Skip to main content
Experimental version.This is the latest in-development version of AI-Implement. Features may change without notice and behavior is not guaranteed. Switch to the latest stable version here.
AI-Implement runs a list of reviewers after it opens a pull request. Each reviewer reads the diff and reports what it finds. A reviewer either gates the merge or runs as advisory. A gating reviewer blocks the merge when it finds a problem. An advisory reviewer reports its findings but does not block the merge.

​
Use Codex for internal review

These controls make different choices: Selecting Codex for the review stage does not install an external Codex review bot. The configured stages divide the work as follows: For example, Claude implementation with Codex review sends code fixes back to Claude. A changed selection applies to a later run; the current run keeps its resolved settings. The configured review model takes precedence over repository model values and each reviewer definition’s model.

​
Built-in reviewers

AI-Implement includes these built-in reviewers:
Turn code-review off only when a trusted external reviewer is in place. See External code reviewers.

​
Which reviewers run

The project’s Reviewers setting selects which reviewers run and which of them gate. A pull request cannot change this setting. When the setting is blank, both built-in reviewers run and gate. The project settings page describes this control.

​
Define a reviewer

You can add your own reviewer in .ai-implement/config.yml in your target repo. Each entry has these fields:
  • id — required
  • prompt — required
  • model — optional legacy model override; configured review-stage selection takes precedence
These rules decide which definition AI-Implement uses and whether it gates: A pull request cannot rewrite the reviewer that judges it. To add a gating reviewer:
1

Merge the definition to the default branch

Merge the reviewer’s definition in .ai-implement/config.yml to the default branch.
2

Select the reviewer

Select the reviewer in the project settings.
3

Turn on Gates

Turn on Gates for the reviewer.
The ids claude-review-summary and legacy-post-push-review are reserved. AI-Implement ignores a config entry whose id matches a built-in reviewer or an image-baked reviewer.

​
Reviewers with code

A fork of AI-Implement can add a reviewer that runs code. The fork bakes the reviewer module into the runner image at custom/reviewers/<id>.ts. A reviewer with code overrides a built-in reviewer that has the same id. A target repository cannot supply reviewer code. See Runner image and Customize pipeline behavior.

​
Turn limit

These turn limits apply to Claude reviewers. AI-Implement takes the limit from the first of these that is set:
  • the reviewer’s Max turns in the project settings
  • the built-in default
  • Review Max Turns in Settings → Retry Policy (default 30)
When a reviewer reaches its turn limit, the review is incomplete and reports REVIEWER_TURNS_EXHAUSTED. See A reviewer reached its turn limit. Codex reviewers use the configured review stage’s elapsed timeout. Max turns and Review Max Turns do not cap Codex. See Timeouts and limits.

​
What the reviewer returns

Each reviewer posts a report on the pull request. The report starts with the reviewer’s label and one result:
  • approved — the reviewer found no problem
  • a finding count — the number of problems the reviewer found
  • incomplete; no approval — the reviewer did not reach a verdict
After the result, the report shows a Checks: list. Each check has one of these results:
  • Passed
  • Failed
  • Not verified
  • Not applicable
A reviewer that does not finish shows {label}: review incomplete (required) or {label}: review incomplete (advisory).

​
Source inspection and executed checks

Codex review has restricted read-only access to repository text. It can read and search, but cannot edit files or run tests and shell commands. Read each check’s evidence. A source inspection can support a check without executing a test. Not verified means the check was not established; it is not a test pass. Automated verdicts do not provide a required human GitHub approval. See Approval is a comment.