Customize
Configure automatic approval
Choose the model that reviews actions when permission rules need a decision.
In auto mode, Vera checks tool actions against permission rules. When those rules cannot decide, a separate classifier evaluates whether to allow the action. If it cannot return a usable decision, the tool does not run.
An action the rules allow runs without a notice. So does a low-risk allow
from the classifier; the end of the turn shows how many there were, such as
2 auto-approved. Other classifier decisions add a line to the conversation.
An allow line sits just above the action it allowed.
When the classifier allows a tool from an extension, the rest of the
conversation runs that extension operation without asking it again, the same
as answering Session on a prompt. Vera’s own operations, such as git.push,
are classified on every call. Review granted permissions in Ctrl+P lists
and removes these grants.
The classifier is separate from the model working on your task. Changing the conversation model does not change the configured classifier.
A managed service is on the roadmap. It will be opt-in and will run Vera fully automatically at a choice of usage tiers. Vera will pick the models, including the classifier, to meet the tier’s expected performance, for users who would rather not tune models themselves.
Choose a classifier
- Open
/modelsand choose Defaults. - Under Dedicated jobs, choose classifier.
- Select its model and reasoning effort.
The assignment applies to the next classification, including in existing conversations. No host restart is needed.
A weak classifier can approve an action it should have blocked. Use a model at the level of Claude Haiku 4.5 or better. To find one, set the Intelligence cutoff in the model picker’s Filter and sort (see Models) so that only well-ranked models are listed.
Set a failsafe
The Classifier entry in /settings sets the same classifier as Defaults,
plus an optional failsafe model. Changing the classifier in Defaults keeps the
failsafe. Use the session model clears the classifier and its failsafe,
so the session model reviews actions instead.
The stored key is model_assignments.reviewer: the first model is the
classifier and the second is the failsafe. The interface calls the job
classifier. The reviewer block in config.json holds only review settings
such as timeout_ms and two_tier. A reviewer block that names a model
fails to load, with a message that says to remove it.
Understand the result
An allow notice for a medium-risk or higher action includes the assessed
risk, authorization, and reason. Low-risk allows are counted instead. While
the turn runs, the status line shows the count. Click the count at the end of
the turn to list each auto-approved action with the classifier’s reason under
it, and click it again to close the list. Long commands are cut short there;
the action’s own tool row has the full command. An action shows no reason when
the classifier gave none. Ctrl+T opens or closes every list along with
the tool details. The count and its list are not saved, so a reopened
conversation does not show them.
A denial says the action was denied. A timeout or provider error names the
tool, confirms that the action did not run, names the classifier model, and
points to /settings then Classifier to pick another.
The classifier receives user turns and tool calls. It does not receive tool results.
Recover from an unavailable classifier
A failsafe handles a classifier that errors or cannot be reached. Set it in
/settings under Classifier. If the provider itself is the problem, such as
an exhausted credit balance, pick a classifier on another provider there or in
Defaults.
Inspect review logs
Each classifier request is recorded in
~/.vera/runtime/logs/reviewer.jsonl. The log includes the prompt, response, grades, and latency.
This is a private diagnostic file containing conversation material sent to the classifier provider. Treat it as sensitive when sharing a report.