cade-cli 0.15.4__tar.gz → 0.15.6__tar.gz

{cade_cli-0.15.4 → cade_cli-0.15.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cade-cli
-Version: 0.15.4
+Version: 0.15.6
 Summary: Cade - The CLI Agent from Arcade.dev
 Project-URL: Homepage, https://arcade.dev
 Project-URL: Documentation, https://docs.arcade.dev
@@ -43,7 +43,7 @@ Classifier: Topic :: Software Development
 Classifier: Topic :: Software Development :: Code Generators
 Classifier: Typing :: Typed
 Requires-Python: >=3.11
-Requires-Dist: agent-library<1.0.0,>=0.14.0
+Requires-Dist: agent-library<1.0.0,>=0.14.1
 Requires-Dist: anthropic<1.0.0,>=0.34.0
 Requires-Dist: arcade-core<5.0.0,>=4.7.0
 Requires-Dist: arcade-mcp-server<2.0.0,>=1.22.0
@@ -123,7 +123,7 @@ The install script detects your OS/arch, downloads the matching prebuilt binary
 `~/.local/bin` — no GitHub token required. Pin a version or install location:
 
 ```bash
-CADE_VERSION=0.15.4 CADE_INSTALL_DIR=/usr/local/bin \
+CADE_VERSION=0.15.5 CADE_INSTALL_DIR=/usr/local/bin \
   sh -c "$(curl -fsSL https://arcadeagent.dev/install)"
 ```
 

{cade_cli-0.15.4 → cade_cli-0.15.6}/README.md

@@ -35,7 +35,7 @@ The install script detects your OS/arch, downloads the matching prebuilt binary
 `~/.local/bin` — no GitHub token required. Pin a version or install location:
 
 ```bash
-CADE_VERSION=0.15.4 CADE_INSTALL_DIR=/usr/local/bin \
+CADE_VERSION=0.15.5 CADE_INSTALL_DIR=/usr/local/bin \
   sh -c "$(curl -fsSL https://arcadeagent.dev/install)"
 ```
 

{cade_cli-0.15.4 → cade_cli-0.15.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "cade-cli"
-version = "0.15.4"
+version = "0.15.6"
 description = "Cade - The CLI Agent from Arcade.dev"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -38,7 +38,7 @@ dependencies = [
     "prompt-toolkit>=3.0.52,<4.0.0",
     "httpx>=0.27.0,<1.0.0",
     "filelock>=3.0.0,<4.0.0",
-    "agent-library>=0.14.0,<1.0.0",
+    "agent-library>=0.14.1,<1.0.0",
     "croniter>=2.0.0,<4.0.0",
     "python-telegram-bot>=21.0,<22.0",
     "keyring>=24.0,<26.0",

cade_cli-0.15.6/src/cade_mcp_local/web/extract.py

@@ -0,0 +1,115 @@
+"""Secondary-model extraction for WebFetch.
+
+Single non-streaming ``Provider.complete()`` call so the same code path
+works for Anthropic, OpenAI, and Ollama users alike. The extraction model is
+resolved from (in priority order) an explicit override, ``CADE_WEB_FETCH_MODEL``,
+the user's configured default model, then ``DEFAULT_EXTRACT_MODEL``. When the
+preferred model's provider isn't configured (e.g. a Claude default on an
+OpenAI-only setup), extraction falls back to whatever provider the user *did*
+configure rather than failing outright.
+"""
+
+from __future__ import annotations
+
+import os
+
+from cade_mcp_local.errors import WebFetchError
+from cade_mcp_local.web.prompts import make_secondary_model_prompt
+from cadecoder.providers import get_provider_for_model, provider_registry
+from cadecoder.providers.base import Provider, ProviderRequest, ProviderType
+
+DEFAULT_EXTRACT_MODEL = "claude-haiku-4-5-20251001"
+
+# Cheap/fast per-provider extraction models, used only as a fallback when the
+# preferred model's provider is unavailable but another provider is registered.
+# Extraction is a one-shot summarization pass, so a small model is sufficient.
+_FALLBACK_MODEL_BY_PROVIDER: dict[ProviderType, str] = {
+    ProviderType.ANTHROPIC: DEFAULT_EXTRACT_MODEL,
+    ProviderType.OPENAI: "gpt-5-mini",
+    ProviderType.OLLAMA: "llama3.1",
+}
+
+
+def _configured_default_model() -> str | None:
+    """Return the user's configured default model, or None if unavailable.
+
+    Reads the global settings directly so the extraction step honors whatever
+    provider/model the user set up during onboarding.
+    """
+    try:
+        from cadecoder.core.config import _load_global_settings
+
+        return _load_global_settings().model or None
+    except Exception:  # noqa: BLE001 - config is best-effort here; fall back to default
+        return None
+
+
+def _resolve_extraction_target(model: str | None) -> tuple[str, Provider]:
+    """Resolve the extraction model and a usable provider.
+
+    Args:
+        model: An explicit model override, if any.
+
+    Returns:
+        A ``(model, provider)`` tuple ready to run an extraction.
+
+    Raises:
+        LookupError: If no LLM provider is registered at all.
+    """
+    chosen = (
+        model
+        or os.environ.get("CADE_WEB_FETCH_MODEL")
+        or _configured_default_model()
+        or DEFAULT_EXTRACT_MODEL
+    )
+
+    provider = get_provider_for_model(chosen)
+    if provider is not None:
+        return chosen, provider
+
+    # The preferred model's provider isn't configured. Fall back to whatever
+    # provider the user did configure (the registry's default) so non-preapproved
+    # fetches still work on OpenAI/Ollama-only setups.
+    fallback = provider_registry.get_default()
+    if fallback is None:
+        raise LookupError("no LLM provider is registered")
+
+    fallback_model = _FALLBACK_MODEL_BY_PROVIDER.get(fallback.provider_type, chosen)
+    return fallback_model, fallback
+
+
+async def extract_with_model(
+    prompt: str,
+    markdown: str,
+    *,
+    model: str | None = None,
+    url: str = "",
+) -> tuple[str, str]:
+    """Run a one-shot extraction. Returns (answer, model_used).
+
+    ``url`` is only used to make the error message useful when no provider is
+    available or the provider rejects the request.
+    """
+    try:
+        chosen, provider = _resolve_extraction_target(model)
+    except LookupError:
+        raise WebFetchError(
+            url or "(no url)",
+            "no provider available for extraction; set ANTHROPIC_API_KEY/OPENAI_API_KEY, "
+            "configure a model via cade onboarding, or override CADE_WEB_FETCH_MODEL",
+        ) from None
+
+    full_prompt = make_secondary_model_prompt(prompt, markdown)
+    request = ProviderRequest(
+        messages=[{"role": "user", "content": full_prompt}],
+        model=chosen,
+        max_tokens=2048,
+        temperature=0.2,
+        stream=False,
+    )
+    try:
+        response = await provider.complete(request)
+    except Exception as e:
+        raise WebFetchError(url or "(no url)", f"extraction model failed: {e}") from e
+
+    return (response.content or "").strip(), chosen

{cade_cli-0.15.4 → cade_cli-0.15.6}/src/cadecoder/cli/commands/mem.py

@@ -43,6 +43,10 @@ def _apply_memory_env() -> Path:
     os.environ["DOCUMENTS_PATH"] = str(librarian_dir / "documents")
     os.environ["SOURCES_CONFIG_PATH"] = str(librarian_dir / "sources.json")
     os.environ["HOME"] = str(persona_dir)
+    # Rebrand librarian's usage hints (e.g. "Add a source with: libr add ...")
+    # to "cade mem". Must be set before librarian.cli is imported, which only
+    # happens later via _load_librarian_app().
+    os.environ["LIBRARIAN_PROG_NAME"] = "cade mem"
     return persona_dir
 
 

{cade_cli-0.15.4 → cade_cli-0.15.6}/src/cadecoder/ui/display.py

@@ -465,7 +465,6 @@ _PRIMARY_ARG_KEYS: tuple[str, ...] = (
     # Generic last-resort labels.
     "name",
     "title",
-    "tool_name",
 )
 
 
@@ -811,6 +810,18 @@ def get_display_tool_name(tool_name: str, args: dict[str, Any]) -> str:
 
 def get_tool_context(tool_name: str, args: dict[str, Any]) -> str:
     """Return parenthesized context for the tool (used by session for 'Calling tool' line)."""
+    # Unwrap the Arcade meta-tools so the context describes the *inner* tool
+    # (e.g. the doc title) instead of echoing the wrapper's tool_name arg —
+    # which produced redundant headlines like
+    # ``GoogleDocs_CreateDocumentFromText("GoogleDocs_CreateDocumentFromText")``.
+    if tool_name in ("Arcade_UseTool", "Arcade_ExecuteTool"):
+        inner_name = get_display_tool_name(tool_name, args)
+        if inner_name and inner_name != tool_name:
+            inner_args = args.get("inputs")
+            if not isinstance(inner_args, dict):
+                inner_args = {}
+            return get_tool_context(inner_name, inner_args)
+
     server, tool = _parse_tool_name(tool_name)
     handler = _find_registry_handler(server, tool)
     if handler:

{cade_cli-0.15.4 → cade_cli-0.15.6}/src/cadecoder/voice/audio.py

@@ -52,6 +52,34 @@ def _rms(audio: np.ndarray) -> float:
     return float(np.sqrt(np.mean(audio.astype(np.float64) ** 2)))
 
 
+# Substrings that identify a Bluetooth headset reported as an input device.
+# Such devices frequently return all-zero samples on macOS because the input
+# route fails to switch to the headset mic — a routing issue, not a permission
+# one — so the silence error gives different advice in that case.
+_BLUETOOTH_MIC_HINTS: Final[tuple[str, ...]] = (
+    "airpods",
+    "beats",
+    "bluetooth",
+    "headphones",
+    "headset",
+    "buds",
+    "powerbeats",
+)
+
+
+def _looks_like_bluetooth_mic(device_name: str) -> bool:
+    """Return True when *device_name* looks like a Bluetooth headset.
+
+    Args:
+        device_name: The human-readable input device name from PortAudio.
+
+    Returns:
+        True if the name matches a known Bluetooth-headset substring.
+    """
+    name = device_name.lower()
+    return any(hint in name for hint in _BLUETOOTH_MIC_HINTS)
+
+
 def check_mic_permission(
     device: str | int | None = None,
     sample_rate: int = DEFAULT_SAMPLE_RATE,
@@ -103,6 +131,27 @@ def check_mic_permission(
         import platform
 
         if platform.system() == "Darwin":
+            if _looks_like_bluetooth_mic(device_name):
+                # Bluetooth headsets (AirPods, Beats, etc.) commonly report
+                # silence on macOS because the input route fails to switch to
+                # the headset's mic — a routing problem, not a permission one.
+                return False, (
+                    f"Microphone '{device_name}' returns silence (all zeros).\n"
+                    "Bluetooth headsets (AirPods, Beats, etc.) often can't be "
+                    "used as a mic on macOS.\n"
+                    "\n"
+                    "Try one of these:\n"
+                    "  • Switch input to the built-in mic: System Settings > "
+                    "Sound > Input > MacBook/Built-in Microphone\n"
+                    "  • Or pin a specific device in ~/.cade/config/cade.toml "
+                    'under [voice]:  audio_input_device = "MacBook Pro Microphone"\n'
+                    "    (type /devices in the chat to list available devices)\n"
+                    "  • Or disconnect/reconnect the headset and try again\n"
+                    "\n"
+                    "If you intended to use the built-in mic, also confirm your "
+                    "terminal has mic access in System Settings > Privacy & "
+                    "Security > Microphone."
+                )
             return False, (
                 f"Microphone '{device_name}' returns silence (all zeros).\n"
                 "This usually means macOS is blocking microphone access.\n"

{cade_cli-0.15.4 → cade_cli-0.15.6}/src/cadecoder/voice/stt.py

@@ -7,8 +7,6 @@ Models are downloaded from HuggingFace on first use and cached automatically.
 import asyncio
 import logging
 import os
-import tempfile
-import wave
 
 import numpy as np
 
@@ -17,6 +15,10 @@ from cadecoder.core.errors import VoiceError
 from cadecoder.core.logging import log
 from cadecoder.voice.audio import DEFAULT_SAMPLE_RATE
 
+# Whisper always operates at 16 kHz. When given a numpy array (rather than a
+# file path) mlx-whisper does NOT resample, so the samples must already be 16 kHz.
+_WHISPER_SAMPLE_RATE = 16_000
+
 
 def _suppress_hf_logging() -> None:
     """Suppress noisy HuggingFace Hub download/progress logging.
@@ -109,17 +111,21 @@ class WhisperSTT:
             raise VoiceError("STT model not loaded — call _ensure_loaded() first")
 
         try:
-            # mlx_whisper.transcribe accepts a file path or numpy array.
-            # Write to a temp WAV file for maximum compatibility across
-            # mlx-whisper versions.
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=True) as tmp:
-                self._write_wav(tmp.name, audio, sample_rate)
-                result = self._mlx_whisper.transcribe(  # type: ignore[union-attr]
-                    tmp.name,
-                    path_or_hf_repo=self.model,
-                    language=self.language,
-                    fp16=False,
-                )
+            # Pass the raw float32 samples directly. Writing a temp WAV and
+            # passing the *path* makes mlx-whisper shell out to `ffmpeg` to
+            # decode it (not bundled in the frozen binary, and often absent on
+            # user machines). Whisper accepts a 16 kHz mono float32 array with
+            # no external tools — the recorder already produces exactly that.
+            samples = np.asarray(audio, dtype=np.float32).flatten()
+            if sample_rate != _WHISPER_SAMPLE_RATE:
+                samples = _resample(samples, sample_rate, _WHISPER_SAMPLE_RATE)
+
+            result = self._mlx_whisper.transcribe(  # type: ignore[union-attr]
+                samples,
+                path_or_hf_repo=self.model,
+                language=self.language,
+                fp16=False,
+            )
 
             text = result.get("text", "").strip()
             log.debug(f"Transcribed {len(audio) / sample_rate:.1f}s -> '{text[:80]}'")
@@ -128,13 +134,19 @@ class WhisperSTT:
         except Exception as exc:
             raise VoiceError(f"Transcription failed: {exc}") from exc
 
-    @staticmethod
-    def _write_wav(path: str, audio: np.ndarray, sample_rate: int) -> None:
-        """Write a float32 numpy array to a 16-bit PCM WAV file."""
-        # Convert float32 [-1.0, 1.0] to int16
-        pcm = (audio * 32767).clip(-32768, 32767).astype(np.int16)
-        with wave.open(path, "wb") as wf:
-            wf.setnchannels(1)
-            wf.setsampwidth(2)  # 16-bit
-            wf.setframerate(sample_rate)
-            wf.writeframes(pcm.tobytes())
+
+def _resample(samples: np.ndarray, src_rate: int, dst_rate: int) -> np.ndarray:
+    """Linearly resample mono float32 audio to ``dst_rate`` (no external deps).
+
+    Adequate for speech fed to Whisper. Only used on the rare path where the
+    recorder's rate differs from Whisper's 16 kHz.
+    """
+    if src_rate == dst_rate or samples.size == 0:
+        return samples
+    duration = samples.shape[0] / float(src_rate)
+    n_out = int(round(duration * dst_rate))
+    if n_out <= 0:
+        return np.array([], dtype=np.float32)
+    x_old = np.linspace(0.0, duration, num=samples.shape[0], endpoint=False)
+    x_new = np.linspace(0.0, duration, num=n_out, endpoint=False)
+    return np.interp(x_new, x_old, samples).astype(np.float32)

cade_cli-0.15.4/src/cade_mcp_local/web/extract.py

@@ -1,54 +0,0 @@
-"""Secondary-model extraction for WebFetch.
-
-Single non-streaming ``Provider.complete()`` call so the same code path
-works for Anthropic, OpenAI, and Ollama users alike. Default model is
-``claude-haiku-4-5-20251001``; override via ``CADE_WEB_FETCH_MODEL``.
-"""
-
-from __future__ import annotations
-
-import os
-
-from cade_mcp_local.errors import WebFetchError
-from cade_mcp_local.web.prompts import make_secondary_model_prompt
-from cadecoder.providers import get_provider_for_model
-from cadecoder.providers.base import ProviderRequest
-
-DEFAULT_EXTRACT_MODEL = "claude-haiku-4-5-20251001"
-
-
-async def extract_with_model(
-    prompt: str,
-    markdown: str,
-    *,
-    model: str | None = None,
-    url: str = "",
-) -> tuple[str, str]:
-    """Run a one-shot extraction. Returns (answer, model_used).
-
-    ``url`` is only used to make the error message useful when the
-    provider rejects the request.
-    """
-    chosen = model or os.environ.get("CADE_WEB_FETCH_MODEL") or DEFAULT_EXTRACT_MODEL
-    provider = get_provider_for_model(chosen)
-    if provider is None:
-        raise WebFetchError(
-            url or "(no url)",
-            f"no provider available for extraction model {chosen!r}; "
-            "set ANTHROPIC_API_KEY/OPENAI_API_KEY or CADE_WEB_FETCH_MODEL",
-        )
-
-    full_prompt = make_secondary_model_prompt(prompt, markdown)
-    request = ProviderRequest(
-        messages=[{"role": "user", "content": full_prompt}],
-        model=chosen,
-        max_tokens=2048,
-        temperature=0.2,
-        stream=False,
-    )
-    try:
-        response = await provider.complete(request)
-    except Exception as e:
-        raise WebFetchError(url or "(no url)", f"extraction model failed: {e}") from e
-
-    return (response.content or "").strip(), chosen