Skip to main content

​
E2E Mock Testing Architecture

To ensure the reliability of the internal/proxy layer, ONR uses a robust E2E Mock testing system. This system validates the end-to-end transformation behavior, including req_map, routing, resp_map, and sse_parse, without actually hitting real upstream providers (saving costs and avoiding flaky tests).

​
Scope and Goals

  • Test Entry Point: open-next-router/internal/proxy/e2e_mock_test.go
  • Key Validations:
    1. Are requests correctly rewritten and routed to the target upstream paths?
    2. Are responses accurately transformed back into the OpenAI-compatible format?
    3. Are SSE termination events (data: [DONE]) ordered correctly and unique?

​
How It Works

  1. Downstream Request Simulation
    • We use gin.CreateTestContext and httptest.NewRecorder to simulate a client requesting ONR.
    • The test directly enters the proxy chain via Client.ProxyJSON(...) without spinning up a full onr process.
  2. Upstream Service Mocking
    • httptest.Server simulates the upstream provider APIs.
    • It returns JSON or SSE text based on the test scenario.
    • The handler asserts the upstream path, query parameters, and critical request fields to ensure the DSL rewrite worked.
  3. DSL Configuration Isolation
    • Tests construct minimal provider DSL strings inline (providerConf*).
    • These are dynamically written to t.TempDir() and loaded using dslconfig.Registry.ReloadFromDir().
    • Each test only loads the provider needed for that scenario, preventing pollution from global configurations.

​
Directory Structure & Conventions

Test data is organized under open-next-router/internal/proxy/testdata/:
  • fixtures/: Contains downstream request bodies (Input samples).
  • mock_upstream/: Contains upstream response bodies (Original JSON/SSE).
  • Naming Convention: Files are organized by provider/scenario_name for easy scaling.

​
Currently Supported Scenarios

  1. Anthropic /v1/messages (Non-streaming)
    • Translates OpenAI chat request -> Anthropic messages request.
    • Translates Anthropic JSON -> OpenAI chat completion JSON.
  2. Anthropic /v1/messages (Streaming tool_use)
    • Translates Anthropic SSE tool_use events -> OpenAI tool_calls chunks.
    • Validates finish_reason=tool_calls and a single [DONE].
  3. Gemini (Streaming)
    • Translates OpenAI chat request -> Gemini :streamGenerateContent?alt=sse.
    • Validates Gemini SSE -> OpenAI chat chunks (including the usage chunk).
  4. Azure Responses (Streaming)
    • Translates OpenAI chat request -> /openai/responses?api-version=2025-04-01-preview.
    • Validates [DONE] does not arrive earlier than subsequent tokens.

​
Running the Tests

To run the mock E2E tests, navigate to the open-next-router directory and execute: 
go test ./internal/proxy -run E2EMock -v
Or run the full proxy layer test suite: 
go test ./internal/proxy

​
Golden Baselines

E2E test cases perform both critical field assertions and full Golden output comparisons. Golden files are located at: open-next-router/internal/proxy/testdata/golden/ To avoid false positives caused by timestamps or random IDs, the test standardizes dynamic fields (like created, chatcmpl_*) before comparing. Updating Golden Files: Only do this when you intentionally change the transformation behavior. 
cd open-next-router
UPDATE_GOLDEN=1 go test ./internal/proxy -run E2EMock -v
After updating, run a normal regression to ensure baselines are consistent: 
go test ./internal/proxy -run E2EMock -v