README.md
Rama main · réplica de lectura del repositorio original.
# ReviewRouter
AI code review for GitHub pull requests, powered by [OpenCodeReview](https://github.com/alibaba/open-code-review).
Run reviews in GitHub Actions with your own AI providers; manage findings, models
and review rules in one dashboard.
[Set up your reviewer now](https://reviewrouter.org) with a guided setup.
[Self-host](./docs/operations/review-router-self-hosted-end-to-end.md) · [Documentation](./docs/README.md)
## Dashboard preview
<img width="1136" height="681" alt="ReviewRouter dashboard" src="https://github.com/user-attachments/assets/18b8e7bd-88a3-4235-bfb2-2110ae284ce7" />
<img width="721" height="630" alt="ReviewRouter review comment" src="https://github.com/user-attachments/assets/3def12b4-ad47-4b85-95b7-5cc585e4ff72" />
<img width="1145" height="737" alt="ReviewRouter Accounts" src="https://github.com/user-attachments/assets/5b7b1383-683b-4266-9da5-74c48f1d94c9" />
<img width="1135" height="737" alt="ReviewRouter usage" src="https://github.com/user-attachments/assets/447a4f0c-5527-41bd-a08e-6bb540eb3943" />
<img width="1135" height="737" alt="ReviewRouter statistics" src="https://github.com/user-attachments/assets/9a6d86ae-1884-45f4-9c89-3c5851bd2ccf" />
## Features
### Memory, Skills and MCP
- **Memory (Hindsight).** Reuse approved workspace-, repository- and user-scoped
rules, observations and exceptions when a memory backend is configured.
- **PR conversations.** On qualified interaction workflows, respond to inline
review threads using the verified discussion. Authorized maintainer replies can
also propose memory changes.
- **Memory controls.** Browse and add knowledge, retire outdated entries or
request permanent forgetting. Workspace admins can require confirmation
before AI memory changes.
- **Skills.** Create versioned Markdown review instructions, import/export
`SKILL.md`, share across workspaces with permission and assign them to repositories.
Each review captures its selected instructions.
- **Skills catalog.** Preview [Universal Agent Plugins](https://universal-agent-plugins.com/)
skills, check static security results (optionally AI-assisted) and import
reviewed text without installing third-party scripts.
- **MCP tools.** Connect remote HTTPS MCP servers, discover and approve tools,
and select them for compatible reviews. Test Dashboard credentials in the UI;
verify GitHub Secrets in the repository workflow. Built-in tools cover
GitHub issues and docs.
### AI reviews
- **Inline code review.** Severity-rated, line-specific findings and summaries
attached to the exact PR commit.
- **Multiple reviewers.** Choose models and accounts, run reviewers in parallel
and require agreement before publishing inline findings.
- **AI providers and accounts.** Connect Codex, Claude Code, OpenRouter,
MiMo Token Plan or Kimi Code via supported subscriptions, OAuth or API keys.
Manage accounts and choose models from the available provider catalogs.
- **Review settings.** Workspace defaults and per-repository overrides for models,
reasoning effort, language, severity, comment limits and parallelism. Workspace
admins also control automatic reviews of draft PRs.
- **PR descriptions.** Optionally generate or update descriptions with the
configured primary reviewer.
- **Safe publication.** Stale runs cannot publish findings for newer code.
Failed or incomplete review coverage is reported, not counted as success.
### Dashboard and insights
- **Review history.** Search pull requests by title, number or repository;
filter by status and date, inspect previous runs and reruns, and open the GitHub PR.
- **Statistics.** Published reviews, distinct PRs and first-recorded issues by
severity, with trends, top repositories and 7/30/90-day comparisons. Drill into
individual findings and open their PRs.
- **Usage.** Provider attempts, known tokens and costs by provider, model
and run, with filters and explicit coverage when data is missing.
- **Operations.** Repository health, queue failures, audit events,
metadata-only diagnostics and workspace/repository access management.
### Setup and hosting
- **GitHub App setup.** Select repositories, create an inspectable setup PR
for workflow changes and verify readiness in the dashboard.
- **Organization-wide checks.** In eligible GitHub organizations, configure a
shared required workflow through GitHub Rulesets for selected or all repos.
Requires a supported GitHub plan and organization permissions.
- **Code and credentials.** By default, checkout and tools run in GitHub Actions,
repository credentials stay in CI, and the chosen AI provider receives review
context. ReviewRouter stores results and operational metadata, not raw source files.
- **Hosting options.** Use the hosted service or
[self-host with the guided Docker Compose installer](./docs/operations/review-router-self-hosted-end-to-end.md).
Availability varies by provider, installed workflows and deployment. See
[capabilities and limits](./docs/capabilities.md).
## PR comment commands
With the **ReviewRouter Interaction** workflow installed, `/rr help`, `/rr status`
and `/rr review` work in the PR conversation; replies to a ReviewRouter
inline finding can use `/rr skip`, `/rr unskip`, `/rr explain`, `/rr fixed`
and `/rr suggest fix`.
- `/rr fixed` checks **one** existing finding against the current HEAD and
reports whether it is fixed, still present or cannot be verified.
- `/rr suggest fix` may produce **one human-applied GitHub code suggestion**
for a verified current changed line; it does not commit, push or resolve threads.
These two focused commands require a qualified managed Conversation Gateway,
an authorized repository maintainer, a same-repository PR and a direct reply
to the original verified inline finding. Publishing focused replies also
requires `REVIEW_ROUTER_DISCUSSION_MODE=suggest` (or `discussion_mode: suggest` in
a reusable workflow); the default `off` does not publish the response.
Provider and workflow availability vary; some standalone Action installs only
listen to inline review replies.
See the [complete command reference](https://github.com/777genius/review-router#pr-comment-commands)
for all supported commands, permissions, Memory options and limitations.
## Start with GitHub
1. [Install the GitHub App](https://reviewrouter.org) and select repositories.
2. Create and merge the review setup PR from the dashboard.
3. Connect a supported AI provider and open a pull request.
You approve the workflow changes before reviews begin. Provider usage counts
against your own subscription or API billing.
## Self-host
**Agent-first is the default.** On a Linux x64 server, prepare Docker
Compose v2, Git, curl, and Node.js 24.21+ or 26.10+. Point two DNS names
(e.g., `review.example.com` and `api.review.example.com`) to the server
and allow inbound ports 80/443. Default Memory also requires Python 3 and tar.
Give your coding agent shell access to that server and run:
```bash
bash -c 'installer=$(mktemp) && curl -fsSL https://raw.githubusercontent.com/777genius/review-router-ai/main/scripts/install-self-hosted.sh -o "$installer" && bash "$installer" "$@"; result=$?; rm -f "$installer"; exit "$result"' -- \
--web-url https://review.example.com \
--repo OWNER/REPOSITORY
```
Replace the example domain and repository. The API domain defaults to
`https://api.review.example.com`; configure its DNS record too.
The installer generates private credentials, starts services and HTTPS,
and returns **one JSON status** with the next action. It resumes without
rebuilding a healthy installation. Agent mode needs no interactive terminal;
to follow prompts yourself, add `--interactive` to the same command.
You still approve the GitHub App, connect the provider through **GitHub Secrets**,
and approve the workflow setup PR. In `--review-only` mode, install the canonical
workflow manually instead. Subscription-account Gateway is a separate service.
A running server is **not** proof of a completed AI review.
This moving-branch command is for evaluation. For production,
[pin an approved source commit](./docs/operations/review-router-self-hosted-end-to-end.md#agent-first-installation-default).
See [installation, resume and first-review steps](./docs/operations/review-router-self-hosted-end-to-end.md#guided-installation).
## Review engine
ReviewRouter uses [OpenCodeReview](https://github.com/alibaba/open-code-review).
Its [published benchmarks](https://github.com/alibaba/open-code-review#benchmark)
are upstream results, not independent ReviewRouter measurements.
## Current status
| Integration | What to expect |
| ----------- | ------------------------------------------------------------------ |
| GitHub | App installation, setup PRs, inline review and a shared dashboard. |
| GitLab | In development. |
| Self-hosted | Guided Docker Compose installation on your infrastructure. |
Live PR progress comments require deployment-specific rollout. See
[capabilities and limits](./docs/capabilities.md) for availability, the review
runtime boundaries and privacy details.
## Learn more
- [Self-hosting guide](./docs/operations/review-router-self-hosted-end-to-end.md)
- [Review settings](./docs/review-settings.md)
- [Development and checks](./docs/development.md)
- [All documentation](./docs/README.md)
This repository contains the dashboard, API, worker and review orchestration.
The public GitHub Action entrypoint lives in
[777genius/review-router](https://github.com/777genius/review-router).
## Licensing
ReviewRouter original contributions are distributed under [Apache License 2.0](LICENSE). It uses [OpenCodeReview](https://github.com/alibaba/open-code-review) (also Apache-2.0); third-party components keep their separate notices. See [`NOTICE`](NOTICE), the bundled font license and the action's third-party license inventory.