Open WebUI 0.12.4 makes skill loading selective and clarifies context compaction

Open WebUI 0.12.4 adds a way to load selected skills only when a model needs them, and makes context compaction easier to follow in chat. The release also extends compaction to temporary and short chats that might otherwise hit a context-length error.

Open WebUI Watch analysis

What happened

The project’s 0.12.4 release notes say the version was released on 25 October. With the new skill-loading option enabled, selected skills are listed for the model to load when needed, rather than having their full instructions sent with every message. Skills mentioned directly in a message still load immediately. The option is off by default and requires the built-in Skills tool.

Compaction now leaves a “Context compacted” marker in the chat. Users can open it to see the summary; the marker also appears in a reply when compaction happened between tool calls. Open WebUI says compaction now runs in temporary chats and in short chats that have already crossed the context threshold, helping avoid context-length errors on follow-ups, regenerations or edits.

The release also adds administrator controls to hide individual interface options from users, and hides options that cannot work under the user’s current permissions or configuration. Hidden controls retain their saved values, and features behind them remain usable where applicable.

Why it matters

Loading skill instructions only when needed can reduce repeated prompt overhead for people who use selected skills, while leaving direct mentions immediate. The compaction changes make it clearer when older messages have been summarised and address cases where even a short exchange can run out of context after a large tool output.

For administrators, the interface changes offer a tidier way to limit what users see without treating every unavailable control as a mysterious broken button. The trade-off is that skill loading is opt-in and depends on the built-in Skills tool.

Our read

This is a practical release for people running Open WebUI: less unnecessary skill text in each message, clearer signs of what has been summarised, and fewer dead-end controls. If you use selected skills, check whether the built-in Skills tool is enabled and decide whether to turn on on-demand loading. Context has a habit of filling up at the least convenient moment; at least now the chat can show its working.

What to watch

• Whether on-demand skill loading reduces prompt overhead in everyday use.
• How often the new compaction behaviour prevents context errors in temporary and short chats.
• Whether administrators find the new visibility controls easier to manage across different permissions.

Discussion spark: Should AI workspaces load selected skills only when needed by default, or is the predictable behaviour of sending their instructions with every message worth the extra context?

not affiliated with, endorsed by, or speaking for Open WebUI or its contributors