agentwire-dev 1.34.0__tar.gz → 1.35.0__tar.gz

agentwire_dev-1.35.0/.github/workflows/tts-smoke.yml

@@ -0,0 +1,104 @@
+# Kokoro default-tier TTS smoke tests (#269).
+#
+# Guards the three packaging promises behind "one-click good voice":
+#   1. the kokoro import path stays torch-free on every OS
+#   2. the dependency set resolves to wheels for Intel macs (onnxruntime
+#      publishes arm64-only current wheels; pip must fall back cleanly)
+#   3. a fresh install can actually synthesize audio
+name: tts-smoke
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - "pyproject.toml"
+      - "agentwire/tts/**"
+      - ".github/workflows/tts-smoke.yml"
+  pull_request:
+    paths:
+      - "pyproject.toml"
+      - "agentwire/tts/**"
+      - ".github/workflows/tts-smoke.yml"
+
+jobs:
+  torchless-import:
+    name: torch-less import (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Install base package
+        run: pip install .
+      - name: Import kokoro engine without torch
+        # Windows is scoped to this import only — the portal/CLI use
+        # fcntl/pty and don't support Windows; the kokoro path must still
+        # install and import there.
+        run: >
+          python -c "import sys;
+          import agentwire.tts.engines.kokoro;
+          import agentwire.tts.local;
+          assert 'torch' not in sys.modules, 'torch leaked into the kokoro import path';
+          assert 'agentwire.tts.engines.chatterbox' not in sys.modules, 'chatterbox eagerly imported';
+          print('torch-free kokoro import OK')"
+
+  intel-mac-wheels:
+    name: Intel-mac wheel resolution
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Resolve kokoro-onnx wheels for macosx x86_64
+        # Verifies pip can resolve the full kokoro-onnx dependency tree to
+        # binary wheels for Intel macs (macOS 13+). Current onnxruntime
+        # releases are arm64-only; resolution must fall back to an older
+        # x86_64 wheel (kokoro-onnx allows >=1.20.1).
+        run: |
+          pip download kokoro-onnx \
+            --platform macosx_13_0_x86_64 \
+            --only-binary=:all: \
+            --python-version 3.12 \
+            -d /tmp/intel-mac-wheels
+          ls /tmp/intel-mac-wheels
+
+  synthesis:
+    name: synthesis smoke (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Cache Kokoro model files
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/kokoro_onnx
+          key: kokoro-model-files-v1.0-fp16
+      - name: Install base package
+        run: pip install .
+      - name: Synthesize a sentence
+        run: |
+          python - <<'EOF'
+          from agentwire.tts.engines.kokoro import KokoroEngine
+          from agentwire.tts.base import TTSRequest
+          from agentwire.tts.audio import pcm_float_to_wav_bytes
+
+          KokoroEngine.download_models()
+          engine = KokoroEngine()
+          result = engine.generate(
+              TTSRequest(text="AgentWire speaks with Kokoro out of the box.", voice="af_heart")
+          )
+          wav = pcm_float_to_wav_bytes(result.audio, result.sample_rate)
+          assert wav[:4] == b"RIFF" and len(wav) > 50_000, f"suspicious WAV ({len(wav)} bytes)"
+          print(f"synthesized {len(wav)} bytes of WAV at {result.sample_rate}Hz")
+          EOF

{agentwire_dev-1.34.0 → agentwire_dev-1.35.0}/CHANGELOG.md

@@ -16,6 +16,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **Kokoro-82M is the default voice** — `tts.backend: default` now speaks via an in-process Kokoro engine (CPU ONNX, torch-free, 32 presets / 8 languages), bundled with the base install (`kokoro-onnx`, py<3.14). The portal downloads the model (~200 MB, one-time, atomic) in the background on first start with a progress toast; browser `speechSynthesis` covers speech until ready and remains the last-resort fallback (model unavailable / py3.14+). No-browser playback upgraded from `say`/`espeak` to Kokoro on local speakers. New `agentwire tts warm` pre-downloads for headless setups; `agentwire tts status` reports default-tier model state; `/api/voice-status` exposes `tts.kokoro {state, percent}`; `/api/voices` + `voiceclone list` surface the presets. The `custom` shim tier is unchanged (chatterbox venvs now need Python 3.13+ — numpy 2 gate). `audio_done` timing now parses real WAV duration instead of a 2x-off stereo estimate. Issue [#269](https://github.com/dotdevdotdev/agentwire-dev/issues/269)
 - **Shareable conversation handoffs** — `agentwire handoff` produces two artifacts from one conversation: `ai-handoff.md` (XML-tagged markdown a teammate can paste into another LLM) and `show-the-story.html` (single-file presentation with tabs and scroll-slides) — issue [#157](https://github.com/dotdevdotdev/agentwire-dev/issues/157)
   - In-conversation agent does the distillation (free — uses existing context, no fresh LLM call); CLI/MCP just renders deterministically via Jinja2
   - `/handoff` slash command walks the agent through filling the template, picking a vibe-matched theme (palette, fonts, motion), and rendering

{agentwire_dev-1.34.0 → agentwire_dev-1.35.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentwire-dev
-Version: 1.34.0
+Version: 1.35.0
 Summary: Multi-session voice web interface for AI coding agents
 Project-URL: Homepage, https://agentwire.dev
 Project-URL: Repository, https://github.com/dotdevdotdev/agentwire-dev
@@ -688,8 +688,10 @@ Requires-Dist: aiohttp-jinja2>=1.6
 Requires-Dist: aiohttp>=3.9.0
 Requires-Dist: av>=12.0.0
 Requires-Dist: jinja2>=3.1.0
+Requires-Dist: kokoro-onnx>=0.4.9; python_version < '3.14'
 Requires-Dist: markdown>=3.5
 Requires-Dist: mcp>=1.2.0
+Requires-Dist: pydantic>=2.0.0
 Requires-Dist: python-dotenv>=1.0.0
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: requests>=2.28.0
@@ -706,7 +708,7 @@ Requires-Dist: soundfile>=0.12.0; extra == 'stt'
 Requires-Dist: useful-moonshine-onnx>=20250101; extra == 'stt'
 Requires-Dist: uvicorn>=0.24.0; extra == 'stt'
 Provides-Extra: tts
-Requires-Dist: chatterbox-tts>=0.1.6; extra == 'tts'
+Requires-Dist: chatterbox-tts>=0.1.6; (python_version >= '3.13') and extra == 'tts'
 Requires-Dist: fastapi>=0.104.0; extra == 'tts'
 Requires-Dist: faster-whisper>=1.0.0; extra == 'tts'
 Requires-Dist: pydantic>=2.0.0; extra == 'tts'
@@ -778,7 +780,7 @@ agentwire portal start
 
 **Requirements:** Python 3.10+, tmux, ffmpeg, Claude Code
 
-**Honest setup time:** under a minute to a working voice portal (instant mode: Chrome speech in, browser voice out — robotic but real). ~15 minutes for the full experience: cloned voices via a self-hosted TTS shim, Whisper-grade transcription, phone-from-anywhere (certs + token).
+**Honest setup time:** under a minute to a working voice portal with a genuinely good voice — Kokoro-82M runs on CPU out of the box (one-time ~200 MB model download in the background; the browser voice covers the wait). ~15 minutes for the full experience: cloned voices via a self-hosted TTS shim, Whisper-grade transcription, phone-from-anywhere (certs + token).
 
 > **Network & trust model.** The portal binds `127.0.0.1` by default — local only. To use it from your phone, set `server.host: 0.0.0.0` in `~/.agentwire/config.yaml`. Non-loopback binds require an auth token (auto-generated on first start; print it with `agentwire portal token`) — your phone prompts for it once, then remembers it. Origin checks reject cross-site browser requests on every bind. Still: keep it on a trusted LAN. Never port-forward it or run it on a public-facing VPS — for internet access use Cloudflare Tunnel + Zero Trust. Details in [SECURITY.md](SECURITY.md).
 
@@ -836,7 +838,7 @@ Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access co
 Hold the mic button, speak your request, release. In instant mode the transcript appears for a quick glance — Enter sends it to Claude Code.
 
 **4. Listen:**
-Agent responses are spoken back — browser voice out of the box, or a real TTS model behind a custom shim.
+Agent responses are spoken back — Kokoro neural voice out of the box, or any TTS model behind a custom shim.
 
 ---
 
@@ -887,9 +889,9 @@ All decisions logged for audit trails.
 
 Two tiers, both sides:
 
-**`default` (zero setup, what a fresh install gets):** Chrome speech recognition in, browser speechSynthesis out, OS voice when no browser is connected. No GPU, no models, no certs.
+**`default` (zero setup, what a fresh install gets):** Chrome speech recognition in, **Kokoro-82M out** — a genuinely good neural voice (top of the TTS Arena at 82M params), 32 preset voices across 8 languages, pure CPU, identical on every OS. The model (~200 MB) downloads in the background on first portal start; browser speechSynthesis covers speech until it's ready (and remains the last-resort fallback). No GPU, no certs, no commands.
 
-**`custom` (bring your own model):** any HTTP shim implementing the [voice shim contract](docs/wiki/voice/shim-contract.md) — ~30 lines wraps anything (Deepgram, whisper.cpp, an expressive emotion-tag model). The bundled servers are reference shims:
+**`custom` (bring your own model):** any HTTP shim implementing the [voice shim contract](docs/wiki/voice/shim-contract.md) — ~30 lines wraps anything (Deepgram, whisper.cpp, an expressive emotion-tag model). Voice cloning, GPU engines, emotion control live here. The bundled servers are reference shims:
 
 ```yaml
 # ~/.agentwire/config.yaml
@@ -910,7 +912,7 @@ Shims can declare capabilities (emotion tags, style instructions) via `GET /capa
 
 Instant mode already needs nothing — just don't press the mic. Agent speech
 plays through the browser; mute the tab (or close it — with no browser
-connected the OS voice handles notifications, which you can silence at the
+connected, speech plays on local speakers, which you can silence at the
 system level).
 
 </details>

{agentwire_dev-1.34.0 → agentwire_dev-1.35.0}/README.md

@@ -59,7 +59,7 @@ agentwire portal start
 
 **Requirements:** Python 3.10+, tmux, ffmpeg, Claude Code
 
-**Honest setup time:** under a minute to a working voice portal (instant mode: Chrome speech in, browser voice out — robotic but real). ~15 minutes for the full experience: cloned voices via a self-hosted TTS shim, Whisper-grade transcription, phone-from-anywhere (certs + token).
+**Honest setup time:** under a minute to a working voice portal with a genuinely good voice — Kokoro-82M runs on CPU out of the box (one-time ~200 MB model download in the background; the browser voice covers the wait). ~15 minutes for the full experience: cloned voices via a self-hosted TTS shim, Whisper-grade transcription, phone-from-anywhere (certs + token).
 
 > **Network & trust model.** The portal binds `127.0.0.1` by default — local only. To use it from your phone, set `server.host: 0.0.0.0` in `~/.agentwire/config.yaml`. Non-loopback binds require an auth token (auto-generated on first start; print it with `agentwire portal token`) — your phone prompts for it once, then remembers it. Origin checks reject cross-site browser requests on every bind. Still: keep it on a trusted LAN. Never port-forward it or run it on a public-facing VPS — for internet access use Cloudflare Tunnel + Zero Trust. Details in [SECURITY.md](SECURITY.md).
 
@@ -117,7 +117,7 @@ Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access co
 Hold the mic button, speak your request, release. In instant mode the transcript appears for a quick glance — Enter sends it to Claude Code.
 
 **4. Listen:**
-Agent responses are spoken back — browser voice out of the box, or a real TTS model behind a custom shim.
+Agent responses are spoken back — Kokoro neural voice out of the box, or any TTS model behind a custom shim.
 
 ---
 
@@ -168,9 +168,9 @@ All decisions logged for audit trails.
 
 Two tiers, both sides:
 
-**`default` (zero setup, what a fresh install gets):** Chrome speech recognition in, browser speechSynthesis out, OS voice when no browser is connected. No GPU, no models, no certs.
+**`default` (zero setup, what a fresh install gets):** Chrome speech recognition in, **Kokoro-82M out** — a genuinely good neural voice (top of the TTS Arena at 82M params), 32 preset voices across 8 languages, pure CPU, identical on every OS. The model (~200 MB) downloads in the background on first portal start; browser speechSynthesis covers speech until it's ready (and remains the last-resort fallback). No GPU, no certs, no commands.
 
-**`custom` (bring your own model):** any HTTP shim implementing the [voice shim contract](docs/wiki/voice/shim-contract.md) — ~30 lines wraps anything (Deepgram, whisper.cpp, an expressive emotion-tag model). The bundled servers are reference shims:
+**`custom` (bring your own model):** any HTTP shim implementing the [voice shim contract](docs/wiki/voice/shim-contract.md) — ~30 lines wraps anything (Deepgram, whisper.cpp, an expressive emotion-tag model). Voice cloning, GPU engines, emotion control live here. The bundled servers are reference shims:
 
 ```yaml
 # ~/.agentwire/config.yaml
@@ -191,7 +191,7 @@ Shims can declare capabilities (emotion tags, style instructions) via `GET /capa
 
 Instant mode already needs nothing — just don't press the mic. Agent speech
 plays through the browser; mute the tab (or close it — with no browser
-connected the OS voice handles notifications, which you can silence at the
+connected, speech plays on local speakers, which you can silence at the
 system level).
 
 </details>

{agentwire_dev-1.34.0 → agentwire_dev-1.35.0}/agentwire/__init__.py

@@ -1,3 +1,3 @@
 """AgentWire - Multi-session voice web interface for AI coding agents."""
 
-__version__ = "1.34.0"
+__version__ = "1.35.0"