Hi,
Thanks for laying out the questions. Here is how this works in Studio 3.1.4.
In the ABP Agent panel, open the model menu (the Select Model button) with Auto turned off.
Max off: the conversation stays within a 300K context limit. Reasoning Effort in the same menu is yours to choose. Max on: Studio uses the full context window of the selected model and sets reasoning to the highest level. An infinity icon next to the input means Max is on. If the selected model has a smaller context window than 300K, that model’s window is what applies. Changing models inside the same chat does not clear what the agent has already read. The limit covers the agent’s working context: your messages, the files and tool results it is still using, and the current plan.
If you want to learn "How to do {x} in ABP Studio", you can ask it to APB Agent, it knows almost everything ABP Studio can do and how to use the interface.
For day-to-day work, leave Max off and scope the agent to the module or folder you are changing. Studio summarizes older parts of a long chat on its own. Ideally you should not have seen that warning, we will work on it. When you see “The conversation is too large…”, start a new chat and do not carry the old conversation into it for now. Besides, it may be cost efficent, depends on the session.
Check the files on disk and the git diff before you ask the agent to retry. A tool step already shown as completed in the chat is on disk. Treat the step that was running when the timeout or the connection error appeared as unfinished. You do not need to rerun a migration or a generator to check.
The plan file remains in the solution under .abpstudio/plans/. A new chat does not open that plan by itself, so it can report that there is no active plan while the file is still there. Attach that file in the new chat, or give the agent its path, and ask it to continue the steps that are not finished. The checkboxes in the file show what is already done.
@harryhopkins
We released a new version after the latest fixes about this ticket. And soon we will release another version with major changes and it will include significant token/cost optimization of ABP Agent, aimed for next week.
@harryhopkins
Yes, it is related with prompt caching again. But this time the cache break is from another source. We're hardly working on cover all possible cache breaks for a couple of days and we beleive that we managed it. We are planning a new release on monday.
Let me know if you need more information.
Since we can not see content of your requests, we tracked input/output token counts from your account and we detected that it was related with context compressions on long running sessions. We will inform you.
Hi again @harryhopkins
I am investigating the case. But it is not related with the previous issue, as i see.
Hi @harryhopkins
By checking your logs, we've confirmed that prompt cache discount is not applied as it should be, due to a bug in ABP Studio. We will refund your AI credits (partial or full) and release a new version asap.
Hi again,
We decided to bring the ABP Studio MCP Server back in future releases, fyi.
Hi again,
We decided to bring the ABP Studio MCP Server back in future releases, fyi.
Hi @AlessandroLittera
We've removed the ABP Studio MCP server support after the launch of ABP Agent.
We also removed the document but somehow it is still present. We'll check it. https://github.com/abpframework/abp/pull/25518/
Yes, there are rules & skills for ABP: https://github.com/abpframework/abp/tree/dev/.agents/skills https://github.com/abpframework/abp/tree/dev/ai-rules
Hi @PDarioJer
We've removed the ABP Studio MCP server support after the launch of ABP Agent.
We also removed the document but somehow it is still present. We'll check it. https://github.com/abpframework/abp/pull/25518/
