[Dhirenderchoudhary]

Adds a Workflows page to the dashboard (under Build, behind deployment:list like Sandbox). You drag nodes onto a canvas, connect them, pick a deployed model for each, and run the whole pipeline. Each node shows its status, time in ms and output.

Nodes: audio input (record or upload), text input, speech to text, LLM (with a system prompt), text to speech, text output, audio output. Two built-in templates:

• Banking voice agent: mic -> STT -> LLM -> TTS -> audio out, with a demo bank system prompt and fake customer data.
• Text chat: text -> LLM -> text.

How it works:

• src/lib/workflow/nodeKinds.ts is one registry table. Each node kind declares its input and output type (audio or text), its config fields and a run function, so adding a node type is one entry.
• engine.ts rejects bad connections (type mismatch, self loop, second input, cycle), runs nodes in topological order and stops at the first error.
• Calls go to the inference service with the user's token and x-sandbox: true, the same as Sandbox. STT/TTS use /v1/audio/transcriptions and /v1/audio/speech from Add speech-to-text and text-to-speech endpoints #428.
• Workflows save to localStorage. There's no backend storage yet.
• The ready-deployments query moved out of Sandbox.tsx into useReadyDeployments so both pages share it. Sandbox behaves the same.

Needs #428 to run audio nodes (and for the model dropdowns to tell STT/TTS from LLMs).

Testing:

• 13 new vitest tests for the registry, connection rules, run order, the voice template run (with mocked fetch checking all three requests and that outputs flow node to node) and storage. Full dashboard suite: 432 passed.
• npm run build passes. npm run lint and tsc already fail on main (156 lint errors, 89 tsc lines). This adds none; the new files lint clean.
• Ran it in Chromium against a local stack with Add speech-to-text and text-to-speech endpoints #428, speaches and Ollama, nothing mocked: logged in, loaded the voice template, uploaded a recorded question. STT 576 ms -> "What is the balance on my checking account?", LLM 1.2 s -> "Your checking balance is $4,820.55.", TTS 4.8 s on CPU -> playable audio.

Live latency per node (ticking while a node runs, then p50/p95 over recent runs, plus total and time to first audio):

[Priyanshu-u07]

Tested locally: the 440 dashboard tests and the build pass, and the API calls match #428. Two fixes before merge:

"Stop and run" uses the state from when Record was clicked (recorder.onstop keeps that render's setAudio). Reproduced with a test that renders the Workflows page with a fake mic and inference server:

• Record while a run is in flight, let it finish, then Stop and run: the recording is stored but never runs. Easy to hit in a live demo when asking the next question while the previous answer is still processing.
• Change the LLM model while recording: the old model is sent.

Calling setAudio through a ref fixes both (verified with the same test). Please also replace the silent if (running) return; in run() with a toast, so stopping during a run isn't silent.

TTS autofill can pick the STT model: the TTS pattern matches bare speech, so deployments ["Speech Recognition", "Kokoro"] put both nodes on "Speech Recognition". Tightening both patterns and skipping deployments already given to another node fixes it. Default speaches names aren't affected.

[Priyanshu-u07]

Pushed the fixes for both bugs above. Looks good to merge now.