Feature Lab testing

Run each SDK feature in the existing Copilot playground and understand the verification limits.

The Feature Lab lives in the existing examples/playground application. It uses the workspace SDK packages and shares the dashboard's API keys settings. Open the hosted lab at /playground/features. Each feature links to its API guide. The playground runs SDK source from main; check package availability when using an npm installation.

Start the playground

From the repository root, with Node.js 22+ and the pinned pnpm 11.5.1:

pnpm install
pnpm build:sdk
pnpm --filter playground dev:features

Open the local Feature Lab. Select API keys to configure your providers. Local development can also use privately supplied environment keys; the status panel reports availability without displaying the keys. A key entered in the dialog takes precedence.

Provider	Configuration	Examples
OpenAI	`OPENAI_API_KEY`	Parameters, fallback, Decisions, Responses, image tools and all media
Anthropic	`ANTHROPIC_API_KEY`	Image tools and system prompts
OpenRouter	`OPENROUTER_API_KEY`	Decisions using Jev through the gateway
Direct Jev/TypeSafe	`TYPESAFE_AI_API_KEY`	Direct Decisions
Cloudflare	`CLOUDFLARE_AUTH_TOKEN` and `CLOUDFLARE_ACCOUNT_ID`	Workers AI Decisions

The dialog saves keys in browser storage and sends only the selected provider's credentials in headers to the playground server. Production routes require caller keys and never use shared server environment keys. Do not commit keys. Example runs make real provider calls and use your provider quota.

Test every feature

Start with Run core checks. It runs ten cases: parameters, fallback, Responses, four Decisions providers, two image-tool providers, and Anthropic system prompts. Missing credentials produce skipped, not passed. Media examples run separately.

Example	Action	Expected result
Reasoning parameters	Run with the default reasoning setting, then select `none` and run again	Both answer `OK`; default reasoning omits temperature with a warning, `none` retains it
Compatibility fallback	Run the deliberate invalid-parameter example	One fallback to GPT-4o, zero retries of the invalid request, and an answer
Native Decisions	Run the sample ticket with each configured provider	Named outage/team/impact answers with probabilities; inspect Response details & usage
Responses & replay	Run with Streaming on, then off	The follow-up recalls `ORCHID-742`
Images in tool calls	Run the default image with OpenAI, then Anthropic	The model identifies blue; one tool execution and saved-history replay pass
System prompt compatibility	Run with Anthropic	Both `systemPrompt` and system-role tests return `SYSTEM_OK`
Voice generation · TTS	Enter a sentence, run, and press play	Audible generated speech; Download audio saves the file
Voice input · STT	Upload audio, or record and stop, then run	Transcript matches the speech; no-file runs use a synthetic TTS/STT round trip
Realtime voice	Start a call and follow the controls below	Connected audio, transcripts and controlled tool execution
Image generation & editing	Generate the sample, then upload an image with an edit instruction	A preview and downloadable result for each operation

Uploads are limited to 4 MiB. Microphone recording is capped at 30 seconds; stop before the cap. The SDK's underlying upload/recording limits are documented separately in audio and image tools.

Decision thresholds in this example are illustrative, not calibrated for your business. A low-confidence answer or named refusal can correctly require human review. OpenRouter Decisions does not support image state. Cache misses are valid; this runner does not establish cache savings or production classification accuracy.

Realtime voice controls

Click Start voice call, allow microphone access, and wait for connected.
Speak, check the transcript, and listen for a response. Use Enable audio if your browser blocks playback. The response uses an AI-generated voice.
Click Ask for demo status, then Allow tool. Expect “Blue balloon ready” and one tool execution.
Ask again and Deny tool. The execution count should not increase.
Check Mute/Unmute, Interrupt reply, and End call. Ending the call should release the microphone and close the connection.

The realtime guide explains server setup, permissions, timeouts and cleanup. Run these steps on the actual browsers and audio devices your application supports.

Test inside Copilot chat

Select Responses or Images and click Open Copilot chat. Ask it to inspect the sample image, then ask it to remember ORCHID-742 and recall that code in a later turn. Compare Streaming on/off; the image example also allows selecting Anthropic. This exercises CopilotProvider, tools and conversation history through the real chat UI.

Export results downloads the focused example history with checks, responses and timings. It excludes audio/image bytes, but can include your prompts and model outputs. History clears on reload. Realtime has a separate transcript/event panel and is not included in this export.

What has been verified

On 17 October 2026, the local production build passed eight core provider cases; direct Jev and Cloudflare were skipped because their credentials were unavailable. OpenRouter's Jev result does not certify the direct Jev endpoint.

Nine browser checks passed, covering the core runner, parameter modes, Responses replay, TTS playback, STT round trip/upload, image generation/editing and layout. Seven integration checks passed, covering OpenAI/Anthropic chat in JSON/SSE, synthetic browser recording, a real WebRTC call and browser errors. The SDK, Copilot and route suite passed 532 automated tests separately.

Realtime used synthetic audio with a real provider connection; input transcription, incoming audio energy, approval, mute and disconnect were observed. Tool denial and disconnect during pending approval were tested with protocol fixtures. Physical microphone/speaker quality, Safari/Firefox/mobile devices, production latency and the deployed public URL still need testing. Mobile layout checks do not certify a physical mobile browser.

Deploy the existing project

Build from the workspace root:

pnpm exec turbo run build --filter=playground...

Deploy examples/playground to its existing project. The docs application's /playground/:path* proxy already covers the Feature Lab and its API routes. Deploy the documentation application separately, then verify the public route, key entry, one core run and the media flows. A successful local production build does not mean either deployment has happened.