OpenCode v1.18.17 Changelog

This release is a maintenance update that focuses on making the AI chat experience more reliable. It matters most to developers who use specific AI models (like GitHub Copilot, DeepSeek, or Meta's Muse) because it fixes several behind-the-scenes glitches to make those tools run smoothly.

Fixes

• Smarter chat summaries: When your chat gets too long, the app "compacts" it to save space. Previously, it might have accidentally deleted your most recent messages while summarizing. Now, it keeps your newest messages safe and writes clearer summaries for smaller AI models.
• Fixed MERGE Gateway models: Added missing "reasoning variants" (special steps an AI takes to think through a problem) so these specific model options work the way they are supposed to.
• Stopped "retry storms": If a connection fails, the app tries again. Before, it could get stuck trying over and over too fast, causing a traffic jam. It now puts a limit on retries and adds "jitter" (random timing delays) to prevent this.
• Enabled PDF attachments for Copilot: If you use GitHub Copilot models that are designed to understand images and documents, you can now successfully attach PDFs for the AI to read.
• Fixed DeepSeek V4 Flash settings: Applied the correct "sampling defaults" (settings that control how creative or predictable the AI's text is) for this specific model.
• Corrected Meta Muse prompts: Routed all models in the "Muse family" to use the correct Meta system prompt (the hidden instructions that tell the AI how to behave).
• Fixed the default model picker: The desktop app was sometimes using an old, saved setting when picking your default AI. It now checks the server for the current default model instead.
• Updated Chinese translations: Improved the Chinese words used for developer terms so they make more sense to the people using them.

Breaking changes

No breaking changes in this release.