ReviewMe
Review Agent Zero’s tasks against your requirements, with configurable simple or full-agent reviews.
README
ReviewMe
An optional review suite for Agent Zero. Before the main agent finishes a task or marks a goal complete or blocked, ReviewMe compares the proposed outcome with your original request, later instructions and available work evidence. A concrete missing requirement can send the agent back to finish the work.
Why ReviewMe exists
ReviewMe is an official plugin created by the Agent Zero development team following our experience with several coding benchmarks. Those evaluations highlighted Agent Zero as one of the most token-light harnesses we tested, but also showed that it can sometimes finish with insufficient verification or unmet requirements. ReviewMe adds an optional, configurable review before completion, so users can balance that efficiency with a closer check of the result.
Settings
| Setting | Default | Behavior |
|---|---|---|
| Review instructions | Built-in review guidance | Customize what the reviewer checks in either mode. Leave blank to use the default instructions. |
| Review mode | Single model | Single model checks only the conversation history for unmet requirements or requests. Full agent also inspects the actual files and other outputs. |
| Reviewer model | Main | Select Main or Utility from the inherited or selected preset. |
| Reviewer preset | Current chat models | Select a saved global preset without changing the solving agent's models. |
| Reviewer agent profile | Current profile | Applies only to full-agent reviews; uses that profile's normal prompts and tool permissions. |
| Maximum reviews | 1 | Between 1 and 3 per new user instruction. A second review can check the repair. |
If a saved preset or profile is removed or renamed, select its replacement in ReviewMe settings. An unavailable selection skips the review and records a configuration error; it does not silently choose another paid model.
One review is the economical default; full-agent mode can involve many model/tool calls. Profile review-only instructions are guidance, not a sandbox.
Completion policy
The reviewer must identify a concrete unmet requirement, quote actual user content and propose an authorized next action before requesting continuation. Later explicit scope changes take precedence. Valid blockers, missing input, stop requests and equivalent solutions can be allowed. The review should not invent deliverables or demand unrelated tests.
Requirements entered through the goal editor are retained separately from model-authored goal rewrites. Later editor changes and removals are observed before model turns and at completion/goal-tool boundaries. Paused goals cannot be marked complete or blocked by the review path. Enable the plugin before starting work; it cannot recover original instructions already lost before activation.
License
MIT; see LICENSE.