Asking other models
A run can ask any model your account has a credential for. Ask in plain words, in standing instructions, an automation's prompt, or a comment:
Before posting a review, send each finding to gpt-sol and grok and ask them to refute it. Drop any finding you cannot re-verify after a refutation.The model you ask works as a helper of the run:
- It has the run's tools, including its shell, and its edits land in the same checkout.
- Pushing, posting to GitHub, and the run's own settings stay with the run that asked.
- Its answer goes back to that run, which decides how to act on it. Nothing posts a finding because another model agreed with it, and nothing drops one because another model disagreed.
A request can name an effort, such as gpt-sol at high effort. The level has to be one that model offers, and the console lists them. A model with no effort control, such as one on Bedrock or Vertex, refuses the request.
Which models a run can ask
The repository console lists them under Standing instructions → Models a run can ask, with the effort levels each one takes. The list covers credentials stored in Pullfrog:
- API keys.
- Connected subscriptions: Claude, ChatGPT, and Grok. A subscription pays before an API key for the same provider, the same order a run uses for its own model.
Other credentials work too:
- A key set as a GitHub Actions secret, which the console cannot see.
- Amazon Bedrock, Vertex AI, Azure OpenAI, and OpenAI-compatible endpoints. Each offers the one model it is set up with, named after it, such as
bedrock/us.anthropic.claude-opus-4-7orazure/<deployment>, and a named endpoint goes by its name. The model is often set in the workflow file, so these appear only once a run starts.
Two kinds of credential cannot be asked:
- A provider reached through a gateway, including keyless auth.
- The OpenRouter key Pullfrog mints for Router, OSS, and trial runs. Those runs can still ask a model you hold your own key for.
Limits
- A run can ask at most 12 times.
- Each call stops after 9 minutes.
- Each call starts fresh. It cannot see the run's conversation, so the question has to include what the other model needs, such as the finding, the file and the line.
- A subscription at its usage limit is refused at once, with the time it resets. An API key at its provider's limit can take the full 9 minutes to fail.
- Your code goes to the provider of every model you ask, under that provider's terms.

