openclaw - 💡(How to fix) Fix feat: keyword-based message routing for pre-turn model selection [1 participants]

Summary

Add a zero-cost pre-turn model routing mechanism that selects the LLM based on keyword/pattern matching in the incoming message, before any LLM call is made.

Motivation

Different tasks benefit from different models:

• Code review / PRs → high-capability coding model (e.g. claude-sonnet)
• Research / web search → large context model (e.g. gemini-flash)
• Quick Q&A / config → fast local model (e.g. qwen35)
• Legal / patent → high-reasoning model (e.g. opus)

Currently switching models requires a manual /model command. This feature enables automatic routing based on message content with zero additional cost.

Proposed Design

Extend the existing model.tasks infrastructure with a new messageRouting config:

{
  agents: {
    defaults: {
      model: {
        primary: "vllm/qwen35",
        tasks: {
          messageRouting: {
            rules: [
              { match: ["code review", "PR", "diff", "refactor"], model: "github-copilot/claude-sonnet-4.6" },
              { match: ["research", "search", "find"], model: "google-gemini-cli/gemini-3-flash-preview" },
              { match: ["patent", "legal", "prior art"], model: "anthropic/claude-opus-4-6" },
            ],
            default: "vllm/qwen35",
          },
        },
      },
    },
  },
}

• First-match-wins rule ordering
• Case-insensitive substring matching
• Zero LLM cost (pure string matching, no classification call)
• Falls back to agent primary model if no rule matches
• Configurable per-agent or at defaults level
• Supports model aliases and full provider/model strings
• Can also be specified in AGENTS.md for documentation/awareness

Implementation (prototype in fork)

A working implementation is in PR (see below) with:

• resolveMessageRoutingModel(cfg, agentId, messageText) in agent-scope.ts
• MessageRoutingRule, MessageRoutingConfig types in types.agents-shared.ts
• 5 unit tests (all passing)