voice-vibecoder 2.20.2__tar.gz → 2.21.0__tar.gz

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/CHANGELOG.md

@@ -2,6 +2,22 @@
 
 <!-- version list -->
 
+## v2.21.0 (2026-04-12)
+
+### Features
+
+- **session**: Narrate background tool events to voice user
+  ([`9c28a4f`](https://github.com/snokam/voice-vibecoder/commit/9c28a4f2f6a3f2bc6e6243f11a629ab0e369d7ee))
+
+
+## v2.20.3 (2026-02-23)
+
+### Bug Fixes
+
+- Skip transcript and instance replay for text-only clients
+  ([`ee30512`](https://github.com/snokam/voice-vibecoder/commit/ee30512f1bebf7e8009d815a872d939aeffacf8b))
+
+
 ## v2.20.2 (2026-02-23)
 
 ### Bug Fixes

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: voice-vibecoder
-Version: 2.20.2
+Version: 2.21.0
 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

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "voice-vibecoder"
-version = "2.20.2"
+version = "2.21.0"
 description = "Olaf The Vibecoder — a voice-controlled coding assistant using OpenAI Realtime API + Claude Code"
 readme = "README.md"
 requires-python = ">=3.10"

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/src/voice_vibecoder/__init__.py

@@ -3,7 +3,7 @@
 A voice-controlled coding assistant using OpenAI Realtime API + Claude Code.
 """
 
-__version__ = "2.20.2"
+__version__ = "2.21.0"
 
 from voice_vibecoder.app_config import VoiceConfig
 from voice_vibecoder.ui.screen import VoiceCodingScreen

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/src/voice_vibecoder/server.py

@@ -222,17 +222,20 @@ class WebSocketBridge:
         if not self._session:
             return
 
-        # Send transcript history (voice conversation)
-        transcripts = self._session.registry.get_transcripts()
-        logger.info(f"[SYNC_STATE] Sending {len(transcripts)} transcripts to frontend")
-        for entry in transcripts:
-            self._send({
-                "type": "transcript",
-                "role": entry["role"],
-                "text": entry["text"],
-            })
+        # Send transcript history (skip for text-only — they start fresh)
+        if not self._session.text_only:
+            transcripts = self._session.registry.get_transcripts()
+            logger.info(f"[SYNC_STATE] Sending {len(transcripts)} transcripts to frontend")
+            for entry in transcripts:
+                self._send({
+                    "type": "transcript",
+                    "role": entry["role"],
+                    "text": entry["text"],
+                })
 
-        # Send instance status and history for all instances
+        # Send instance status and history (skip for text-only)
+        if self._session.text_only:
+            return
         for inst in self._session.registry.get_all():
             self._send({
                 "type": "instance_status",

{voice_vibecoder-2.20.2 → voice_vibecoder-2.21.0}/src/voice_vibecoder/session.py

@@ -13,6 +13,7 @@ via the voice provider so ChatGPT can summarize it to the user.
 
 import asyncio
 import logging
+import time
 from pathlib import Path
 from typing import Any, Callable
 
@@ -49,6 +50,70 @@ logger = logging.getLogger(__name__)
 RECONNECT_MAX_ATTEMPTS = 5
 RECONNECT_DELAYS = [2, 4, 8, 16, 30]
 
+# Debounce between spoken tool-progress narrations. Prevents the model from
+# interrupting itself mid-filler when many tool events fire in quick succession.
+TOOL_NARRATION_DEBOUNCE_SEC = 6.0
+
+
+# Markers used by code-provider runners to tag the kind of tool event:
+#   🔧  tool invocation (tool_call)
+#   ✓   tool completed successfully (tool_result)
+#   ❌  tool failed (tool_result)
+#   ↳   streaming output preview, NOT a completion — skip narration
+_TOOL_CALL_MARKER = "\U0001f527"  # 🔧
+_TOOL_OK_MARKER = "\u2713"        # ✓
+_TOOL_ERR_MARKER = "\u274c"       # ❌
+
+
+def _friendly_tool_hint(line: str) -> str | None:
+    """Turn a registry tool-event line into a short natural-language hint.
+
+    Lines look like `"tool_call:🔧 jira_search({...})"` or
+    `"tool_result:✓ jira_search: 3 issues"`. Returns None for lines that
+    aren't recognized tool starts/completions (e.g. streaming previews)
+    so the caller skips narration.
+    """
+    if not line or ":" not in line:
+        return None
+    category, rest = line.split(":", 1)
+    if category not in ("tool_call", "tool_result"):
+        return None
+    rest = rest.strip()
+    # Must be one of the known markers — otherwise it's a preview or
+    # something we don't want to narrate.
+    if category == "tool_call":
+        if not rest.startswith(_TOOL_CALL_MARKER):
+            return None
+        verb = "using"
+    else:
+        if rest.startswith(_TOOL_OK_MARKER):
+            verb = "finished"
+        elif rest.startswith(_TOOL_ERR_MARKER):
+            verb = "had trouble with"
+        else:
+            return None
+    rest = rest[1:].lstrip()
+    for sep in ("(", ":", " "):
+        if sep in rest:
+            rest = rest.split(sep, 1)[0]
+    raw = rest.strip()
+    if not raw:
+        return None
+    name = raw.split(".", 1)[-1].replace("_", " ").replace("-", " ").strip()
+    if not name:
+        return None
+    if verb == "using":
+        return f"using {name}"
+    if verb == "finished":
+        return f"finished {name}, preparing answer"
+    return f"had trouble with {name}, trying another way"
+
+
+_TOOL_NARRATION_PROMPT = (
+    "[agent background status: {hint}. Let the user know naturally in "
+    "whatever language they are speaking. Don't guess at the results.]"
+)
+
 
 class RealtimeSession:
     """Manages a single voice session with multi-instance support."""
@@ -89,6 +154,7 @@ class RealtimeSession:
         self._response_pending = False
         self._fallback_task: asyncio.Task | None = None
         self._loop: asyncio.AbstractEventLoop | None = None
+        self._last_tool_narration: float = 0.0
 
         # Create instance registry and wire callbacks
         self._registry = InstanceRegistry(user_email=user_email, user_jwt=user_jwt)
@@ -322,6 +388,38 @@ class RealtimeSession:
     def _on_registry_output(self, instance_id: str, line: str) -> None:
         if self._on_agent_output:
             self._on_agent_output(instance_id, line)
+        self._maybe_narrate_tool_event(line)
+
+    def _maybe_narrate_tool_event(self, line: str) -> None:
+        """Speak a short filler phrase when a background tool event fires.
+
+        Without this, a caller hears 30–60s of silence while the agent runs
+        MCP tools. Debounced to TOOL_NARRATION_DEBOUNCE_SEC so many rapid
+        tool events coalesce into at most one spoken hint. Called from
+        background threads, so the injection is bounced to the session's
+        asyncio loop.
+        """
+        if self.text_only or not self._running:
+            return
+        if not isinstance(line, str):
+            return
+        hint = _friendly_tool_hint(line)
+        if not hint:
+            return
+        now = time.monotonic()
+        if (now - self._last_tool_narration) < TOOL_NARRATION_DEBOUNCE_SEC:
+            return
+        loop = self._loop
+        if loop is None:
+            return
+        self._last_tool_narration = now
+        text = _TOOL_NARRATION_PROMPT.format(hint=hint)
+        try:
+            asyncio.run_coroutine_threadsafe(
+                self._provider.inject_message(text), loop,
+            )
+        except Exception as e:
+            logger.warning("Tool narration inject failed: %s", e)
 
     def _on_registry_status_change(self, instance_id: str, status: InstanceStatus) -> None:
         if self._on_instance_status: