docs: add a voice agent guide to Learn

[joshua-mo-143]

Summary

Adds a Learn guide that builds a terminal voice agent on Venice, walking through the reference implementation at joshua-mo-143/venice-voice-agent-demo.

Venice has no realtime speech-to-speech socket, which reads like a gap until you notice a voice agent is three ordinary HTTP calls in a loop. This is the same STT → LLM → TTS shape as the existing LiveKit Agents guide, with the framework removed so the three requests are visible.

New page: learn/voice-agent.mdx → /learn/voice-agent. Bylined Joshua Mo, 27 August 2026.

The sections that got the most space are the ones the demo is actually opinionated about:

• include_venice_system_prompt: False and disable_thinking, and what each costs if left on
• Yielding sentences rather than tokens, which is what lets TTS start before the model finishes, plus why the splitter waits for trailing whitespace and how it avoids breaking on Dr. and U.S.
• response_format="pcm" with streaming: True, and the status / content-type / first-chunk checks — a JSON error body written into a raw sink plays as a burst of noise
• Writing to a raw output stream on sample boundaries, since HTTP chunk boundaries don't respect them and a split 16-bit sample byte-shifts everything after it
• The queue-and-thread handoff, the lazily created PcmPlayer, and abort() versus close()

Also includes a privacy section, since this is the first Learn guide that turns on a microphone. Recordings are assembled in memory and given a WAV header with wave rather than being written to a temp file, so there's nothing on disk to clean up.

Supporting changes

• learn.mdx — card in the Projects row. Projects rather than Walkthroughs because every existing Projects entry is repo-backed and every Walkthrough is notebook- or inline-based.
• docs.json — English nav path only.
• llms.txt — one entry under Learn.

Per agents.md, locale copies are untouched for Mintlify to regenerate after merge.

Note on the third commit

The demo repo switched local audio from PipeWire (pw-record / pw-play) to sounddevice / PortAudio after the guide was first written, so it's no longer Linux-only. The third commit rewrites the audio section against the new code and re-checks the surrounding claims — most visibly the privacy section, which described a temp file that no longer exists.

Test plan

Verified:

• All 14 internal doc links resolve to real .mdx files
• docs.json still parses as valid JSON
• Every icon name used already appears elsewhere in the repo (Tabler set)
• Model IDs (nvidia/parakeet-tdt-0.6b-v3, zai-org-glm-5-2, tts-kokoro, af_sky) match those already documented in guides/integrations/livekit-agents.mdx and guides/media/*
• Every symbol the guide names exists in the demo's current audio.py; setup commands and the sounddevice>=0.5.6 pin match its pyproject.toml and .env.example
• No stale PipeWire references left anywhere in the repo
• Demo repo's 18 tests pass at 67dd1e2
• Mintlify Deployment and link-rot checks pass (vale-spellcheck skips)

Worth a reviewer's eye:

• The 14 min read-time estimate on the card, scaled by byte count against neighbouring guides (lands between the RAG bot at 9 min and the Rust gateway at 18 min)
• Whether linking the reference repo under a personal account is still the convention — it matches the four existing joshua-mo-143/venice-*-demo links, but flagging it
• The guide documents specific model IDs while agents.md says to resolve them at runtime. There's a callout saying so, but the code uses env-var defaults like the other guides.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
veniceai	🟢 Ready	View Preview	Aug 27, 2026, 8:17 PM