validators in src/gateway/protocol/index.ts; server methods in src/gateway/server.ts. The generated JSON Schema is dist/protocol.schema.json; docs note the raw schema is usually published at raw.githubusercontent.com/openclaw/openclaw/main/dist/protocol.schema.json (verify availability). citeturn3view0turn8search1
• Gateway frames: Request/response/event frames are the canonical transport shape; connect is the first request. citeturn2view0
• Skills + ClawHub: Skills are folders with SKILL.md (plus supporting files), loaded from <workspace>/skills and ~/.openclaw/skills with workspace precedence. ClawHub is the public registry; CLI installs into ./skills by default. citeturn0search0turn0search1
• Browser behavior: The OpenClaw browser is an agent‑controlled, isolated Chromium profile with deterministic tab control, snapshots/screenshots, and action execution. Internally it uses CDP and Playwright for advanced actions. This is more than content fetch; treat browser workloads as potentially complex unless a deterministic fetch‑only path is confirmed. citeturn6search0

Phase 1 — Rule Engine v1

• Add RuleEngine + RuleDecision types.
• Migrate tiers/profiles/domain routing decisions to RuleEngine.
• Keep backwards compatibility with existing config.

Phase 2 — OpenClaw Pre-Router

• Implement method/payload-based classifier.
• Respect explicit tags from OpenClaw skill.
• Provide opt-in config to enable/disable classifier.
• Map OpenClaw category → channel when no channel is provided.

Phase 3 — Provider + Skill

• No OpenClaw code changes required: users configure a custom provider in OpenClaw to route LLM calls to Cascadeflow.
• OpenClaw skill for explicit tags and guidance.
• Failover channel support and docs.
• Optional Cascadeflow config file to map OpenClaw categories → dedicated channel models with per-channel strategies.
• Expose GET /stats on the OpenAI server for skill-driven savings/latency reporting.

Phase 3.5 — Multi-turn + Tool Reliability

• Detect multi-turn prompts even when full message history is not supplied.
• Boost conversation domain detection for multi-turn formatted prompts.
• Normalize OpenAI-format tool schemas into Cascadeflow’s universal format.
• Add regression tests for multi-turn detection and tool normalization.

Phase 4 — Validation

• E2E OpenClaw tests (latency, savings, accuracy).
• Must meet acceptance criteria; if not, block launch and iterate.
• Verify heartbeat/cron routing hits dedicated channels when configured.

Open Questions

• How OpenClaw exposes skill tagging and routing metadata (docs confirm skills are injected into prompts; explicit tag surface still to confirm). citeturn0search1
• Exact OpenClaw native categories and their stability (validate against schema + server methods list).
• Whether the published dist/protocol.schema.json exists on GitHub or is only available in releases (docs suggest a raw link but availability may vary). citeturn3view0turn8search1
• Any OpenClaw constraints around provider/skill execution on Pi vs Mac mini.