OpenClawNet.E2ETests

End-to-end validation that a chat request through the gateway uses an enabled skill from the K-1b registry, with Azure OpenAI as the live model.

What runs here#

Test	Live?	What it proves
`SkillRegistryRoundtripTests.Skills_Endpoints_RoundTripPerAgentEnable`	API only	Boots the gateway, lists the seeded system skills via `GET /api/skills`, flips a per-agent enable through `PATCH /api/skills/enabled`, and re-reads the skill to confirm the override persisted.
`AzureOpenAiSkillsChatTests.Chat_BaselineWithoutSkills_StreamsAssistantContent`	Live AOAI	Hits `POST /api/chat/stream` with no skills enabled and asserts the NDJSON stream contains non-empty assistant content. Baseline that proves Azure OpenAI auth + the streaming pipeline are wired.
`AzureOpenAiSkillsChatTests.Chat_WithEnabledSkill_RespectsSkillInstruction`	Live AOAI	Installs the `banana-suffix` fixture skill via `POST /api/skills`, enables it for the default agent, runs a chat turn, and asserts the model output contains the literal word `BANANA` — i.e. the per-agent skill overlay reached the model. Note: today this auto-skips with a diagnostic message because `/api/chat/stream` bypasses `ChatClientAgent`, so `OpenClawNetSkillsProvider` never fires on the streaming path. See `.squad/decisions/inbox/dylan-e2e-skills-chat-wiring-gap.md`. The hard assertion is authored — once the wiring gap is closed the `Skip.IfNot` falls through and the test asserts BANANA strictly.

All tests are tagged:

[Trait("Category", "Live")]
[Trait("Layer",    "E2E")]
[SkippableFact]

So the default dotnet test filter (Category!=Live) runs zero tests in this project. The skill-registry round-trip is intentionally tagged Live so the whole project is opt-in behind a single filter; it does not actually call Azure and will pass even when AOAI is unconfigured (so long as the filter includes it).

Required environment variables (for live runs)#

The two AOAI tests skip cleanly with Skip.IfNot(...) unless these are set:

Variable	Required?	Notes
`AZURE_OPENAI_ENDPOINT`	yes	e.g. `https://my-resource.openai.azure.com/`
`AZURE_OPENAI_DEPLOYMENT` or `AZURE_OPENAI_DEPLOYMENT_NAME`	yes	The deployment name to call.
`AZURE_OPENAI_KEY` or `AZURE_OPENAI_API_KEY`	yes¹	API key auth.
`AZURE_OPENAI_AUTH_MODE`	optional	Set to `integrated` to use `DefaultAzureCredential` instead of an API key.

¹ Either the key or AZURE_OPENAI_AUTH_MODE=integrated is required.

As a developer convenience the tests also fall back to the gateway's Model:Endpoint / Model:ApiKey / Model:DeploymentName / Model:AuthMode user secrets (UserSecretsId c15754a6-dc90-4a2a-aecb-1233d1a54fe1) — the same secrets the existing LiveLlmTests and AzureOpenAILiveTests use.

Running#

All commands assume $env:NUGET_PACKAGES="$env:USERPROFILE\.nuget\packages2" is set first (squad convention).

# Build only.
dotnet build tests\OpenClawNet.E2ETests --verbosity quiet
 
# Default — runs the full repo suite WITHOUT the Live E2E tests (0 tests in this project).
dotnet test tests\OpenClawNet.E2ETests --filter "Category!=Live" --nologo --verbosity quiet
 
# Live run — requires the env vars above.
dotnet test tests\OpenClawNet.E2ETests --filter "Category=Live" --nologo --verbosity quiet

When the env vars are not set, the live filter run still succeeds — the AOAI tests skip with a Skip.IfNot reason explaining what to set.

Hermetic isolation#

Each test class spins up its own GatewayE2EFactory, which:

Creates a temp folder under %TEMP%\openclawnet-e2e\<guid>\.
Sets OPENCLAWNET_STORAGE_ROOT to that folder so the gateway's skills layers (skills/system, skills/installed, skills/agents) all resolve inside the temp scope.
Deletes the folder on Dispose.
Replaces the EF Core SQLite factory with an InMemory one so no DB file is written.
Each test additionally tears down any per-agent enabled.json it touched and disables any skill it enabled, so test ordering doesn't matter.

Skill fixture#

TestData/Skills/banana-suffix/SKILL.md is the fixture used by E2E-3. It instructs the model to end every reply with the literal uppercase word BANANA. The test installs it through the gateway's POST /api/skills endpoint (Installed-layer authoring per K-1b L-2) so the create path is also exercised end-to-end.