Repository navigation
Completions Not Refreshing Ghost Text #2
Description
I'm running Eclipse 025-06 (4.36) for with the java developer package and the latest version of AI Coder (1.0.0.202507142020) using a recent version of Ollama and the qwen2.5-coder:3b from the example https://eclipse.hetzge.de/aicoder/.
The completion generation works both when "Enable autocomplete" is enabled or the manual trigger is used. I've verified this by watching the AI Coder History view and inspected the 'Generated' result. However, the Ghost text does not appear when I am actually working in the Java text editor. It only shows up when I click away from it (like into the package explorer or into the history, basically just removing focus from the editor).
Separately, (and maybe an ignorant question) the completions work with the completions qwen model, but any other model I try to use I get a '400' response from ollama. What criteria can I use to determine what models should be compatible?
Awesome plugin!
Activity
hetzge commentedon Oct 24, 2025
Hi @skcussm,
Ghost text refresh problem:
It seems there is a problem with refreshing the editor content when the completion is ready to show. Which operating system do you use? I expect Windows, because there I could monitor similar behavior. The ghost text in this case is also refreshed if you scroll the editor. It could be that this problem is already solved in the latest (not yet officially released) version of the AI Coder plugin. I will retest that on Windows soon.
If you want, you can try the latest build by using the following Eclipse update site: https://hetzge.github.io/aicoder/eclipse/beta1
(There is also a new inline edit feature that can be triggered with Ctrl + Shift + 1)
Other completion models:
You can detect models that are compatible with fill-in-the-middle code completion by looking at the model’s prompt template. For example, for your qwen-2.5 model you can find it here: https://ollama.com/library/qwen2.5-coder:3b/blobs/1e65450c3067
You will see the usage of prefix and suffix:
Thank you for the feedback
skcussm commentedon Oct 24, 2025
Awesome! Indeed, I am on Windows 11.
The issue is fixed in the beta. Thank you for the quick response.
The beta looks like a huge update! Lot's of new features. When do you plan on releasing? Is that feature set complete for the beta? Are there known issues?
I see you you've added the OpenAI provider. I also use LM Studio to run models locally which provides openai compatible v1/[models|responses|chat/completions, completions|embeddings] endpoints. Do you see any issues using all the features with LM Studio as openai provider?
I see there is nothing by default right now in the MCP configuration. I've worked with some other plugins like AssistAI and Code-Intelligence which provide some eclipse tools that can be called by the models. Are you planning to add something like this, or is there another standard eclipse tool set that can be exposed through that configuration?
I'm excited to see all the updates, and again, thank you for the response!
hetzge commentedon Oct 24, 2025
Great that the beta is working for you.
Release is planned soon, but it's a hobby project "soon" ;). I plan to implement and improve more of the context feature (add/remove files to the context, context prioritization, and so on...). There are still some hard-to-fix bugs, like flickering line-height issues and other refresh/rerender problems.
The OpenAI provider currently only works for the other actions than inline completion, as far as I know. I think I need to make some small changes (use the completions endpoint instead of chat completions) to get inline completion working with LM Studio, too. I will investigate that because I am currently also experimenting with LM Studio.
The MCP client is currently implemented only with the prompts feature (and only supports static prompts). The idea is to share prompts from a server with your colleagues. MCP tools are not planned at the moment because AI Coder has no agent or chat feature.
hetzge commentedon Oct 26, 2025
@skcussm I just pushed a small update to the beta update site. With this update, you can use fill-in-the-middle with LM Studio via the OpenAI API. I successfully tested it with the qwen2.5-coder model.
skcussm commentedon Oct 27, 2025
That's great! I'll check it out when I have a chance. It's great seeing this capability for eclipse being developed. I work on fully isolated networks, and so far this is the only option I've found for eclipse that supports offline models (without a big upfront hardware/license cost). It's looking really solid.