OpenLive 0.2.0: coding agents over ACP, voice cloning, and a hardened live pipeline

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: CI
 
 on:
   push:
-    branches: [main]
+    branches: [main, v2-pipeline-agents]
   pull_request:
 
 jobs:
@@ -23,3 +23,39 @@ jobs:
 
       - name: Typecheck
         run: pnpm typecheck
+
+      - name: Test
+        run: pnpm test
+
+  # Windows parity is code-level (no hardware QA): prove the repo installs,
+  # typechecks, tests, and produces an unsigned Windows build on windows-latest.
+  windows:
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: pnpm/action-setup@v6
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: 22
+          cache: pnpm
+
+      - run: pnpm install --frozen-lockfile
+
+      - name: Typecheck
+        run: pnpm typecheck
+
+      - name: Test
+        run: pnpm test
+
+      - name: Pack web + agent
+        run: |
+          pnpm --filter @openlive/desktop pack:web
+          pnpm --filter @openlive/desktop pack:agent
+
+      # --publish never: this is a build-only smoke test. Without it,
+      # electron-builder's publish auto-detection demands a GH_TOKEN in
+      # pull_request runs and fails the job after a successful build.
+      - name: Build Windows installer (unsigned)
+        run: pnpm --filter @openlive/desktop exec electron-builder --win --publish never

.gitignore

@@ -16,3 +16,6 @@ apps/desktop/dist/
 
 # demo videos are hosted as GitHub assets, not committed (100MB limit + no inline playback)
 *.mp4
+
+# vendored voice runtime assets (copied from node_modules by copy-voice-assets.mjs)
+apps/web/public/vad/

CHANGELOG.md

@@ -4,6 +4,111 @@ All notable changes to OpenLive are recorded here. The newest version is on top.
 Releases before 0.1.9 predate this file — see the
 [GitHub releases](https://github.com/katipally/openlive/releases) for those.
 
+## [0.2.0] - 2026-07-16
+
+### Fixed
+- **"The web service keeps crashing" on launch (#6).** A crashed or force-killed
+  run could leave the app's server processes alive (Windows especially), still
+  holding OpenLive's ports — every later launch then died in an EADDRINUSE
+  respawn loop that relaunching never fixed. The app now clears its own stale
+  server processes from those ports at startup (other apps' processes are left
+  alone), and the crash dialog says what to check if it ever still happens.
+- **The workspace you picked is now the workspace the agent gets.** A bind race
+  on new conversations could silently strand the session with no agent and no
+  folder while the top bar showed both (the server's boot-time bind restore could
+  supersede and reverse the client's bind). The server now yields to the client's
+  bind, always receives the folder explicitly (empty means clear, not "keep
+  stale"), and echoes back what it actually bound so the UI can't drift — with a
+  one-shot self-heal re-bind if they ever disagree.
+- **No more silent brain swap.** If a coding agent is selected but can't run
+  (no folder, failed start), OpenLive says exactly that instead of quietly
+  answering with the built-in assistant as if it were the agent.
+- **Speaking to answer a permission no longer cancels it.** Starting to talk
+  while an agent asked for permission counted as barge-in: the ask vanished
+  mid-answer and your words became a new turn. Speech during a pending ask is
+  now the answer ("yes"/"no", matched against the agent's real option ids).
+- **Model/mode pickers stop blinking out.** Resumed sessions kept their model
+  and mode lists (updates during session replay were dropped), the server
+  re-sends them on a same-bind reconnect, and the in-call top bar falls back to
+  the per-agent cache like the lobby always did.
+- **Pre-call verification.** The lobby now checks the project folder actually
+  exists on disk, the built-in brain's provider has an API key, and a microphone
+  is present — gaps surface as chips (with a jump to the right Settings tab)
+  before Start instead of failing mid-call. A missing folder also gets a clear
+  error instead of a baffling "spawn npx ENOENT".
+- **Cross-process data race.** Settings and conversations are written by both the
+  web and agent processes; every read-modify-write now runs under a file lock, so
+  a concurrent save can no longer silently drop the other side's update.
+- **Agent plans and usage now actually show.** The server has always emitted the
+  agent's working plan (ACP plan updates) and context/cost usage — the UI dropped
+  both. Plans render as a live checklist above the transcript; a context/cost chip
+  sits in the top bar.
+- **Permission asks no longer time out silently.** An unanswered agent permission
+  auto-denies after 2 minutes — the prompt now shows a visible countdown and the
+  voice speaks a reminder 30 seconds before the deadline.
+- **Hermes session history.** Discovery was querying columns that don't exist in
+  hermes' database; rewritten against the real hermes-agent 0.18.2 schema.
+- **History with huge session logs.** Reading titles from Codex rollout logs
+  (hundreds of MB) no longer loads whole files into memory.
+- **Workspace file confinement.** The built-in assistant's file tools now refuse
+  symlinks that point outside the workspace, not just `../` escapes.
+- Crash screen follows the OS theme and uses the brand accent.
+
+### Added
+- **Clone Voice — clone your own voice.** A dedicated Settings tab: pick a script (or
+  just talk), record 5–30 seconds with a live level meter, listen back before
+  saving, fix the auto-transcript, and your assistant speaks as you — zero-shot
+  cloning (ZipVoice, Apache-2.0, via sherpa-onnx) running locally in the agent
+  service at ~4x realtime on CPU. Optional ~208 MB download, removable anytime.
+  Profiles preview with any text you type, rename, export/import between
+  machines, play back their original recording, and delete; automatic Kokoro
+  fallback; consent required — clone only your own voice or one you have
+  permission for. The Pipeline TTS stage just picks among your cloned voices.
+- **Persona.** Settings → General gains "Your assistant's style": your own words
+  on how it should behave and speak, applied to the built-in assistant AND every
+  coding agent via its session preamble.
+- **Spoken progress narration (opt-in).** While a coding agent works in silence,
+  OpenLive voices its plan steps ("Step 2 of 4 — …"), throttled and barge-in aware.
+- **Notifications + menu bar.** A tray icon (Open / Mini mode / Quit) and OS
+  notifications when a turn finishes or an agent asks permission while you're in
+  another app — clicking brings OpenLive forward.
+- **Markdown transcript.** Agent replies render as real markdown — code blocks
+  with copy buttons, lists, tables — plus per-message copy and a one-tap export
+  of the whole conversation to a Markdown file.
+- **In-call keyboard shortcuts.** M mute, C camera, S screen share, T activity
+  panel, H history, Cmd/Ctrl-E end call — press `?` for the cheat sheet.
+- **Lobby readiness check.** Picking an agent that isn't installed or signed in
+  shows a one-tap jump to Settings → Agents instead of failing the call.
+- **A real player for voice previews.** Everything Clone Voice plays back — the
+  recorded take, a synthesized preview, the original recording — now goes through
+  a compact seekable player (play/pause, drag to seek, elapsed/total time) instead
+  of fire-and-forget playback. Only one plays at a time.
+- **Agent sign-in that can't strand you.** Sign-in and setup flows open in your
+  terminal; the row now polls while you finish there and flips to Ready by itself.
+  If the terminal can't open (macOS Automation permission), the panel explains the
+  fix and a Copy command button gives you the manual path. Hermes gets an honest
+  "Setup incomplete" state (its wizard was started but no provider picked), a
+  "Finish setup" button, and an Uninstall that removes `~/.hermes` after a warning.
+
+### Changed
+- **Light mode rebuilt.** A stepped warm-paper ladder (no pure white): cards,
+  panels, and popovers now separate cleanly instead of fusing into one white
+  field, and borders are actually visible.
+- **README and docs rewritten** around what OpenLive is: the open voice and
+  vision layer for AI agents — bring your own model, with coding agents over ACP
+  as the flagship integration. Includes an honest note that the pipeline is
+  cascaded, not full-duplex speech-to-speech.
+- **VAD assets are served from the app itself** (vendored at build time) instead
+  of a CDN — the voice loop no longer touches jsdelivr at runtime.
+- **Agents settings shows each CLI's version** and gains an **Update** button;
+  a failed npm install from a root-owned prefix now gets actionable guidance
+  instead of a raw error dump.
+- Slash-command metadata (never surfaced in the UI) removed from the wire protocol.
+- **Settings reorganized.** The Voices tab is now **Clone Voice**. Speaking speed
+  and "Narrate agent progress" moved from Pipeline → Text-to-speech to General
+  under a new **Voice & speech** group, next to voice input — everyday preferences
+  in General, engine choices in Pipeline. Same settings underneath; nothing resets.
+
 ## [0.1.9] - 2026-07-11
 
 ### Added

CONTRIBUTING.md

@@ -30,16 +30,22 @@ pnpm desktop:build:win # build the Windows installer (run on Windows)
 ## Where things live
 
 ```
-apps/desktop     Electron shell: local servers, permissions, window, auto-update
-apps/web         Next.js UI + the on-device voice engine in src/lib/live
-services/agent   the /live WebSocket and the model tools
+apps/desktop     Electron shell: local servers, permissions, window, mini mode,
+                 tray + notifications, auto-update
+apps/web         Next.js UI + the on-device voice engine in src/lib/live + /api
+                 routes (agent install/auth, history discovery, settings)
+services/agent   the /live WebSocket, the ACP coding-agent driver (agents/*),
+                 local voice cloning (voice/*), and the built-in model tools
 packages/harness model adapters (Anthropic / OpenAI Responses / OpenAI Chat), model listing
-packages/shared  the wire protocol and shared types
+packages/shared  the agent registry (single source of agent identity), wire
+                 protocol, shared types
 packages/db      JSON-file store for keys, settings, conversations
 ```
 
-The voice loop (VAD, STT, end-of-turn, TTS, barge-in) is in
-`apps/web/src/lib/live`. The model turn goes out from `services/agent`.
+The voice loop (VAD, STT, end-of-turn, TTS — Kokoro, Supertonic, or a cloned
+voice — and barge-in) is in `apps/web/src/lib/live`. The model turn goes out from
+`services/agent`, which either streams a provider reply or drives a coding agent
+(Claude Code, Codex, Cursor, OpenCode, Hermes) over ACP as a child process.
 
 ## Sending a change