The Go API behind Aevor, a review-first developer tool for working through GitHub issues and repository changes. This repository owns the backend; the React client lives in Aevor/Frontend.
The API connects a user's GitHub account, gives them access to selected repositories, builds bounded code context, and supports an issue-to-change workflow. A proposed change is not published just because an AI service produced it: the user has to approve the change set, the applied result is validated, and publishing is a separate step.
- Account and GitHub access: GitHub OAuth with PKCE, Aevor JWT sessions, and encrypted storage for the GitHub access token.
- Repository workspace: list and select repositories, sync issues / pull requests / commits, and clone selected repositories into a controlled workspace. Clone URLs are restricted by an allowlist; local
file://transport is opt-in. - Code context: discover files, apply bounded filtering, extract content, split it into chunks, and create traceable code representations.
- Index and analysis: query an in-memory metadata index, analyze repository context, inspect potential impact, and retrieve saved analyses.
- Issue workflow: analyze an issue, propose a solution, generate a reviewable change set, approve it, apply it to the workspace, validate it, and explicitly publish a branch / commit / pull request.
- Pull-request review: retrieve pull-request details and run a separate analysis of review feedback.
- Supporting endpoints: health and readiness checks, contribution summary, skills, recommendations, and asynchronous job handling.
The route registration in services/api/cmd/server/main.go is the source of truth for the current HTTP surface. This is an actively developed API, not a claim that the whole product is deployed or production-ready.
React client (Aevor/Frontend)
|
| HTTP + Aevor JWT
v
Go API (this repository)
├── auth / users
├── GitHub API client
├── repository and workspace services
├── discovery → filtering → extraction
│ → chunking → representation
├── metadata index and impact analysis
├── issue / change-set / publishing workflow
└── jobs and webhook handling
|
├── PostgreSQL
├── GitHub
└── separately configured AI service
The metadata index is currently in memory, so it is rebuilt rather than treated as durable storage after an API restart. AI-backed routes need a reachable service configured through AI_SERVICE_URL; the API repository does not bundle that model service.
Requirements: Go 1.26.1 or newer, Docker with Compose, and a GitHub OAuth App.
From the repository root, start PostgreSQL:
docker compose -f deployments/docker/docker-compose.yml up -dConfigure the API:
cd services/api
cp .env.example .envFill in the required database values, GitHub OAuth client ID / secret and callback URL, FRONTEND_URL, WORKSPACE_ROOT, JWT_SECRET, and GITHUB_TOKEN_ENCRYPTION_KEY.
Generate the two cryptographic keys independently:
openssl rand -hex 32
openssl rand -hex 32Use the first value for JWT_SECRET and the second for GITHUB_TOKEN_ENCRYPTION_KEY. Do not reuse a key for both purposes. The OAuth callback should match the configured GITHUB_REDIRECT_URL, normally http://localhost:8080/auth/github/callback.
Run the API from services/api:
go run ./cmd/serverThe API listens on port 8080 by default. The frontend's local Vite server normally runs on port 5173; set FRONTEND_URL to that address.
Run from services/api:
gofmt -w .
go vet ./...
go test -race ./...
go build ./...The opt-in PostgreSQL integration tests require AEVOR_TEST_DATABASE_DSN. Unit tests should run without a live GitHub or AI service by using the injected test clients and fakes.
- User identity comes from the verified Aevor JWT, not from a caller-supplied user ID.
- GitHub access tokens are encrypted before they are stored.
- Repository clones are restricted to configured hosts; local file transport is disabled unless explicitly enabled.
- Repository processing has file, byte, chunk and result limits.
- The indexing and representation endpoints return bounded metadata rather than raw source content.
- Generating a proposal does not publish it. Approval, apply, validation and publishing are distinct operations.
docs/development-state.mdanddocs/todo.mdare dated implementation notes, not live status dashboards.- Product and architecture documents are maintained separately in Aevor/docs.