voice-vibecoder 2.26.0__tar.gz → 2.26.1__tar.gz

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/CHANGELOG.md

@@ -2,6 +2,28 @@
 
 <!-- version list -->
 
+## v2.26.1 (2026-07-22)
+
+### Bug Fixes
+
+- Per-connection dispatch state via ContextVars
+  ([#7](https://github.com/snokam/voice-vibecoder/pull/7),
+  [`1d28e3b`](https://github.com/snokam/voice-vibecoder/commit/1d28e3b045a8d0f7623fb0b0d6faa46aedd09357))
+
+### Chores
+
+- Remove leftover hello-world test artifact
+  ([`1f6f47c`](https://github.com/snokam/voice-vibecoder/commit/1f6f47c98ac203203e8acdce3047a9510072f3fb))
+
+### Documentation
+
+- Add CLAUDE.md so agents on this repo have context
+  ([`c7267cc`](https://github.com/snokam/voice-vibecoder/commit/c7267ccf76fabc2579f53264047ef9de504e8897))
+
+- Sell the real product and delegate non-merge git to agents
+  ([`a83ceb5`](https://github.com/snokam/voice-vibecoder/commit/a83ceb536bdd55e7c2199cd9af4f9571f70ffd07))
+
+
 ## v2.26.0 (2026-07-13)
 
 ### Features

voice_vibecoder-2.26.1/CLAUDE.md

@@ -0,0 +1,62 @@
+# voice-vibecoder (Olaf The Vibecoder)
+
+Voice-controlled coding assistant: OpenAI/Azure Realtime API for conversation,
+Claude Code agents in git worktrees for the actual coding. The user talks;
+Olaf converses and delegates invisibly.
+
+## Product principles
+
+- The conversation must feel human: short answers, no self-interruption,
+  barge-in works, follow the user's language, resume where you left off.
+- The voice bot stays thin: conversation + delegation. Only `merge_branch`
+  touches git natively (speed + protecting main); all other git operations
+  are delegated to the coding agent in plain words.
+- Agents work invisibly. Olaf never says "I'll send this to an agent."
+
+## Architecture map
+
+- `session.py` — conversation-floor state machine. A response is active from
+  ResponseCreated to ResponseDone; background injections (agent results,
+  narration) queue until the floor is free. Cancel on SpeechStarted only when
+  a response is in flight.
+- `audio.py` — sounddevice playback + fallback mic gating (echo gate).
+- `audio_macos.py` — echo-cancelled capture via AVAudioEngine Voice
+  Processing I/O (system AEC). Mic stays open; server VAD handles barge-in.
+  PyObjC gotcha: `varlist.as_buffer(n)` takes an element count, not bytes.
+- `voice_providers/` — provider protocol; `openai.py` speaks both GA
+  (`response.output_audio.*`, nested session schema) and preview wire formats.
+  Azure GA: transcription must be `whisper-1` (`gpt-4o-mini-transcribe`
+  fails server-side); pass `language` for accuracy.
+- `tools/` — the model's tool surface (definitions.py + dispatch.py +
+  handlers.py). New tools need all three plus `_SLOW_TOOLS` in session.py if
+  they do subprocess I/O.
+- `worktrees.py` — worktree lifecycle; always prune stale registrations
+  (deleted directories stay registered in git as "prunable").
+- `ui/` — Textual TUI. `screen.py` orchestrates; panels/callbacks are mixins.
+- `server.py` — websocket bridge for remote clients (sync_state protocol).
+
+## Testing
+
+- `uv run --with pytest pytest tests/` — must pass before any push.
+- Live-endpoint testing: synthesize speech with macOS `say` + `afconvert`
+  to 24kHz PCM16, stream over the websocket using `RealtimeSettings.load()`
+  credentials. Verify conversation behavior against the real API, not just
+  unit tests.
+
+## Releases
+
+- Conventional commits pushed to main ARE the release mechanism:
+  python-semantic-release cuts the version, updates CHANGELOG, publishes to
+  PyPI on every push (feat → minor, fix/perf → patch).
+- Never hand-edit VERSION or CHANGELOG.
+- Downstream consumer: `snokam/monorepo` `olaf/bridge` pins this package —
+  bump it after notable releases.
+
+## Gotchas
+
+- `app.py` is a PEP 723 script with `editable = true` — required, or uv
+  serves stale cached builds of the package.
+- `uv tool` installs live in a separate environment from the checkout; the
+  in-app updater detects install kind via `direct_url.json`.
+- CI runs import/syntax/build only (no pytest) across Python 3.10–3.13; keep
+  code 3.10-compatible, and mark macOS-only deps with `sys_platform == 'darwin'`.

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: voice-vibecoder
-Version: 2.26.0
+Version: 2.26.1
 Summary: Olaf The Vibecoder — a voice-controlled coding assistant using OpenAI Realtime API + Claude Code
 Project-URL: Homepage, https://github.com/snokam/voice-vibecoder
 Project-URL: Repository, https://github.com/snokam/voice-vibecoder
@@ -33,7 +33,18 @@ Description-Content-Type: text/markdown
 
 # Olaf The Vibecoder
 
-A voice-controlled coding assistant that combines OpenAI's Realtime API for voice conversation with Claude Code for autonomous coding tasks.
+Talk to your codebase like a colleague. You speak an idea; Olaf holds a natural
+voice conversation while coding agents (Claude Code and friends) quietly build
+it on separate git branches. Ask to see the diff, say "merge it", and it lands
+on main — the whole loop, hands-free.
+
+```
+you:   "lag en fil med et dikt i repoet"
+olaf:  "Klart, jeg setter i gang!"          → agent builds on branch 'dikt'
+you:   "vis diff"                            → colorized diff in the TUI
+you:   "merge det inn i main"
+olaf:  "Ferdig! Dikt er merget inn i main."
+```
 
 ## Quick Start
 
@@ -47,12 +58,20 @@ On first launch, a setup wizard will guide you through connecting your OpenAI or
 
 ## Features
 
-- Voice conversation via OpenAI Realtime API (supports OpenAI and Azure)
-- Multi-instance Claude Code agents running in separate git worktrees
-- Git diff visualization with file tree and colorized hunks
+- **Natural conversation** — full-duplex voice: interrupt Olaf mid-sentence and
+  it stops and listens (system echo cancellation on macOS, adaptive echo gate
+  elsewhere). Short spoken answers, follows your language, resumes previous
+  conversations where they left off
+- **Invisible agent orchestration** — multi-instance Claude Code agents in
+  separate git worktrees; Olaf narrates progress and delivers results
+- **Voice-driven git** — show diffs, merge branches into main, clean up
+  worktrees, all by voice; conflicts are reported, never half-merged
+- Live file tree per agent (including files the agent just created) and
+  GitHub-style diff view
 - Session persistence across restarts
 - Push-to-talk and voice activity detection modes
-- Multi-language support (English, Norwegian, Swedish)
+- OpenAI and Azure OpenAI realtime (GA voices marin/cedar), Gemini Live
+- Multi-language (English, Norwegian, Swedish) — spoken and transcribed
 
 ## Requirements
 

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/README.md

@@ -1,6 +1,17 @@
 # Olaf The Vibecoder
 
-A voice-controlled coding assistant that combines OpenAI's Realtime API for voice conversation with Claude Code for autonomous coding tasks.
+Talk to your codebase like a colleague. You speak an idea; Olaf holds a natural
+voice conversation while coding agents (Claude Code and friends) quietly build
+it on separate git branches. Ask to see the diff, say "merge it", and it lands
+on main — the whole loop, hands-free.
+
+```
+you:   "lag en fil med et dikt i repoet"
+olaf:  "Klart, jeg setter i gang!"          → agent builds on branch 'dikt'
+you:   "vis diff"                            → colorized diff in the TUI
+you:   "merge det inn i main"
+olaf:  "Ferdig! Dikt er merget inn i main."
+```
 
 ## Quick Start
 
@@ -14,12 +25,20 @@ On first launch, a setup wizard will guide you through connecting your OpenAI or
 
 ## Features
 
-- Voice conversation via OpenAI Realtime API (supports OpenAI and Azure)
-- Multi-instance Claude Code agents running in separate git worktrees
-- Git diff visualization with file tree and colorized hunks
+- **Natural conversation** — full-duplex voice: interrupt Olaf mid-sentence and
+  it stops and listens (system echo cancellation on macOS, adaptive echo gate
+  elsewhere). Short spoken answers, follows your language, resumes previous
+  conversations where they left off
+- **Invisible agent orchestration** — multi-instance Claude Code agents in
+  separate git worktrees; Olaf narrates progress and delivers results
+- **Voice-driven git** — show diffs, merge branches into main, clean up
+  worktrees, all by voice; conflicts are reported, never half-merged
+- Live file tree per agent (including files the agent just created) and
+  GitHub-style diff view
 - Session persistence across restarts
 - Push-to-talk and voice activity detection modes
-- Multi-language support (English, Norwegian, Swedish)
+- OpenAI and Azure OpenAI realtime (GA voices marin/cedar), Gemini Live
+- Multi-language (English, Norwegian, Swedish) — spoken and transcribed
 
 ## Requirements
 

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "voice-vibecoder"
-version = "2.26.0"
+version = "2.26.1"
 description = "Olaf The Vibecoder — a voice-controlled coding assistant using OpenAI Realtime API + Claude Code"
 readme = "README.md"
 requires-python = ">=3.10"

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/src/voice_vibecoder/__init__.py

@@ -3,7 +3,7 @@
 A voice-controlled coding assistant using OpenAI Realtime API + Claude Code.
 """
 
-__version__ = "2.26.0"
+__version__ = "2.26.1"
 
 from voice_vibecoder.app_config import VoiceConfig
 from voice_vibecoder.ui.screen import VoiceCodingScreen

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/src/voice_vibecoder/config.py

@@ -80,6 +80,7 @@ Avbryt/reset:
 Merge:
 - "merge det", "ta det inn i main" → merge_branch. Ved konflikter: foreslå å be \
 agenten merge main inn i sin branch først.
+- Andre git-operasjoner (push, rebase, revert) → send_to_agent med brukerens ord.
 
 Visning:
 - "vis diff" → show_diff
@@ -139,6 +140,7 @@ Cancel/reset:
 Merge:
 - "merge it", "bring it into main" → merge_branch. On conflicts: suggest asking \
 the agent to merge main into its branch first.
+- Other git operations (push, rebase, revert) → send_to_agent with the user's words.
 
 Views:
 - "show diff" → show_diff
@@ -198,6 +200,7 @@ Avbryt/reset:
 Merge:
 - "merga det", "ta in det i main" → merge_branch. Vid konflikter: föreslå att \
 agenten mergar main in i sin branch först.
+- Övriga git-operationer (push, rebase, revert) → send_to_agent med användarens ord.
 
 Visning:
 - "visa diff" → show_diff

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/src/voice_vibecoder/session.py

@@ -12,6 +12,7 @@ via the voice provider so ChatGPT can summarize it to the user.
 """
 
 import asyncio
+import contextvars
 import logging
 import time
 from pathlib import Path
@@ -622,9 +623,14 @@ class RealtimeSession:
                 # External tools are always async — pass user JWT for OBO flows
                 result = await handle_tool_call_async(name, arguments, self._user_jwt)
             elif name in self._SLOW_TOOLS:
-                # Run slow I/O tools in executor to avoid blocking
+                # Run slow I/O tools in executor to avoid blocking. Copy the
+                # context so the worker thread sees this connection's dispatch
+                # state (ContextVars don't cross into executor threads).
                 loop = asyncio.get_event_loop()
-                result = await loop.run_in_executor(None, handle_tool_call, name, arguments)
+                ctx = contextvars.copy_context()
+                result = await loop.run_in_executor(
+                    None, lambda: ctx.run(handle_tool_call, name, arguments)
+                )
             else:
                 # Sync tools can run directly
                 result = handle_tool_call(name, arguments)

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/src/voice_vibecoder/tools/agent_task.py

@@ -5,6 +5,7 @@ background thread, and relaying interactive questions back through voice.
 """
 
 import asyncio
+import contextvars
 import json
 import logging
 import threading
@@ -19,6 +20,17 @@ logger = logging.getLogger(__name__)
 
 ANSWER_TIMEOUT = 300  # 5 min to answer a question
 
+
+def _start_background(target, *args) -> None:
+    """Run ``target`` in a daemon thread that inherits the current context.
+
+    Dispatch state (registry, repo root, enabled agents, ...) lives in
+    ContextVars, which do not propagate to bare threads — so copy the context
+    and run inside it, else the thread reads process defaults.
+    """
+    ctx = contextvars.copy_context()
+    threading.Thread(target=lambda: ctx.run(target, *args), daemon=True).start()
+
 _CLAUDE_JSON = Path.home() / ".claude.json"
 
 # Cached MCP servers — loaded once per session, files rarely change
@@ -168,20 +180,10 @@ def _handle_send_to_agent(message: str, branch: str | None = None, agent: str |
     is_helper = _repo_root and str(instance.worktree_path) == str(_repo_root)
 
     if is_helper and _helper_override:
-        thread = threading.Thread(
-            target=_run_helper_override_background,
-            args=(instance, message, registry),
-            daemon=True,
-        )
-        thread.start()
+        _start_background(_run_helper_override_background, instance, message, registry)
         return "Sent to helper."
 
-    thread = threading.Thread(
-        target=_run_agent_background,
-        args=(instance, message, registry),
-        daemon=True,
-    )
-    thread.start()
+    _start_background(_run_agent_background, instance, message, registry)
 
     agent_label = get_agent(instance.agent_type).label
     return f"Sent to {agent_label}."
@@ -249,12 +251,7 @@ def _handle_send_to_session(message: str, session_name: str | None = None, agent
 
     # Route through helper override if registered (backward compat)
     if _helper_override:
-        thread = threading.Thread(
-            target=_run_helper_override_background,
-            args=(instance, message, registry),
-            daemon=True,
-        )
-        thread.start()
+        _start_background(_run_helper_override_background, instance, message, registry)
         result = f"Sent to session '{instance.display_name}'."
         if is_new_session:
             all_sessions = registry.get_all_sessions()
@@ -263,12 +260,7 @@ def _handle_send_to_session(message: str, session_name: str | None = None, agent
                 result += f" (Active sessions: {', '.join(names)})"
         return result
 
-    thread = threading.Thread(
-        target=_run_agent_background,
-        args=(instance, message, registry),
-        daemon=True,
-    )
-    thread.start()
+    _start_background(_run_agent_background, instance, message, registry)
 
     result = f"Sent to session '{instance.display_name}'."
     if is_new_session:

{voice_vibecoder-2.26.0 → voice_vibecoder-2.26.1}/src/voice_vibecoder/tools/dispatch.py

@@ -11,6 +11,7 @@ tools are automatically included in the tool definitions sent to OpenAI and
 dispatched through ``handle_tool_call_async``.
 """
 
+import contextvars
 import logging
 from pathlib import Path
 from typing import Any, Callable, Awaitable
@@ -21,12 +22,52 @@ from voice_vibecoder.instances import AgentInstance, InstanceRegistry, InstanceT
 
 logger = logging.getLogger(__name__)
 
-_registry: InstanceRegistry | None = None
-_repo_root: Path | None = None
-_agent_timeout: int = 300
-_permission_mode: PermissionMode = "bypass"
-_enabled_agents: list[str] = ["claude"]
-_voice_config: VoiceConfig | None = None
+# Per-connection dispatch state. Held in ContextVars so concurrent realtime
+# sessions in one process each configure and read their own registry/config
+# instead of clobbering shared module globals. The names below are still
+# exposed as module attributes via ``__getattr__`` so callers may keep doing
+# ``from voice_vibecoder.tools.dispatch import _registry`` inside a function.
+_registry_var: contextvars.ContextVar[InstanceRegistry | None] = contextvars.ContextVar(
+    "vv_registry", default=None
+)
+_repo_root_var: contextvars.ContextVar[Path | None] = contextvars.ContextVar(
+    "vv_repo_root", default=None
+)
+_agent_timeout_var: contextvars.ContextVar[int] = contextvars.ContextVar(
+    "vv_agent_timeout", default=300
+)
+_permission_mode_var: contextvars.ContextVar[PermissionMode] = contextvars.ContextVar(
+    "vv_permission_mode", default="bypass"
+)
+_enabled_agents_var: contextvars.ContextVar[list[str] | None] = contextvars.ContextVar(
+    "vv_enabled_agents", default=None
+)
+_voice_config_var: contextvars.ContextVar[VoiceConfig | None] = contextvars.ContextVar(
+    "vv_voice_config", default=None
+)
+
+_SCOPED_VARS: dict[str, contextvars.ContextVar] = {
+    "_registry": _registry_var,
+    "_repo_root": _repo_root_var,
+    "_agent_timeout": _agent_timeout_var,
+    "_permission_mode": _permission_mode_var,
+    "_enabled_agents": _enabled_agents_var,
+    "_voice_config": _voice_config_var,
+}
+
+
+def __getattr__(name: str) -> Any:
+    var = _SCOPED_VARS.get(name)
+    if var is None:
+        raise AttributeError(name)
+    value = var.get()
+    if name == "_enabled_agents" and not value:
+        return ["claude"]
+    return value
+
+
+def _enabled_agents_or_default() -> list[str]:
+    return _enabled_agents_var.get() or ["claude"]
 
 # ── External tool registry ───────────────────────────────────────────────────
 # Maps tool name → async handler(name, arguments, user_jwt) -> str
@@ -87,21 +128,33 @@ def configure(
     user_jwt: str | None = None,
     voice_config: VoiceConfig | None = None,
 ) -> None:
-    global _registry, _repo_root, _agent_timeout, _permission_mode, _enabled_agents, _voice_config
-    _registry = registry
-    _repo_root = repo_root
-    _agent_timeout = agent_timeout
-    _permission_mode = permission_mode
-    _enabled_agents = enabled_agents or ["claude"]
-    _voice_config = voice_config
+    _registry_var.set(registry)
+    _repo_root_var.set(repo_root)
+    _agent_timeout_var.set(agent_timeout)
+    _permission_mode_var.set(permission_mode)
+    _enabled_agents_var.set(enabled_agents or ["claude"])
+    _voice_config_var.set(voice_config)
+
+
+def set_repo_root(repo_root: Path) -> None:
+    """Point the current context's dispatch at another repo (per-connection)."""
+    _repo_root_var.set(repo_root)
+    registry = _registry_var.get()
+    if registry is not None:
+        registry.repo_root = str(repo_root)
+
+
+def current_repo_root() -> Path | None:
+    return _repo_root_var.get()
 
 
 def _default_agent_type() -> str:
     """Return the default agent type from config, or first enabled agent."""
-    if _voice_config and _voice_config.default_agent:
-        logger.info(f"Using default_agent from config: {_voice_config.default_agent}")
-        return _voice_config.default_agent
-    result = _enabled_agents[0] if _enabled_agents else "claude"
+    voice_config = _voice_config_var.get()
+    if voice_config and voice_config.default_agent:
+        logger.info(f"Using default_agent from config: {voice_config.default_agent}")
+        return voice_config.default_agent
+    result = _enabled_agents_or_default()[0]
     logger.info(f"Using fallback agent: {result}")
     return result
 
@@ -120,28 +173,30 @@ def _fuzzy_match_branch(query: str, branches: list[str]) -> str | None:
 
 
 def _resolve_instance(branch: str | None = None, session_name: str | None = None) -> AgentInstance | None:
-    if not _registry:
+    registry = _registry_var.get()
+    if not registry:
         return None
     if session_name:
         return _resolve_session(session_name)
     if branch:
-        inst = _registry.get_by_branch(branch)
+        inst = registry.get_by_branch(branch)
         if inst:
             return inst
         # Fuzzy match against registered branches
-        all_branches = [i.branch for i in _registry.get_all()]
+        all_branches = [i.branch for i in registry.get_all()]
         matched = _fuzzy_match_branch(branch, all_branches)
         if matched:
-            return _registry.get_by_branch(matched)
+            return registry.get_by_branch(matched)
         return None
-    return _registry.get_default()
+    return registry.get_default()
 
 
 def _resolve_session(session_name: str) -> AgentInstance | None:
     """Resolve a session instance by display_name (fuzzy match)."""
-    if not _registry:
+    registry = _registry_var.get()
+    if not registry:
         return None
-    return _registry.get_session_by_name(session_name)
+    return registry.get_session_by_name(session_name)
 
 
 def cancel_instance_task(instance: AgentInstance) -> bool:
@@ -154,19 +209,21 @@ def cancel_instance_task(instance: AgentInstance) -> bool:
 
 
 def cancel_all_tasks() -> int:
-    if not _registry:
+    registry = _registry_var.get()
+    if not registry:
         return 0
     count = 0
-    for inst in _registry.get_all():
+    for inst in registry.get_all():
         if cancel_instance_task(inst):
             count += 1
     return count
 
 
 def cancel_active_task() -> bool:
-    if not _registry:
+    registry = _registry_var.get()
+    if not registry:
         return False
-    inst = _registry.get_default()
+    inst = registry.get_default()
     return cancel_instance_task(inst) if inst else False
 
 

voice_vibecoder-2.26.1/tests/test_openai_events.py

@@ -0,0 +1,109 @@
+"""Event mapping for the OpenAI/Azure provider — GA and preview wire formats."""
+
+import pytest
+
+from voice_vibecoder.voice_providers import (
+    AudioCommitted,
+    AudioDelta,
+    FunctionCall,
+    ResponseCreated,
+    ResponseDone,
+    SpeechStarted,
+    SpeechStopped,
+    TranscriptDone,
+    UserTranscript,
+    VoiceError,
+)
+from voice_vibecoder.voice_providers.openai import OpenAIVoiceProvider
+
+
+@pytest.fixture
+def provider():
+    return OpenAIVoiceProvider(ws_url="wss://x", ws_headers={})
+
+
+@pytest.mark.parametrize("msg_type", ["response.output_audio.delta", "response.audio.delta"])
+def test_audio_delta_ga_and_preview(provider, msg_type):
+    ev = provider._map_event({"type": msg_type, "delta": "QUJD"})
+    assert isinstance(ev, AudioDelta)
+    assert ev.delta == "QUJD"
+
+
+@pytest.mark.parametrize("msg_type", [
+    "response.output_audio_transcript.done", "response.audio_transcript.done",
+])
+def test_transcript_done_ga_and_preview(provider, msg_type):
+    ev = provider._map_event({"type": msg_type, "transcript": "hei"})
+    assert isinstance(ev, TranscriptDone)
+    assert ev.transcript == "hei"
+
+
+def test_lifecycle_events(provider):
+    assert isinstance(provider._map_event({"type": "input_audio_buffer.speech_started"}), SpeechStarted)
+    assert isinstance(provider._map_event({"type": "input_audio_buffer.speech_stopped"}), SpeechStopped)
+    assert isinstance(provider._map_event({"type": "input_audio_buffer.committed"}), AudioCommitted)
+    assert isinstance(provider._map_event({"type": "response.created"}), ResponseCreated)
+    assert isinstance(provider._map_event({"type": "response.done"}), ResponseDone)
+
+
+def test_user_transcript_stripped(provider):
+    ev = provider._map_event({
+        "type": "conversation.item.input_audio_transcription.completed",
+        "transcript": " hei du \n",
+    })
+    assert isinstance(ev, UserTranscript)
+    assert ev.transcript == "hei du"
+
+
+def test_function_call_parses_arguments(provider):
+    ev = provider._map_event({
+        "type": "response.function_call_arguments.done",
+        "call_id": "c1",
+        "name": "merge_branch",
+        "arguments": '{"branch": "dikt"}',
+    })
+    assert isinstance(ev, FunctionCall)
+    assert ev.name == "merge_branch"
+    assert ev.arguments == {"branch": "dikt"}
+
+
+def test_function_call_invalid_json_gives_empty_args(provider):
+    ev = provider._map_event({
+        "type": "response.function_call_arguments.done",
+        "call_id": "c1",
+        "name": "x",
+        "arguments": "{broken",
+    })
+    assert ev.arguments == {}
+
+
+def test_cancel_errors_are_ignorable(provider):
+    ev = provider._map_event({
+        "type": "error",
+        "error": {"message": "Cannot cancel: no active response"},
+    })
+    assert isinstance(ev, VoiceError)
+    assert ev.is_ignorable
+
+
+def test_real_errors_are_not_ignorable(provider):
+    ev = provider._map_event({
+        "type": "error",
+        "error": {"message": "Invalid session configuration"},
+    })
+    assert isinstance(ev, VoiceError)
+    assert not ev.is_ignorable
+
+
+def test_unknown_events_are_dropped(provider):
+    assert provider._map_event({"type": "rate_limits.updated"}) is None
+
+
+def test_ga_session_body_includes_transcription_language():
+    p = OpenAIVoiceProvider(
+        ws_url="wss://x", ws_headers={},
+        transcription_model="whisper-1", transcription_language="no", ga=True,
+    )
+    body = p._session_body("inst", "marin", [], None)
+    assert body["audio"]["input"]["transcription"] == {"model": "whisper-1", "language": "no"}
+    assert body["audio"]["output"]["voice"] == "marin"

voice_vibecoder-2.26.0/index.html

@@ -1,30 +0,0 @@
-<!DOCTYPE html>
-<html lang="en">
-<head>
-    <meta charset="UTF-8">
-    <meta name="viewport" content="width=device-width, initial-scale=1.0">
-    <title>Hello World</title>
-    <style>
-        body {
-            margin: 0;
-            display: flex;
-            justify-content: center;
-            align-items: center;
-            min-height: 100vh;
-            background-color: #f0f4f8;
-            font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
-        }
-
-        h1 {
-            font-size: 3rem;
-            color: #2d3748;
-            background: linear-gradient(135deg, #667eea, #764ba2);
-            -webkit-background-clip: text;
-            -webkit-text-fill-color: transparent;
-        }
-    </style>
-</head>
-<body>
-    <h1>Hello, World!</h1>
-</body>
-</html>