Choose the review level
Choose a faster or more thorough review, and understand your team’s rules.
Choose a level. Your AI tool chooses the model and how much reasoning to use. Grace sends your request and your team’s rules without requiring a particular model name.
This setting belongs to your account. Your profile and the grace-effort skill share the same choice for future reviews. It does not change the settings of the agent writing your code.
Choose your level
In Profile > Review level, choose a value, then select Save level. Wait for Review level saved for your next reviews. The information button lets you compare the choices with either a keyboard or touch.
| Choice | Example in a slash client | What you ask for |
|---|---|---|
| Auto | /grace-effort auto | Let your AI tool choose the model and reasoning based on the work to review. |
| Low | /grace-effort low | Favor speed with a lightweight model and little reasoning. |
| Medium | /grace-effort medium | Balance speed and analysis capability. |
| High | /grace-effort high | Use a capable model and more reasoning. The review may take longer. |
Every level keeps the same required checks. Choosing Low does not remove any rules; choosing High does not guarantee that every problem will be found.
Auto adapts to the work being reviewed. The AI organizing the review considers complexity, risks and uncertainty. It must explain its choice before starting the reviewer. Auto does not always mean Medium and does not promise to finish before a deadline.
To use the command, connect your agent with your Grace account through OAuth.
A project API key cannot change your personal preference.
The installation kit includes the grace-effort skill in French and English. Use your client’s
native syntax: $grace-effort in Codex, /grace-effort where slash skills are supported, or the
skills selector. In T3, use the selector or syntax of the active provider. The existing MCP prompt
also remains available in clients that expose it. Refresh the client’s skills menu after an update;
the presence of files alone does not prove discovery.
Run grace-effort without an argument, using your client’s syntax, to read your choice and the rules for this project. Reading the settings changes nothing.
You have succeeded when your profile and the command show the same saved value. It applies to future reviews; a reviewer already running keeps its settings.
Understand Grace’s message
When the task is prepared, your agent states the requested level and what determines it. For example, if you choose Low and the project requires at least Medium:
Grace — Medium review requested. Your Low preference is raised to meet the minimum review level expected by this project.
After choosing the settings and before every launch, including a retry or second opinion, the agent announces the requested level, the exact selected model, its reasoning effort and the talents and planned checks. Effort keeps the exact label from your AI tool, called native effort. Neither the model nor this effort is inferred from Low, Medium or High.
The following examples are message templates: replace the bracketed placeholders with actual choices.
Grace — Review: requested level Low; selected level Medium, raised by the project minimum; selected model
[exact model identifier]; reasoning[native effort value]. Talents and checks:[served talents and planned checks for this scope].
With Auto, the agent also announces the chosen level and a short reason:
Grace — Review: requested level Auto; chosen level High because the change affects access permissions; selected model
[exact model identifier]; reasoning[native effort value]. Talents and checks:[served talents and planned checks for this scope].
These messages illustrate the expected information; they are not completed reviews. If the organization and project require the same minimum, both are named. One announcement groups reviewers using the same settings; each different combination identifies its scope. Every new attempt is announced, even with the same settings. An unknown value is stated as unavailable, with the limitation encountered. If the agent launches a reviewer without passing the chosen effort, that reviewer uses the main agent's effort. The announcement then states reasoning inherited and level not applied instead of showing the requested level.
Requesting a level does not prove it was used. The agent distinguishes planned settings, configuration observed in your tool and confirmed execution. If the tool returns a different configuration, it corrects the announcement and explains what remains unconfirmed. Some tools offer fewer controls than others. Connecting to Grace does not guarantee these controls are available.
Understand your team’s minimums
A minimum is the lowest level your team accepts for a review. The higher of the organization and project minimums applies. Your account keeps your personal preference.
| Your choice | Team rule | Requested level |
|---|---|---|
| Low | No minimum | Low |
| Low | Organization: Medium | Medium |
| Medium | Project: High | High |
| Auto | Project: Medium | The AI tool chooses Medium or High. |
An owner or administrator can change these rules:
- For the entire organization: Organization > Configuration.
- For one project: Project > Settings > Reviews.
Members can read them. None removes only the minimum for that scope: a project cannot cancel its organization’s minimum.
After changing a value, select Save minimum. The displayed requested level reflects the saved settings; it does not change until you save your new choice.
Grace cannot yet confirm a mandatory minimum
When a minimum is required, Grace keeps the review findings but cannot declare the review compliant: trusted confirmation of the model and reasoning used is not yet available. Showing the right settings is not enough. Consider this limitation before requiring a minimum.
Your saved values are retained
Existing preferences and team minimums keep their values. Every setting now applies to both the model and its reasoning; Auto always adapts to the task. There is no adoption action to perform. Previous review reports keep their original results and decisions.
If settings are out of date
The setting refreshes every 30 seconds while displayed, and when you return to the window. If an error occurs, use Reload settings.
If someone changes the value while you edit, Grace keeps your choice on screen and shows the saved value. Choose Keep saved value or Apply my choice. The command also reports the conflict without silently overwriting the change.
To find your connection or available commands, see Connect a project and MCP tools.