You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
i hate giving permissions. i feel like this precedent was set by incumbents and it's dumb. this is what git is for? idk, maybe this is heavy handed, but i know other tools are opting out of prompting for permission and that's what i would prefer, as a user.
we could just have a full-auto option, a la codex or claude code's shift+tab (kind of); could be enabled by default?
we could also do the codex thing and use macos seatbelt to sandbox open code to your project dir (+ config dirs/etc).
opening this issue to get thoughts from others, i'd love to just rip out permissions altogether but i also hope to not be the only user of open code, soooooo.
Full auto option makes sense, but imo probably shouldnt be enabled by default, pragmatism makes sense for first time users.
On sandboxing, is this about write operations only? If both read and write, Id be concerned about it being a hard limit of the scope as there are cases where Id want to reference a file from a different project (follow this style kind of thing). In case its about both read and write, would we be able to expand scope dynamically? I am sorry if I misunderstood the second suggestion!
Reacted by Adam, Youcef Boumar, Yukai Huang and Adam Smaka
i'd have to look at what approach codex takes w/r/t read vs write, but my first thought was that it would be for both read and write (prevent the tool from reading sensitive system files? idk). i don't feel strongly either way.
Sometimes I like to scroll through individual edits before approving them (mostly to stop Claude from being too eager to please) but I also sometimes like to go YOLO mode so would love the option to configure this or toggle this on/off per project.
Every so often Claude goes wild and wants to "start from scratch" just because it confused itself, so I would like to control the permissions easily.
I only use opencode now but I used Claude Code and Codex, and I enjoyed the way Claude Code does "full auto" quite a bit. I like being able to turn it on and off super easy depending on what I prompt it to do
Activity
indeep99 commentedon May 15, 2025
Full auto option makes sense, but imo probably shouldnt be enabled by default, pragmatism makes sense for first time users.
On sandboxing, is this about write operations only? If both read and write, Id be concerned about it being a hard limit of the scope as there are cases where Id want to reference a file from a different project (follow this style kind of thing). In case its about both read and write, would we be able to expand scope dynamically? I am sorry if I misunderstood the second suggestion!
adamdotdevin commentedon May 15, 2025
i'd have to look at what approach codex takes w/r/t read vs write, but my first thought was that it would be for both read and write (prevent the tool from reading sensitive system files? idk). i don't feel strongly either way.
ezynda3 commentedon May 16, 2025
Sometimes I like to scroll through individual edits before approving them (mostly to stop Claude from being too eager to please) but I also sometimes like to go YOLO mode so would love the option to configure this or toggle this on/off per project.
rekram1-node commentedon May 16, 2025
Every so often Claude goes wild and wants to "start from scratch" just because it confused itself, so I would like to control the permissions easily.
I only use opencode now but I used Claude Code and Codex, and I enjoyed the way Claude Code does "full auto" quite a bit. I like being able to turn it on and off super easy depending on what I prompt it to do
adamdotdevin commentedon May 16, 2025
yeah, i think the shift+tab toggle approach is good
thdxr commentedon Jun 14, 2025
going to close for now as 0.1.x is a rewrite. feel free to open a new issue if it's still a problem
feat: Fix Windows command execution and edit tool issues -- Stable an…
feat: Enhanced specialist domain restrictions and scope violation pre…
chore: update tiara submodule with security fixes
fix: smart retry classification, Bedrock 200K context cap, batch erro…
fix(pr): refresh before create, validate base branch, fence comment b…
fix: smart retry classification, Bedrock 200K context cap, batch erro…
11 remaining items
fix(prevent-sleep): code-review 修复 — 失败不谎报 + 存盘顺序 + 前端 wrapper/store …
fix(telemetry): code-review 第二批 — update_applied 阻塞上报 + 启动期 IO 后台化 + …
sec(app): validate deep-link parameters (directory / providerID / code)
fix(openrust): fix subagent multi-tool-call, parallel streaming, and …
docs: post-refactor review of the event-flow architecture
feat(car): surface failures, honour tool idempotency, add recall and …
feat(car): use targeted recall for context instead of dumping the fac…
docs(smoke): 人工单缩到 5 条;记录两处判错更正与 user 自验结果 [feat: ui-probe-toolkit]
feat(tui): native Arabic/RTL bidi support for prompts and messages
fix(tui): correct mixed-direction BiDi ordering, isolation and wrapping
fix(tui): correct mixed-direction BiDi ordering, isolation and wrapping
deps(core): align @ai-sdk family (provider 3.0.15, provider-utils 4.0…
fix(opencode): a review file decided both ways is partial, not accepted
feat(app): agent changes first, a visible keyboard model, decided dif…
test(app): a file kept and reverted in part is partial in the review e2e