Continue v1.2.19 Changelog

This update is all about making the app more stable and customizable. If you use Continue to code with different AI models—especially "thinking" models that show their reasoning, or local models run through Ollama—this release fixes a lot of annoying glitches so the AI behaves the way you expect it to.

What's new

• Opt-out option for OpenAI's Responses API: OpenAI has a new way for apps to talk to its models (the Responses API). This update adds a setting (useResponsesApi) so you can choose to turn it off if it doesn't work for your setup.
• Better identification with OpenRouter: When Continue talks to OpenRouter (a service that connects you to many different AI models), it now sends extra background info (HTTP-Referer and X-Title headers) to identify itself, which helps keep the connection running smoothly.

Improvements

• Faster autocomplete: The code-completion feature used to quietly leak memory and read the same files twice. It now cleans up after itself better, making it faster and more efficient.
• Less background chatter with Ollama: The app stopped constantly asking Ollama for model details in the background. It now only asks when it actually needs to, saving computer resources.
• Better setup for developers: Under the hood, the app now properly syncs its lock file (a file that keeps track of exact software versions) and no longer skips necessary installs by default.
• Clearer documentation: Added troubleshooting tips for Ollama memory errors and clarified exactly where you can safely put secret passwords (API keys) in your settings.

Fixes

• Smarter "thinking" models: Fixed an issue where AI models that show their step-by-step reasoning would accidentally mix their thoughts into your actual answer. It also stopped the "thinking" animation from staying on screen if the AI had no thoughts to show.
• Tool use fixes: When an AI uses tools (like searching the web or reading a file), it sometimes broke if the AI was thinking but didn't actually write any text. This is now fixed. Also, cleaned up weird backtick formatting from tool instructions.
• Gemini message ordering: Fixed a bug where Google's Gemini models would get confused if you sent multiple messages in a row from the same role (like if the system sent two messages back-to-back). It now merges them properly.
• Respecting your settings: If you manually set a "context length" (how much text the AI can remember at once) in your settings file, the app now actually listens to that limit for local models like llama.cpp and vLLM.
• Duplicate context providers: You can now add multiple of the same type of context provider (like two different GitHub repos) in your settings file without breaking the app.
• API key corrections: If you typed the wrong API key for xAI or Gemini, you used to get stuck. Now, you can easily fix it and enter a valid key.
• Ollama tool errors: If Ollama messes up while trying to use a tool, you now get a helpful, actionable error message instead of a confusing crash.
• OpenRouter and Gemini 3 fixes: Fixed several issues with OpenRouter, including better support for Gemini 3 tools, stopping it from cutting off text, and fixing the autocomplete endpoint.
• Kimi model support: Added support for the "reasoning_content" field so that Moonshot's Kimi models can properly show their thinking process.
• Removed Llama 3.1 405B from Groq: The Groq provider (a fast AI hosting service) no longer shows the massive Llama 3.1 405B model, likely because it's not supported there.
• CLI free trial fix: The command-line tool stopped constantly checking for a free trial if you were using a model that didn't have one.
• JetBrains release fix: Fixed a behind-the-scenes step in building the plugin for JetBrains IDEs so it handles multiple zip files correctly.

Breaking changes

No breaking changes in this release.