Auto review is an experimental access mode for a chat. A model checks each tool call before it runs and blocks the risky ones, so Oda works with wide access while every call passes a review instead of an approval prompt.
Auto review is experimental. Its reviewer can wrongly allow or wrongly deny a call, and anything it allows runs without asking you.
Before a supported tool call runs, the reviewer checks it once and answers with one risk level and one decision, under a fixed policy:
An allowed call runs as usual. A denied call never runs: Oda sees the denial and carries on with narrower work.
Auto review is an advanced feature of Oda and ships switched off: until it is switched on, the mode pickers offer only the three standard modes, Read Only, Workspace Write and Full access (see Workspaces, files and permissions). To use it in a chat:
/. It
carries an EXP badge and applies to the current chat.Enable Auto review (experimental)? explains that the mode runs
without a sandbox, that a review can wrongly allow or wrongly deny a call, and that it spends
extra tokens. Tick the acknowledgement box to confirm; cancelling leaves the chat in its
previous mode.A denied call never runs, and the reason is kept. The denial reads
Auto review rejected tool "<name>"; its body was not executed, with the tool's own name in place;
in the chat, the step stays as a plain row reading Rejected by Auto review, and opening it spells
the reason out. The denial is recorded with the code AUTO_REVIEW_DENIED.
When the reviewer's answer is not exactly one decision, the call is denied too; the failure is
reported as auto-review: reviewer output must be one JSON object or
auto-review: reviewer output does not match the risk/. Either way, Oda carries
on with narrower work; Oda's steps and tool results covers the row.
If the Auto review feature is switched off while a chat is in it, the chat moves to Full access. Switching the feature back on returns the mode to the pickers, but it does not move that chat back to Auto review; the chat stays in Full access. While the feature is off, the mode pickers offer only the three standard modes.
Choose Auto review in the mode picker of the message box, or run /, then
accept the risks in the confirmation. The feature must be switched on first, and the mode applies
to that chat only; switching it on is above.
It ships switched off: until its feature is switched on under Settings > Developer > Advanced features, the pickers offer only Read Only, Workspace Write and Full access. The switch also needs a DataMind account with Full access; see Advanced features.
No. In this mode the review decides: each call is allowed or denied, and a denied call simply does not run. The file limits of the other modes do not apply, so the review is the check on every call.
Nothing of the call runs. The step stays in the chat reading Rejected by Auto review, its reason
is kept with it, and Oda continues with narrower work. The denial carries the code
AUTO_REVIEW_DENIED.
No. Each call is reviewed once and a denial stands: the mode has no override and no manual fallback. If you need the action done, switch the chat to another mode with the picker.
Yes. It is experimental: the reviewer is a model, so it can block a call that was fine or allow one that should not run. Anything it allows runs without asking you, which is why switching the mode on asks you to accept its risks.
Yes. Every reviewed call adds one model request to the chat, and reviews are not cached or retried, so the same task costs more than in another mode. A review that fails, or cannot be sent in full, denies the call.
Every supported tool call is checked once before it runs. Work inside a block of code Oda runs is not reviewed call by call, and there are no exemptions or standing approvals for later calls.
No. The mode applies to the chat you switch; the Permission setting that decides what new chats start in never offers it.
The chat moves to Full access. Switching the feature back on returns the mode to the pickers, but the chat stays in Full access.