Fix strict YAML boolean parsing

agent-mcp-servers/vlm-mcp/vlm_mcp_server/__main__.py

@@ -70,7 +70,7 @@
     load_models_config_from_dict,
     make_vlm,
 )
-from xr_ai_models.config import KIND_OPENAI_COMPAT
+from xr_ai_models.config import KIND_OPENAI_COMPAT, parse_bool
 from xr_ai_models.protocols import VLMService
 
 
@@ -89,7 +89,7 @@ def _make_vlm_from_cfg(cfg: dict[str, Any]) -> tuple[VLMService, float]:
     models_block: dict[str, Any] | None = cfg.get("models")
     vlm_server: str | None = cfg.get("vlm_server")
     vlm_request_timeout_s = float(cfg.get("vlm_request_timeout_s", 60.0))
-    enable_thinking = bool(cfg.get("enable_thinking", False))
+    enable_thinking = parse_bool(cfg.get("enable_thinking", False), "enable_thinking")
 
     if models_block:
         vlm_entry = dict(models_block.get("vlm") or {})

agent-sdk/xr-ai-models/xr_ai_models/config.py

@@ -25,6 +25,25 @@
 
 KIND_OPENAI_COMPAT: ModelKind = "openai_compat"
 
+_TRUE_BOOL_STRINGS = {"1", "true", "yes", "on"}
+_FALSE_BOOL_STRINGS = {"0", "false", "no", "off"}
+
+
+def parse_bool(value: object, key: str) -> bool:
+    """Parse a config boolean without treating every non-empty string as true."""
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in _TRUE_BOOL_STRINGS:
+            return True
+        if normalized in _FALSE_BOOL_STRINGS:
+            return False
+    raise ValueError(
+        f"{key} must be a boolean or one of "
+        f"{sorted(_TRUE_BOOL_STRINGS | _FALSE_BOOL_STRINGS)} (got {value!r})"
+    )
+
 
 @dataclass(frozen=True)
 class LLMSpec:
@@ -171,7 +190,7 @@ def _construct(category: Category, body: dict[str, Any]) -> Spec:
         "base_url": _require_str(body, "base_url"),
         # Remote endpoints (e.g. hosted NIM) have no local /health route; set
         # ``health_check: false`` so the worker readiness gate doesn't block.
-        "health_check": bool(body.get("health_check", True)),
+        "health_check": parse_bool(body.get("health_check", True), "health_check"),
     }
     if "api_key_env" in body:
         common["api_key_env"] = body["api_key_env"]

ai-services/llm/llama_nemotron/llama_nemotron_llm_server/__main__.py

@@ -38,6 +38,7 @@
 from xr_ai_vllm import (
     DEFAULT_IMAGE,
     load_config,
+    parse_bool,
     resolve_model_cache,
     serve,
     setup_hf_env,
@@ -74,9 +75,12 @@ def run() -> None:
     tp_size            = int(cfg.get("tensor_parallel_size", _DEFAULT_TP))
     max_ctx            = int(cfg.get("max_model_len",    _DEFAULT_CTX))
     gpu_mem            = float(cfg.get("gpu_memory_utilization", _DEFAULT_GPU_MEM))
-    enforce_eager      = bool(cfg.get("enforce_eager",   _DEFAULT_EAGER))
+    enforce_eager      = parse_bool(cfg.get("enforce_eager", _DEFAULT_EAGER), "enforce_eager")
     tool_call_parser   = cfg.get("tool_call_parser",     _DEFAULT_TOOL_CALL_PARSER)
-    enable_tool_choice = bool(cfg.get("enable_tool_choice", _DEFAULT_ENABLE_TOOL_CHOICE))
+    enable_tool_choice = parse_bool(
+        cfg.get("enable_tool_choice", _DEFAULT_ENABLE_TOOL_CHOICE),
+        "enable_tool_choice",
+    )
     backend            = cfg.get("vllm_backend",         "pip")
     image              = cfg.get("vllm_image",           DEFAULT_IMAGE)
 

ai-services/llm/nemotron3_nano/nemotron3_nano_llm_server/__main__.py

@@ -40,6 +40,7 @@
     DEFAULT_IMAGE,
     gpu_compute_major,
     load_config,
+    parse_bool,
     resolve_model_cache,
     serve,
     setup_hf_env,
@@ -100,7 +101,7 @@ def run() -> None:
     tp_size       = int(cfg.get("tensor_parallel_size", _DEFAULT_TP))
     max_ctx       = int(cfg.get("max_model_len",   _DEFAULT_CTX))
     gpu_mem       = float(cfg.get("gpu_memory_utilization", _DEFAULT_GPU_MEM))
-    enforce_eager = bool(cfg.get("enforce_eager",  _DEFAULT_EAGER))
+    enforce_eager = parse_bool(cfg.get("enforce_eager", _DEFAULT_EAGER), "enforce_eager")
     parser_url    = cfg.get("parser_url",          _PARSER_URL_DEFAULT)
     backend       = cfg.get("vllm_backend",        "pip")
     image         = cfg.get("vllm_image",          DEFAULT_IMAGE)

ai-services/llm/nemotron_omni/nemotron_omni_llm_server/__main__.py

@@ -47,6 +47,7 @@
     DEFAULT_IMAGE,
     gpu_compute_major,
     load_config,
+    parse_bool,
     resolve_model_cache,
     serve,
     setup_hf_env,
@@ -81,7 +82,8 @@ def run() -> None:
     # nvidia-smi queries the right device.
     cuda_devices = setup_hf_env(cfg, model_cache)
 
-    if cfg.get("use_bf16", False):
+    use_bf16 = parse_bool(cfg.get("use_bf16", False), "use_bf16")
+    if use_bf16:
         model = cfg.get("model_bf16", _MODEL_BF16)
         use_kv_fp8 = False
         logger.info("use_bf16=true → {}", model)
@@ -104,7 +106,7 @@ def run() -> None:
     tp_size       = int(cfg.get("tensor_parallel_size", _DEFAULT_TP))
     max_ctx       = int(cfg.get("max_model_len",    _DEFAULT_CTX))
     gpu_mem       = float(cfg.get("gpu_memory_utilization", _DEFAULT_GPU_MEM))
-    enforce_eager = bool(cfg.get("enforce_eager",   _DEFAULT_EAGER))
+    enforce_eager = parse_bool(cfg.get("enforce_eager", _DEFAULT_EAGER), "enforce_eager")
     prune_rate    = float(cfg.get("video_pruning_rate", _DEFAULT_PRUNE))
     video_fps     = int(cfg.get("video_fps",        _DEFAULT_FPS))
     video_frames  = int(cfg.get("video_num_frames", _DEFAULT_FRAMES))

ai-services/tts/piper/piper_tts_server/__main__.py

@@ -47,6 +47,24 @@
 # Callers/tests can treat this as retry-or-skip rather than a hard failure.
 _EXIT_VOICE_UNAVAILABLE = 3
 
+_TRUE_BOOL_STRINGS = {"1", "true", "yes", "on"}
+_FALSE_BOOL_STRINGS = {"0", "false", "no", "off"}
+
+
+def _parse_bool(value: object, key: str) -> bool:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in _TRUE_BOOL_STRINGS:
+            return True
+        if normalized in _FALSE_BOOL_STRINGS:
+            return False
+    raise ValueError(
+        f"{key} must be a boolean or one of "
+        f"{sorted(_TRUE_BOOL_STRINGS | _FALSE_BOOL_STRINGS)} (got {value!r})"
+    )
+
 
 def _resolve_model_cache(cfg: dict, yaml_dir: Path) -> Path:
     raw = cfg.get("model_cache", "../models")
@@ -202,7 +220,7 @@ def _build_app(cfg: dict, model_cache: Path):
     from pydantic import BaseModel
 
     voice_name = cfg["voice"]
-    use_cuda   = bool(cfg.get("use_cuda", False))
+    use_cuda   = _parse_bool(cfg.get("use_cuda", False), "use_cuda")
     backend    = _PiperBackend(voice_name, model_cache, use_cuda)
 
     app = FastAPI(title="Piper TTS Server", version="0.1.0")

ai-services/vlm-server/vlm_server/__main__.py

@@ -45,6 +45,7 @@
 from xr_ai_vllm import (
     DEFAULT_IMAGE,
     load_config,
+    parse_bool,
     resolve_model_cache,
     serve,
     setup_hf_env,
@@ -81,7 +82,7 @@ def run() -> None:
     tp_size       = int(cfg.get("tensor_parallel_size", _DEFAULT_TP))
     max_ctx       = int(cfg.get("max_model_len",    _DEFAULT_CTX))
     gpu_mem       = float(cfg.get("gpu_memory_utilization", _DEFAULT_GPU_MEM))
-    enforce_eager = bool(cfg.get("enforce_eager",   _DEFAULT_EAGER))
+    enforce_eager = parse_bool(cfg.get("enforce_eager", _DEFAULT_EAGER), "enforce_eager")
     max_images    = int(cfg.get("max_images_per_prompt", _DEFAULT_MAX_IMAGES))
     max_videos    = int(cfg.get("max_videos_per_prompt", _DEFAULT_MAX_VIDEOS))
     backend       = cfg.get("vllm_backend",         "pip")

tests/test_models_config.py

@@ -128,6 +128,30 @@ def test_health_check_defaults_true_and_parses_false(tmp_path) -> None:
     assert nim.base_url == "https://integrate.api.nvidia.com"
 
 
+def test_health_check_parses_quoted_false(tmp_path) -> None:
+    cfg = load_models_config(_write(tmp_path, """
+remote_llm:
+  kind:        openai_compat
+  category:    llm
+  base_url:    https://integrate.api.nvidia.com
+  model_name:  meta/llama-3.1-8b-instruct
+  health_check: "false"
+"""))
+    assert cfg.llm("remote_llm").health_check is False
+
+
+def test_health_check_rejects_unknown_string(tmp_path) -> None:
+    with pytest.raises(ValueError, match="health_check"):
+        load_models_config(_write(tmp_path, """
+remote_llm:
+  kind:        openai_compat
+  category:    llm
+  base_url:    https://integrate.api.nvidia.com
+  model_name:  meta/llama-3.1-8b-instruct
+  health_check: maybe
+"""))
+
+
 def test_vlm_preset(tmp_path) -> None:
     cfg = load_models_config(_write(tmp_path, """
 vlm:

tests/test_piper_tts.py

@@ -20,6 +20,7 @@
 from __future__ import annotations
 
 import asyncio
+import importlib.util
 import json
 import os
 import shutil
@@ -50,6 +51,27 @@
 _EXIT_VOICE_UNAVAILABLE = 3
 
 
+def _load_piper_main_module():
+    spec = importlib.util.spec_from_file_location(
+        "piper_tts_server_main",
+        _PIPER_PROJECT / "piper_tts_server" / "__main__.py",
+    )
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+async def test_piper_config_parses_quoted_use_cuda_false() -> None:
+    mod = _load_piper_main_module()
+    assert mod._parse_bool("false", "use_cuda") is False
+
+
+async def test_piper_config_rejects_unknown_use_cuda_string() -> None:
+    mod = _load_piper_main_module()
+    with pytest.raises(ValueError, match="use_cuda"):
+        mod._parse_bool("sometimes", "use_cuda")
+
+
 class _ServerExited(Exception):
     """Raised when piper_tts_server exits before binding its port.
 

tests/test_vllm_lifecycle.py

@@ -1,16 +1,41 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Unit tests for xr_ai_vllm._lifecycle pure helpers."""
+"""Unit tests for xr_ai_vllm pure helpers."""
 from __future__ import annotations
 
 from unittest.mock import MagicMock, patch
 
 import pytest
 
+from xr_ai_vllm import parse_bool
 from xr_ai_vllm._lifecycle import health_ok, health_url, wait_until_healthy
 
 
+class TestParseBool:
+    @pytest.mark.parametrize(
+        ("value", "expected"),
+        [
+            (True, True),
+            (False, False),
+            ("true", True),
+            ("YES", True),
+            ("1", True),
+            ("on", True),
+            ("false", False),
+            ("No", False),
+            ("0", False),
+            (" off ", False),
+        ],
+    )
+    def test_accepts_booleans_and_known_strings(self, value, expected):
+        assert parse_bool(value, "enforce_eager") is expected
+
+    def test_rejects_unknown_string(self):
+        with pytest.raises(ValueError, match="enforce_eager"):
+            parse_bool("sometimes", "enforce_eager")
+
+
 class TestHealthUrl:
     def test_always_probes_localhost(self):
         # The host param is intentionally ignored — always 127.0.0.1.

tests/test_vlm_mcp.py

@@ -282,6 +282,41 @@ async def test_make_vlm_from_cfg_new_models_block(mock_vlm, png_path: Path):
     assert payload["chat_template_kwargs"] == {"enable_thinking": False}
 
 
+async def test_make_vlm_from_cfg_parses_quoted_enable_thinking_false(mock_vlm, png_path: Path):
+    server, base_url = mock_vlm
+    server.answer = "quoted false"
+
+    vlm, _ = _make_vlm_from_cfg({
+        "models": {
+            "vlm": {
+                "kind":     "preset:cosmos_vlm",
+                "base_url": base_url,
+            },
+        },
+        "enable_thinking": "false",
+    })
+    mcp = build_mcp(vlm)
+    try:
+        result = await mcp.call_tool(
+            "ask_image",
+            {"question": "cosmos q", "image_path": str(png_path)},
+        )
+    finally:
+        await vlm.close()
+
+    assert result.structured_content["result"] == "quoted false"
+    payload = server.requests[0]
+    assert payload["chat_template_kwargs"] == {"enable_thinking": False}
+
+
+async def test_make_vlm_from_cfg_rejects_unknown_enable_thinking_string():
+    with pytest.raises(ValueError, match="enable_thinking"):
+        _make_vlm_from_cfg({
+            "vlm_server":      "http://localhost:8100",
+            "enable_thinking": "sometimes",
+        })
+
+
 async def test_make_vlm_from_cfg_missing_required_keys_raises():
     """Neither models: nor vlm_server: present → ValueError."""
     with pytest.raises(ValueError, match="must specify either"):

tests/test_xr_ai_voicegate.py

@@ -799,6 +799,21 @@ def test_load_voice_gate_config_null_phrases_normalizes_to_empty(tmp_path: pathl
     assert cfg.followup_grace_s == 3.0
 
 
+def test_load_voice_gate_config_parses_quoted_listening_chime_false(tmp_path: pathlib.Path):
+    p = tmp_path / "voice_gate.yaml"
+    p.write_text('magic_phrases: agent\nlistening_chime: "false"\n')
+    cfg = load_voice_gate_config(p)
+    assert cfg.magic_phrases == ("agent",)
+    assert cfg.listening_chime is False
+
+
+def test_load_voice_gate_config_rejects_unknown_listening_chime_string(tmp_path: pathlib.Path):
+    p = tmp_path / "voice_gate.yaml"
+    p.write_text("listening_chime: sometimes\n")
+    with pytest.raises(ValueError, match="listening_chime"):
+        load_voice_gate_config(p)
+
+
 def test_load_voice_gate_config_strips_whitespace_and_drops_empty(tmp_path: pathlib.Path):
     """Case 39: phrase entries are stripped, and empty entries (after
     strip) are dropped — same normalization the inline parser ran."""

utils/xr-ai-vllm/xr_ai_vllm/__init__.py

@@ -50,6 +50,7 @@
 from ._config import (
     gpu_compute_major,
     load_config,
+    parse_bool,
     resolve_model_cache,
     setup_hf_env,
 )
@@ -237,6 +238,7 @@ def stop_persistent_servers(
     "DEFAULT_IMAGE",
     "resolve_model_cache",
     "load_config",
+    "parse_bool",
     "setup_hf_env",
     "gpu_compute_major",
 ]

utils/xr-ai-vllm/xr_ai_vllm/_config.py

@@ -19,6 +19,25 @@
 
 log = logging.getLogger(__name__)
 
+_TRUE_BOOL_STRINGS = {"1", "true", "yes", "on"}
+_FALSE_BOOL_STRINGS = {"0", "false", "no", "off"}
+
+
+def parse_bool(value: object, key: str) -> bool:
+    """Parse a config boolean without treating every non-empty string as true."""
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in _TRUE_BOOL_STRINGS:
+            return True
+        if normalized in _FALSE_BOOL_STRINGS:
+            return False
+    raise ValueError(
+        f"{key} must be a boolean or one of "
+        f"{sorted(_TRUE_BOOL_STRINGS | _FALSE_BOOL_STRINGS)} (got {value!r})"
+    )
+
 
 def resolve_model_cache(cfg: dict, yaml_dir: Path, *, default: str) -> Path:
     """Resolve ``model_cache`` (relative to the YAML dir) and ensure it exists."""