Default permissions allow editing files and executing any commands

[NaikSoftware]

In my opinion very strange that Opencode allows any actions by default. I did not know any other AI tool with the same behaviour. It would be cool to have default permissions set – always ask for editing, and also always ask for critical commands, etc.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Feature Request: Add Prefix Requirements for Permission Prompts #2521: Comprehensive feature request for adding prefix requirements for permission prompts to prevent accidental approvals of sensitive operations
• Feature Request: Yolo and non-Yolo mode (auto-accept) #1813: Feature request for easily switching between 'ask' permissions and 'allow all' (yolo) mode, similar to Claude Code's permission modes
• [FEATURE] add exec tool as more secure alternative to bash tool #2206: Proposal for more secure alternatives to bash tool and general security improvements for command execution
• [SECURITY] allow ignoring files to prevent secrets being leaked from .env to LLM #539: Security-focused issue about preventing secrets from being leaked to LLM through file access controls

Feel free to ignore if none of these address your specific case.

[rekram1-node]

This was done very intentionally by the team they felt like this was a better UX and better default when they made the decision

[NaikSoftware]

This was done very intentionally by the team they felt like this was a better UX and better default when they made the decision

Maybe "BUILD WITHOUT ASK" predefined agent could be an option. I spent much time to research how to setup permissions and define tools I use. I think almost all users managed permissions and I am not alone)

[thdxr]

most people initially think they want permissions and then actually prefer not having them so that's why our defaults work this way

majority of people never turn them on so we won't flip the defaults. let us know if the docs around permissions or the feature itself is lacking in a new issue

[NaikSoftware]

I can't imagine how can work in a real project without permissions. There is one option - Opencode is used only for playing around and forgetting)

[rekram1-node]

I use it everyday in real projects at work, so do my coworkers, none of us use permissions

I think it is just a preference thing, if the llm makes a mistake I just do /undo and start over or try a fresh session

I use git worktrees so sometimes I just have a branch I am using it on and I do more focused work on separate ones

[NaikSoftware]

@rekram1-node you can't undo creating MR with wrong target branch/description or undoing any bash command also impossible. In Claude Code / Windsurf / Warp I can refine result. It much faster and safe

[rekram1-node]

True, i just don't have my agents do anything destructive I generally dont prompt about git commands other than git diff and git log and I've never had it go off the rails for me. Granted I know that is a possibility but I have protecctions in place to ensure that force pushes to main can't happen, etc

[NaikSoftware]

@rekram1-node as an option please provide possibility to generate opencode.json with normal permission handling in /init command. Ask what user needs (and add statistics tracking optionally). Allow anything by default is an stopper for several users because it's dangerous. I recommended Opencode for friends and they are surprised by this behaviour.

[anuramat]

why not wrap it in a sandbox/container?

[NaikSoftware]

why not wrap it in a sandbox/container?

wrap remote services?)) I want to allow access to them, but with explicit consent. In addition in many cases you should disallow edit and prompt to do it in right way

[anuramat]

I want to allow access to them, but with explicit consent

I mean you can, but why force everybody else to add boilerplate to their configs? if you are going to give your AI access to a prod db, you're probably motivated enough to rtfm and figure out the permissions

in many cases you should disallow edit and prompt to do it in right way

in my experience it's better/faster/easier to either let the agent iterate or undo and re-prompt (or just do it yourself)