Custom Python action output not reaching the model in portal chat (works via API

Bugs / Site Issues
1

Summary
A custom Python action runs correctly and its full result appears in the action run log, but in the portal chat the model receives nothing, or only the first part of the result. The same agent, action and prompt give complete answers when run through the API.

Setup

  • Private Python action that calls an external search API and returns plain text (about 3,500 characters: six numbered results with title, link and excerpt).
  • Agent on claude-haiku-5-5. I also tested claude-sonnet-5-5 with streaming off.
  • Action attached with a trigger prompt, three actions attached in total (Gold plan).

What happens in the portal chat

  1. The run log shows all six results, but the model says it received 0 characters or “no results”.
  2. In another run the model reported only the first result and the start of the second, as if the output were cut at roughly 800 to 1,000 characters.
  3. A trivial test action with no network, returning six fixed lines instantly, ran correctly in the logs, but the model saw nothing in all three return styles (plain string, dictionary with a result key, delayed by 8 seconds). So it is not size, format or timing.
  4. Built-in Google Search results do reach the model in the same chat.
  5. Through the API (run in the background, then fetch the completion), the same agent receives the full results and uses all of them.

Also noticed

  • Every action result in the logs is followed by “No files found.”
  • With reasoning effort “high” on Haiku, answers needing several action calls often end with an empty reply. With “medium” the same request completes.

What I have ruled out
Model (Haiku and Sonnet), streaming on or off, output size, output format, delay, relevance threshold and memory buffer settings.

Questions

  1. Is there a size limit, filter or summarising step on action output in the portal chat that doesn’t apply to the API? Can it be changed?
  2. What does “No files found.” mean, and does it affect what the model receives?
  3. Is there a runtime setting for the agent (for example a basic or advanced runtime) that controls whether custom action output is passed to the model? I can’t see it in Studio.
  4. Any tips to get a custom Python action’s full output into the portal chat?

Thank you. Any hint helps.