agent-cli 0.104.0__tar.gz → 0.105.1__tar.gz

{agent_cli-0.104.0 → agent_cli-0.105.1}/.github/workflows/docker.yml

@@ -90,7 +90,7 @@ jobs:
           cache-from: type=gha
           cache-to: type=gha,mode=max
 
-  # Single-target images (transcribe-proxy)
+  # Single-target images (proxies and diarization)
   build-single-target:
     runs-on: ubuntu-latest
     permissions:
@@ -106,8 +106,16 @@ jobs:
             dockerfile: docker/rag-proxy.Dockerfile
           - image: memory-proxy
             dockerfile: docker/memory-proxy.Dockerfile
+          - image: diarization
+            dockerfile: docker/diarization.Dockerfile
 
     steps:
+      - name: Free disk space
+        run: |
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          df -h
+
       - name: Checkout repository
         uses: actions/checkout@v7.0.1
         with:
@@ -153,3 +161,81 @@ jobs:
           labels: ${{ steps.meta.outputs.labels }}
           cache-from: type=gha
           cache-to: type=gha,mode=max
+
+  smoke-diarization:
+    needs: build-single-target
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    permissions:
+      contents: read
+      packages: read
+
+    steps:
+      - name: Free disk space
+        run: |
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          df -h
+
+      - name: Log in to Container Registry
+        uses: docker/login-action@v4.5.1
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Resolve image reference
+        id: image
+        run: |
+          tag="sha-${GITHUB_SHA::7}"
+          echo "ref=${REGISTRY}/${GITHUB_REPOSITORY}-diarization:${tag}" >> "$GITHUB_OUTPUT"
+
+      - name: Pull published image
+        run: docker pull "${{ steps.image.outputs.ref }}"
+
+      - name: Verify package version
+        env:
+          IMAGE_REF: ${{ steps.image.outputs.ref }}
+          RELEASE_TAG: ${{ github.event.release.tag_name }}
+        run: |
+          actual_version=$(
+            docker run --rm --entrypoint /app/.venv/bin/python "$IMAGE_REF" \
+              -c 'import importlib.metadata; print(importlib.metadata.version("agent-cli"))'
+          )
+          if [[ "$GITHUB_EVENT_NAME" == "release" ]]; then
+            expected_version="${RELEASE_TAG#v}"
+            if [[ "${actual_version%%+*}" != "$expected_version" ]]; then
+              echo "Expected agent-cli $expected_version, got $actual_version" >&2
+              exit 1
+            fi
+          fi
+          echo "Verified agent-cli $actual_version"
+
+      - name: Verify health and API routes
+        env:
+          IMAGE_REF: ${{ steps.image.outputs.ref }}
+        run: |
+          container_id=$(docker run --detach --publish 61337:61337 "$IMAGE_REF")
+          trap 'docker rm --force "$container_id" >/dev/null 2>&1 || true' EXIT
+
+          ready=false
+          for _ in {1..30}; do
+            if curl --fail --silent --connect-timeout 1 --max-time 2 \
+              http://localhost:61337/health | jq --exit-status '.status == "healthy"' >/dev/null; then
+              ready=true
+              break
+            fi
+            sleep 2
+          done
+          if [[ "$ready" != "true" ]]; then
+            docker logs "$container_id"
+            exit 1
+          fi
+
+          if ! curl --fail --silent --show-error --connect-timeout 1 --max-time 2 \
+            http://localhost:61337/openapi.json | jq --exit-status \
+              '.paths["/health"].get and .paths["/transcribe"].post and .paths["/diarize"].post' \
+              >/dev/null; then
+            docker logs "$container_id"
+            exit 1
+          fi

{agent_cli-0.104.0 → agent_cli-0.105.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: agent-cli
-Version: 0.104.0
+Version: 0.105.1
 Summary: A suite of AI-powered command-line tools for text correction, audio transcription, and voice assistance.
 Project-URL: Homepage, https://github.com/basnijholt/agent-cli
 Author-email: Bas Nijholt <bas@nijho.lt>

{agent_cli-0.104.0 → agent_cli-0.105.1}/agent_cli/server/cli.py

@@ -403,6 +403,17 @@ def whisper_cmd(  # noqa: C901, PLR0912, PLR0915
             ),
         ),
     ] = False,
+    max_new_tokens: Annotated[
+        int,
+        typer.Option(
+            "--max-new-tokens",
+            min=1,
+            help=(
+                "Maximum output tokens for autoregressive transformers ASR models. "
+                "Increase for unusually long audio"
+            ),
+        ),
+    ] = 4096,
     ttl: Annotated[
         int,
         typer.Option(
@@ -602,6 +613,7 @@ def whisper_cmd(  # noqa: C901, PLR0912, PLR0915
             compute_type=compute_type,
             default_language=default_language,
             trust_remote_code=trust_remote_code,
+            max_new_tokens=max_new_tokens,
             ttl_seconds=ttl,
             cache_dir=cache_dir,
             backend_type=resolved_backend,  # type: ignore[arg-type]

{agent_cli-0.104.0 → agent_cli-0.105.1}/agent_cli/server/whisper/backends/base.py

@@ -32,6 +32,7 @@ class BackendConfig:
     cache_dir: Path | None = None
     default_language: str | None = None
     trust_remote_code: bool = False
+    max_new_tokens: int = 4096
 
 
 class InvalidAudioError(ValueError):

{agent_cli-0.104.0 → agent_cli-0.105.1}/agent_cli/server/whisper/backends/transformers.py

@@ -6,6 +6,7 @@ import asyncio
 import logging
 import tempfile
 import time
+import traceback
 import wave
 from concurrent.futures import ProcessPoolExecutor
 from dataclasses import dataclass
@@ -47,6 +48,10 @@ _MODEL_MAP: dict[str, str] = {
 _REMOTE_CODE_MODEL_PREFIXES = ("CohereLabs/cohere-transcribe",)
 
 
+class TranscriptionTruncatedError(RuntimeError):
+    """Raised when an autoregressive ASR model exhausts its output budget."""
+
+
 def _resolve_model_name(model_name: str) -> str:
     """Resolve a model name to a HuggingFace repo."""
     if "/" in model_name:
@@ -220,6 +225,22 @@ def _move_inputs_to_device(inputs: Any) -> Any:
     return {k: v.to(_state.device) if hasattr(v, "to") else v for k, v in inputs.items()}
 
 
+def _qwen_generation_was_truncated(generated_ids: Any, max_new_tokens: int) -> bool:
+    """Return whether Qwen exhausted its token budget without emitting EOS."""
+    if generated_ids.shape[1] < max_new_tokens:
+        return False
+
+    generation_config = getattr(_state.model, "generation_config", None)
+    eos_token_ids = getattr(generation_config, "eos_token_id", None)
+    if eos_token_ids is None:
+        return True
+    if isinstance(eos_token_ids, int):
+        eos_token_ids = [eos_token_ids]
+
+    last_token_id = generated_ids[:, -1:].tolist()[0][0]
+    return last_token_id not in eos_token_ids
+
+
 def _transcribe_cohere_asr(
     *,
     audio_array: Any,
@@ -276,6 +297,7 @@ def _transcribe_qwen3_asr(
     task: str,
     initial_prompt: str | None,
     duration: float,
+    max_new_tokens: int,
 ) -> dict[str, Any]:
     """Transcribe with Qwen3-ASR's native transformers interface."""
     if task != "transcribe":
@@ -294,11 +316,19 @@ def _transcribe_qwen3_asr(
     with torch.inference_mode():
         output_ids = _state.model.generate(
             **inputs,
-            max_new_tokens=256,
+            max_new_tokens=max_new_tokens,
             do_sample=False,
         )
 
     generated_ids = output_ids[:, inputs["input_ids"].shape[1] :]
+    if _qwen_generation_was_truncated(generated_ids, max_new_tokens):
+        msg = (
+            "Qwen3-ASR transcription reached "
+            f"max_new_tokens={max_new_tokens} before completion. "
+            "Restart the server with a higher --max-new-tokens value "
+            "or split the audio into shorter chunks."
+        )
+        raise TranscriptionTruncatedError(msg)
     parsed = _state.processor.decode(generated_ids, return_format="parsed")[0]
     language = parsed["language"] or effective_language or "unknown"
 
@@ -399,8 +429,8 @@ def _transcribe_with_generate(
     )
 
 
-def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
-    """Run transcription in subprocess. Reuses model from _state."""
+def _transcribe_with_loaded_model(kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Dispatch transcription using the model retained in this worker."""
     if _state.model is None or _state.processor is None:
         msg = "Model not loaded in subprocess. Call _load_model_in_subprocess first."
         raise RuntimeError(msg)
@@ -416,6 +446,7 @@ def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
             task=task,
             initial_prompt=kwargs.get("initial_prompt"),
             duration=_read_wav_duration(wav_path),
+            max_new_tokens=kwargs.get("max_new_tokens", 4096),
         )
 
     audio_array, sample_rate, duration = _read_wav_audio(wav_path)
@@ -449,6 +480,44 @@ def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
     )
 
 
+def _clear_exception_frames(exc: BaseException) -> None:
+    """Release helper locals retained by an exception and its chained failures."""
+    pending = [exc]
+    seen: set[int] = set()
+    while pending:
+        current = pending.pop()
+        if id(current) in seen:
+            continue
+        seen.add(id(current))
+        traceback.clear_frames(current.__traceback__)
+        pending.extend(
+            error for error in (current.__cause__, current.__context__) if error is not None
+        )
+        if isinstance(current, BaseExceptionGroup):
+            pending.extend(current.exceptions)
+
+
+def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Transcribe and return unused CUDA memory while keeping model weights loaded."""
+    if not _state.device or not _state.device.startswith("cuda"):
+        return _transcribe_with_loaded_model(kwargs)
+
+    import torch  # noqa: PLC0415
+
+    try:
+        return _transcribe_with_loaded_model(kwargs)
+    except BaseException as exc:
+        # Failed helper frames otherwise keep temporary tensors live until the
+        # executor serializes the exception, which happens after our finally.
+        # Clear locals without losing the traceback's function names and lines.
+        _clear_exception_frames(exc)
+        raise
+    finally:
+        # Helper frames have returned, so their temporary tensors are now free.
+        # Only unused allocator blocks are released; model weights stay resident.
+        torch.cuda.empty_cache()
+
+
 class TransformersWhisperBackend:
     """ASR backend using HuggingFace transformers.
 
@@ -548,6 +617,7 @@ class TransformersWhisperBackend:
             "default_language": self._config.default_language,
             "task": task,
             "initial_prompt": initial_prompt,
+            "max_new_tokens": self._config.max_new_tokens,
         }
 
         try:

{agent_cli-0.104.0 → agent_cli-0.105.1}/agent_cli/server/whisper/model_manager.py

@@ -30,6 +30,7 @@ class WhisperModelConfig(ModelConfig):
     backend_type: BackendType = "auto"
     default_language: str | None = None
     trust_remote_code: bool = False
+    max_new_tokens: int = 4096
 
 
 class WhisperModelManager:
@@ -50,6 +51,7 @@ class WhisperModelManager:
                 cache_dir=config.cache_dir,
                 default_language=config.default_language,
                 trust_remote_code=config.trust_remote_code,
+                max_new_tokens=config.max_new_tokens,
             ),
             backend_type=config.backend_type,
         )

{agent_cli-0.104.0 → agent_cli-0.105.1}/docker/docker-compose.yml

@@ -240,10 +240,10 @@ services:
     restart: unless-stopped
 
   # Optional transcription proxy with local diarization. Start explicitly with:
-  # docker compose -f docker/docker-compose.yml up --build diarization
+  # docker compose -f docker/docker-compose.yml up diarization
   # Shares port 61337 with transcribe-proxy; run one of them at a time.
   diarization:
-    image: agent-cli-diarization
+    image: ghcr.io/basnijholt/agent-cli-diarization:latest
     build:
       context: ..
       dockerfile: docker/diarization.Dockerfile

{agent_cli-0.104.0 → agent_cli-0.105.1}/docs/commands/server/transcribe-proxy.md

@@ -211,15 +211,13 @@ docker compose -f docker/docker-compose.yml --profile cpu up transcribe-proxy
 
 ### Docker with diarization
 
-Build the image that includes pyannote, PyTorch, and FFmpeg:
+Run the published image, which includes pyannote, PyTorch, and FFmpeg:
 
 ```bash
-docker build -f docker/diarization.Dockerfile -t agent-cli-diarization .
-
 docker run --rm -p 61337:61337 \
   -e HF_TOKEN \
   -v agent-cli-diarization-cache:/home/transcribe/.cache \
-  agent-cli-diarization
+  ghcr.io/basnijholt/agent-cli-diarization:latest
 ```
 
 This runs `/diarize` on CPU. For NVIDIA GPU inference, add `--gpus all -e DIARIZATION_DEVICE=cuda` to `docker run`. For `/transcribe`, also provide your ASR environment variables, for example `-e ASR_PROVIDER=openai -e ASR_OPENAI_BASE_URL=http://your-whisper-host:10301/v1 -e OPENAI_API_KEY=dummy`.
@@ -228,11 +226,17 @@ The optional Compose service can run beside the existing Whisper server:
 
 ```bash
 # HF_TOKEN must be exported in the shell or set in docker/.env.
-docker compose -f docker/docker-compose.yml --profile cpu up --build whisper-cpu diarization
+docker compose -f docker/docker-compose.yml --profile cpu up whisper-cpu diarization
 ```
 
 Use `diarization` in place of `transcribe-proxy`, since both default to port 61337. Set `DIARIZATION_PORT` to use a different host port. The model cache persists in `agent-cli-diarization-cache`. GPU instructions are included beside the service in the Compose file.
 
+To build from source instead, add `--build` to the Compose command or run:
+
+```bash
+docker build -f docker/diarization.Dockerfile -t agent-cli-diarization .
+```
+
 ### Environment Variables
 
 Configure the proxy using environment variables (priority: env var > config file > default):

{agent_cli-0.104.0 → agent_cli-0.105.1}/docs/commands/server/whisper.md

@@ -55,6 +55,7 @@ Run a local ASR server with automatic backend selection based on your platform:
 - **OpenAI-compatible API** at `/v1/audio/transcriptions` - drop-in replacement for OpenAI's Whisper API
 - **Wyoming protocol** for [Home Assistant](https://www.home-assistant.io/) voice integration (Wyoming is the standard protocol for local voice services)
 - **TTL-based memory management** - models unload after idle period, freeing RAM/VRAM
+- **CUDA cache cleanup** - the transformers backend releases unused GPU cache after each transcription while keeping model weights loaded
 - **Multiple models** - run different model sizes with independent TTLs
 - **Background preloading** - downloads start at startup without blocking; use `--preload` to wait
 - **Multi-platform support** - automatically uses the optimal backend for your hardware (`auto` switches to `nemo` for Parakeet models)
@@ -130,6 +131,7 @@ agent-cli server whisper \
 | `--cache-dir` | - | Custom directory for downloaded models (default: HuggingFace cache) |
 | `--default-language` | - | Fallback language code for requests that omit `language`. Required for models that do not support language auto-detection (for example Cohere Transcribe). |
 | `--trust-remote-code` | `false` | Allow Hugging Face model repositories to execute custom Python code. Known supported remote-code ASR models are trusted automatically. |
+| `--max-new-tokens` | `4096` | Maximum output tokens for autoregressive transformers ASR models. Increase for unusually long audio |
 | `--ttl` | `300` | Seconds of inactivity before unloading model from memory. Set to 0 to keep loaded indefinitely |
 | `--preload` | `false` | Load model(s) immediately at startup instead of on first request. Useful for reducing first-request latency |
 | `--host` | `0.0.0.0` | Network interface to bind. Use `0.0.0.0` for all interfaces |
@@ -311,6 +313,8 @@ agent-cli server whisper \
 
 Qwen3-ASR supports transcription, automatic language detection, and context or hotwords through the OpenAI-compatible `prompt` field. Translation and timestamped subtitle output are not currently supported.
 
+The default output budget is 4096 tokens, which accommodates long recordings. For unusually long or dense speech, increase it with `--max-new-tokens`. If Qwen3-ASR exhausts the configured budget, the request fails with a clear error instead of returning a silently truncated transcript.
+
 To install it as the standard `whisper` background daemon:
 
 ```bash

{agent_cli-0.104.0 → agent_cli-0.105.1}/docs/installation/docker.md

@@ -59,7 +59,7 @@ The Docker setup provides:
 | ----------------------- | --------------------------------- | ----------- | ------------------------------ |
 | **whisper**             | agent-cli-whisper (custom)        | 10300/10301 | Speech-to-text (Faster Whisper)|
 | **tts**                 | agent-cli-tts (custom)            | 10200/10201 | Text-to-speech (Kokoro/Piper)  |
-| **diarization**        | agent-cli-diarization (local build) | 61337       | Speaker timing and labeled transcripts (optional) |
+| **diarization**        | agent-cli-diarization               | 61337       | Speaker timing and labeled transcripts (optional) |
 | **transcribe-proxy**    | agent-cli-transcribe-proxy        | 61337       | ASR proxy for iOS/external apps|
 | **rag-proxy**           | agent-cli-rag-proxy               | 8000        | Document-aware chat (RAG)      |
 | **memory-proxy**        | agent-cli-memory-proxy            | 8100        | Long-term memory chat          |

{agent_cli-0.104.0 → agent_cli-0.105.1}/macos/appcast.xml

@@ -4,6 +4,41 @@
         <title>Agent CLI Updates</title>
         <link>https://github.com/basnijholt/agent-cli</link>
         <description>Signed Agent CLI macOS app updates.</description>
+        <item>
+            <title>0.105.0</title>
+            <pubDate>Tue, 08 Sep 2026 10:29:52 +0000</pubDate>
+            <link>https://github.com/basnijholt/agent-cli</link>
+            <sparkle:version>204</sparkle:version>
+            <sparkle:shortVersionString>0.105.0</sparkle:shortVersionString>
+            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
+            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
+            <description sparkle:format="markdown"><![CDATA[## What’s Changed
+
+* Prevent Qwen long-audio truncation (#641) @basnijholt
+* Publish diarization Docker image (#640) @basnijholt
+
+]]></description>
+            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.105.0/AgentCLI.zip" length="18454913" type="application/octet-stream" sparkle:edSignature="Z3253GuNAV6GC2V4Qib0Bm7m8bSdlWHPiXeHIZNa06qAJaY+qRXzrSPRyT8iZW5GnkZOJHciIbqvkkoJ9/DoCw=="/>
+        </item>
+        <item>
+            <title>0.104.0</title>
+            <pubDate>Mon, 07 Sep 2026 21:38:19 +0000</pubDate>
+            <link>https://github.com/basnijholt/agent-cli</link>
+            <sparkle:version>203</sparkle:version>
+            <sparkle:shortVersionString>0.104.0</sparkle:shortVersionString>
+            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
+            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
+            <description sparkle:format="markdown"><![CDATA[## What’s Changed
+
+* Add speaker diarization API and Docker image (#639) @basnijholt
+* Fix Docker image publishing and runtime extras (#638) @basnijholt
+* ⬆️ Update actions/deploy-pages action to v5.0.1 (#637) [@renovate[bot]](https://github.com/apps/renovate)
+* ⬆️ Update docker/setup-buildx-action action to v4.3.0 (#635) [@renovate[bot]](https://github.com/apps/renovate)
+
+
+]]></description>
+            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.104.0/AgentCLI.zip" length="18454336" type="application/octet-stream" sparkle:edSignature="p7BVeR6YS/LBGO61o0xR8GE9beHf+4uaz0c28HrHX4oyWjVZSBgwI6jIjR2Mn23HNLS65CfECbp2NSJoIc+ABw=="/>
+        </item>
         <item>
             <title>0.103.0</title>
             <pubDate>Tue, 01 Sep 2026 19:34:50 +0000</pubDate>
@@ -174,36 +209,5 @@
 ]]></description>
             <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.100.0/AgentCLI.zip" length="22908645" type="application/octet-stream" sparkle:edSignature="2xn/3yQOqyqwCQ5871EAgFra/D/z6UrDl1B+ZxUh7PnOxJpL4Oe44UzQgjFB+LJXzqrY6Kmqvu9GBmk+3aZ1Ag=="/>
         </item>
-        <item>
-            <title>0.99.0</title>
-            <pubDate>Fri, 05 Jun 2026 22:57:28 +0000</pubDate>
-            <link>https://github.com/basnijholt/agent-cli</link>
-            <sparkle:version>194</sparkle:version>
-            <sparkle:shortVersionString>0.99.0</sparkle:shortVersionString>
-            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
-            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
-            <description sparkle:format="markdown"><![CDATA[## What’s Changed
-
-* Add live transcription preview (#571) @basnijholt
-
-]]></description>
-            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.99.0/AgentCLI.zip" length="22924494" type="application/octet-stream" sparkle:edSignature="COI8nqDd0cBtOO9jw/QRmsucsxE0SPiSgNF/o+s/nv3FCjSXy0C+gZX/bIFtCYyz/DYkff+IaHaLzhw8Z5aEBQ=="/>
-        </item>
-        <item>
-            <title>0.98.5</title>
-            <pubDate>Fri, 05 Jun 2026 20:33:00 +0000</pubDate>
-            <link>https://github.com/basnijholt/agent-cli</link>
-            <sparkle:version>193</sparkle:version>
-            <sparkle:shortVersionString>0.98.5</sparkle:shortVersionString>
-            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
-            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
-            <description sparkle:format="markdown"><![CDATA[## What’s Changed
-
-* Fix macOS F1 shortcut recorder crash (#584) @basnijholt
-* Retry Fn hotkeys after Accessibility approval (#583) @basnijholt
-
-]]></description>
-            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.98.5/AgentCLI.zip" length="22904207" type="application/octet-stream" sparkle:edSignature="ZUCHwTaFNGvFW3w/cq1Wf5OzB0e/1mjUQjCksmNaDJy9muTY1r5dRcDuIk15hp7iB4ELNAUUDqqa+PMi/tyiAA=="/>
-        </item>
     </channel>
 </rss>

{agent_cli-0.104.0 → agent_cli-0.105.1}/tests/test_api_integration.py

@@ -129,9 +129,47 @@ def test_server_whisper_command_in_cli() -> None:
     assert "--model" in clean_output
     assert "--ttl" in clean_output
     assert "--wyoming-port" in clean_output
+    assert "--max-new-tokens" in clean_output
     assert "nemo" in clean_output
 
 
+@patch("uvicorn.run")
+def test_server_whisper_propagates_max_new_tokens(mock_uvicorn_run: MagicMock) -> None:
+    """Whisper server should propagate the generation limit to model config."""
+    runner = CliRunner()
+    registry = MagicMock()
+    registry.default_model = "Qwen/Qwen3-ASR-1.7B-hf"
+
+    with (
+        patch("agent_cli.core.deps._check_and_install_extras", return_value=[]),
+        patch("agent_cli.server.cli._check_whisper_deps"),
+        patch("agent_cli.server.cli._check_transformers_audio_model_deps"),
+        patch(
+            "agent_cli.server.whisper.model_registry.create_whisper_registry",
+            return_value=registry,
+        ),
+    ):
+        result = runner.invoke(
+            cli_app,
+            [
+                "server",
+                "whisper",
+                "--model",
+                "Qwen/Qwen3-ASR-1.7B-hf",
+                "--backend",
+                "transformers",
+                "--no-wyoming",
+                "--max-new-tokens",
+                "2048",
+            ],
+        )
+
+    assert result.exit_code == 0
+    registered_config = registry.register.call_args.args[0]
+    assert registered_config.max_new_tokens == 2048
+    mock_uvicorn_run.assert_called_once()
+
+
 @patch("uvicorn.run")
 def test_server_whisper_accepts_asr_wyoming_port_alias(mock_uvicorn_run: MagicMock) -> None:
     """Whisper server should accept the client-style ASR Wyoming port alias."""

{agent_cli-0.104.0 → agent_cli-0.105.1}/tests/test_server_whisper.py

@@ -52,6 +52,7 @@ class TestModelConfig:
         assert config.backend_type == "auto"
         assert config.default_language is None
         assert config.trust_remote_code is False
+        assert config.max_new_tokens == 4096
 
     def test_custom_values(self) -> None:
         """Test custom configuration values."""
@@ -64,6 +65,7 @@ class TestModelConfig:
             cpu_threads=8,
             default_language="en",
             trust_remote_code=True,
+            max_new_tokens=2048,
         )
         assert config.model_name == "small"
         assert config.device == "cuda:0"
@@ -73,6 +75,7 @@ class TestModelConfig:
         assert config.cpu_threads == 8
         assert config.default_language == "en"
         assert config.trust_remote_code is True
+        assert config.max_new_tokens == 2048
 
     def test_zero_ttl_disables_auto_unload(self) -> None:
         """TTL 0 should be accepted as keep-loaded mode."""