Switching between models to control costs

[tall dust]

Can before_model_resolve fire after message:transcribed for the same message — enabling same-turn model switching based on message content?

I want to auto-switch to a more capable model (gpt-5.4) when a WhatsApp message contains travel-related content, without a one-message lag. My current thinking:

message:transcribed hook detects travel keywords, writes a flag file
before_model_resolve hook reads the flag, overrides model to gpt-5.4
Does the message pipeline sequence allow this — i.e. does before_model_resolve run after transcription for the same message? Or is there a better pattern for content-aware model switching?

Running OC 4.15, WhatsApp channel, OpenAI provider.

[final plover]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[formal ridge]

Short answer: don’t do the flag-file handoff from message:transcribed to before_model_resolve.

I checked the current source. The order is roughly:

1. media understanding/transcription updates the finalized inbound context
2. OpenClaw emits message:transcribed and message:preprocessed
3. the agent prompt is built
4. before_model_resolve runs before resolveModel()

But the important bit: message:transcribed is fired via a fire-and-forget internal hook. So although it is scheduled before the agent run, OpenClaw does not await your handler before model resolution. A flag file can race and give you flaky same-turn behavior.

Better pattern: put the routing logic directly in before_model_resolve and inspect event.prompt. By the time that hook runs, the prompt should already include the preprocessed/transcribed user content for that turn, so you can keyword/classifier-route there and return modelOverride / providerOverride directly.

So for same-turn content-aware switching:

[formal ridge]

api.on("before_model_resolve", async (event) => {
  const text = event.prompt.toLowerCase();

  if (/\b(flight|hotel|itinerary|travel|airport)\b/.test(text)) {
    return {
      providerOverride: "openai",
      modelOverride: "<your higher-cost travel model>",
    };
  }
});

Use message:transcribed / message:preprocessed for observation, logging, side effects, or metrics, not as a dependency for model selection.

Docs:
https://docs.openclaw.ai/automation/hooks

[formal ridge]

https://docs.openclaw.ai/concepts/agent-loop
https://docs.openclaw.ai/nodes/audio

[tall dust]

Could you share a minimal working plugin skeleton that uses api.on("before_model_resolve", ...)?

We have the hook logic but need the surrounding plugin structure — package.json, how to import api, and the install command. Just the bare minimum to get it running locally.

[formal ridge]

Yep. Minimal native hook plugin shape:

travel-model-router/
├── package.json
├── openclaw.plugin.json
└── index.ts

// package.json
{
  "name": "travel-model-router",
  "version": "0.1.0",
  "type": "module",
  "openclaw": {

[formal ridge]

    "extensions": ["./index.ts"]
  }
}

// openclaw.plugin.json
{
  "id": "travel-model-router",
  "name": "Travel Model Router",
  "description": "Routes travel-related turns to a higher-capability model.",
  "configSchema": {
    "type": "object",
    "additionalProperties": false,
    "properties": {}

[formal ridge]

  }
}

// index.ts
import { definePluginEntry } from "openclaw/plugin-sdk/plugin-entry";

export default definePluginEntry({
  id: "travel-model-router",
  name: "Travel Model Router",
  description: "Routes travel-related turns to a higher-capability model.",

  register(api) {
    api.on("before_model_resolve", async (event) => {

[formal ridge]

      const text = event.prompt.toLowerCase();

      if (/\b(flight|hotel|itinerary|airport|travel|trip)\b/.test(text)) {
        api.logger.info("travel terms detected; overriding model");

        return {
          providerOverride: "openai",
          modelOverride: "<your-travel-model-id>"
        };
      }
    });
  }
});

Install locally during dev: