The permission validation that prevents commands from accessing paths outside the project directory was accidentally disabled, which could allow commands like 'cd ../' to escape the workspace. This restores the security check that keeps your commands safely contained within your project boundaries.
The patch tool now works seamlessly alongside other file editing tools with improved
error handling and a more intuitive permission system. Users will experience:
- More reliable patch application with better error messages
- Consistent permission prompts that match other editing tools
- Smoother integration when applying complex multi-file changes
- Better feedback on what changes are being made before applying patches
This refactoring leverages the robust patch parsing engine while making the tool
feel native to the opencode workflow, reducing friction when making bulk changes
to your codebase.
This release has a bunch of minor breaking changes if you are using opencode plugins or sdk
1. storage events have been removed (we might bring this back but had some issues)
2. concept of `app` is gone - there is a new concept called `project` and endpoints to list projects and get the current project
3. plugin receives `directory` which is cwd and `worktree` which is where the root of the project is if it's a git repo
4. the session.chat function has been renamed to session.prompt in sdk. it no longer requires model to be passed in (model is now an object)
5. every endpoint takes an optional `directory` parameter to operate as though opencode is running in that directory
The concept of mode has been deprecated, there is now only the agent field in the config.
An agent can be cycled through as your primary agent with <tab> or you can spawn a subagent by @ mentioning it. if you include a description of when to use it, the primary agent will try to automatically use it
Full docs here: https://opencode.ai/docs/agents/
Dax Raad