Review agent

OverviewCopy Anchor Link

Review agent is an Inspect-powered automated reviewer. It can apply built-in and organization-specific review skills, re-review new commits, and surface actionable findings before human review.

Each review includes a general code reviewer and separate reviewers for security, logic, API contracts, data integrity, and performance. A test-design reviewer can also run when enabled. Repository rules add specialist reviews to this panel. Use them for checks specific to your codebase, or to get a second review with another model.

ModelsCopy Anchor Link

Set the organization defaults in Settings → Review agent. You can select a model and a reasoning effort for each of these roles:

• Orchestrator: the main review session. It plans the review, dispatches the reviewers, and writes the review. System default: gpt-6.1-sol with high effort.
• Generalist reviewer: reviews the whole change. System default: gpt-6.1-sol with xhigh effort.
• Sub-reviewers: the security, logic, API contracts, data integrity, performance, and test-design reviewers, and repository review skills. System default: gpt-6.1-sol with high effort.

A role that you do not set uses its system default. If you select a different model and no effort, the role uses high effort. If your Inspect deployment cannot run gpt-6.1-sol for a role that uses the system default, the orchestrator uses gpt-5.6-terra with high effort, and reviewers without a selected model use the orchestrator's settings.

Organizations that selected one model and effort before these roles existed keep that choice for all three roles until an admin changes the settings.

Model and effort choices are saved when a review session starts. Restart Review agent to apply a new selection to an existing run.

Repository rulesCopy Anchor Link

Add .inspect/reviews/rules.yml to set repository defaults and specialist reviewers. Inspect reads it from the pull request's head commit.

version: 1
model: gpt-5.6-sol
reasoning_effort: high
rules:
  - id: security-sol
    skill: security
    model: gpt-5.6-sol
    reasoning_effort: xhigh
    files: ['src/**']
    condition: Authentication or authorization behavior changes
  - id: security-opus
    skill: security
    model: claude-opus-5
    files: ['src/**']
    condition: Authentication or authorization behavior changes

The root model and reasoning_effort fields are optional. Each field applies to all three roles and replaces the organization setting for that field. A rule's model and reasoning_effort override those settings for that specialist. A rule field that is not set uses the sub-reviewer setting. Supported reasoning efforts are low, medium, high, and xhigh. Restart an existing review to apply changes to its model or effort defaults.

Each rule needs a unique id and a skill. The optional files list limits the rule to pull requests that change matching paths. The optional condition tells Review agent when to run the skill. Omit both to run the skill on every review.

The example runs the security skill twice, with separate models and distinct rule IDs. Rules that select the same skill, model, and reasoning effort are combined into one reviewer; rules with different model or effort overrides stay separate. These specialist reviews are additional to the built-in security review.

The skill must exist in the Inspect skill catalog or at .claude/skills/<skill>/SKILL.md in the repository. For this example, provide a security skill with your organization's checks. Skill names with a namespace, such as plugin:security, are currently skipped. Per-rule model and effort overrides are supported by hosted and Laminar reviewers. Select models available to your organization and effort levels supported by the reviewer runtime. An unavailable model or unsupported effort can reject the review launch or follow-up; Inspect does not silently replace the selection.