feat: supertonic TTS shared workflow + daily-news voice artifact

[Copilot]

Wraps the supertone-inc/supertonic-py on-device TTS SDK as a reusable shared workflow and wires it into daily-news to produce a spoken digest uploaded as a run artifact.

shared/supertonic.md (new)

• Pre-agent steps: installs supertonic[serve], starts HTTP server on 127.0.0.1:7788, polls readiness up to 3 min (first run downloads ~400 MB model from HuggingFace)
• Model caching: actions/cache stores ~/.cache/supertonic3/ keyed by runner OS and package version — subsequent runs skip the ~400 MB download entirely; cache is automatically invalidated when the supertonic package version changes
• Agent prompt: full curl reference for both endpoints, voice/language/quality tables, artifact upload walkthrough

# native endpoint
curl -X POST http://127.0.0.1:7788/v1/tts \
  -H 'content-type: application/json' \
  -d '{"text":"...", "voice":"F1", "lang":"en", "steps":8}' \
  -o output.wav

# OpenAI-compatible alias
curl -X POST http://127.0.0.1:7788/v1/audio/speech \
  -H 'content-type: application/json' \
  -d '{"model":"supertonic-3", "input":"...", "voice":"F1", "response_format":"wav"}' \
  -o output.wav

daily-news.md (updated)

• Imports shared/supertonic.md
• timeout-minutes 30 → 50 to cover model download + synthesis
• Adds 🔊 Voice Summary phase: agent composes a 2–4 sentence spoken digest, synthesizes WAV via the local server (jq-safe quoting to avoid shell escaping issues), uploads via upload_artifact safe-output, and embeds the download link in the discussion

[github-actions]

Hey @Copilot 👋 — great work adding the Supertonic TTS shared workflow and wiring it into daily-news! The PR description is thorough, the implementation is focused, and the jq-safe quoting pattern for the TTS payload is a nice touch.

One thing that would strengthen this before it leaves draft:

• No validation step — shared/supertonic.md does real work in its pre-agent steps (pip install, server startup, up-to-3-min readiness poll), but there is no smoke-test that confirms the server can actually synthesise audio before the agent tries to call it. A lightweight end-to-end health check right after the readiness loop would surface broken installs or network issues early rather than letting the agent fail mid-run.

If you'd like a hand, assign this prompt to your coding agent:

In .github/workflows/shared/supertonic.md, add a post-startup validation step that runs immediately after the readiness-poll loop.
The step should:
1. Send a minimal TTS request (short phrase, e.g. "test") to (127.0.0.1/redacted) with voice F1, lang en, steps 4.
2. Verify the response is a non-empty WAV file (check that the output file size is greater than 0 bytes).
3. Print a success line including the output file size.
4. Exit non-zero (fail the step) if the health check does not pass, so the job fails fast before the agent starts.

Generated by ✅ Contribution Check · 445.3 AIC · ⌖ 14.6 AIC · ⊞ 24.7K · ◷

[pelikhan]

@copilot store downloaded model in an actions/cache

[Copilot]

@copilot store downloaded model in an actions/cache

Done. Two new steps added to shared/supertonic.md (commit feat: cache supertonic model in actions/cache):

1. Restore (actions/cache/restore@v5.0.5) — runs after install, before the server starts. Cache key: supertonic3-model-${{ runner.os }}-${{ steps.install-supertonic.outputs.version }} with a restore-keys fallback on OS alone.
2. Save (actions/cache/save@v5.0.5) — runs after server readiness is confirmed, guaranteeing the model is fully on disk before the snapshot is taken.

The install step now also writes the package version to $GITHUB_OUTPUT so the key is automatically invalidated whenever the supertonic package version changes.