feat: model capabilities lined up in tui menu

2 files changed, 63 insertions, 20 deletions

diff --git a/app/converter/clients/__init__.py b/app/converter/clients/__init__.py
index fd3df64..68b988d 100644
--- a/app/converter/clients/__init__.py
+++ b/app/converter/clients/__init__.py
@@ -44,6 +44,7 @@ from .audiocpp import (
     AUDIOCPP_VOICE_SPEAKER,
     AudioCppFamilyProfile,
     AudioCppTTSClient,
+    audiocpp_entry_supports_design,
     audiocpp_entry_voice_capability,
     audiocpp_family_spec_tasks,
     audiocpp_family_voice_policy,
@@ -77,6 +78,7 @@ __all__ = [
     "AUDIOCPP_VOICE_OPTIONAL", "AUDIOCPP_VOICE_NONE",
     "AudioCppFamilyProfile", "AUDIOCPP_DEFAULT_FAMILY_PROFILE",
     "AUDIOCPP_FAMILY_PROFILES", "audiocpp_entry_voice_capability",
+    "audiocpp_entry_supports_design",
     "audiocpp_family_spec_tasks", "audiocpp_family_voice_policy",
     "audiocpp_request_error",
 ]
diff --git a/app/converter/clients/audiocpp.py b/app/converter/clients/audiocpp.py
index c441047..a1888bc 100644
--- a/app/converter/clients/audiocpp.py
+++ b/app/converter/clients/audiocpp.py
@@ -102,24 +102,23 @@ AUDIOCPP_VOICE_REQUIRED = "required"  # clone-only: a reference voice is mandato
 AUDIOCPP_VOICE_OPTIONAL = "optional"  # tts + clone: blank voice means plain TTS
 AUDIOCPP_VOICE_NONE = "none"          # pure TTS: no cloning, no voice at all
 
-# Spec-task cache for audiocpp_family_voice_policy (family -> tasks or
-# None for unknown). The form consults the policy on every menu render,
-# so each family's spec is read at most once per process.
-_FAMILY_SPEC_TASKS: Dict[str, Optional[Set[str]]] = {}
+# Spec cache (family -> parsed spec dict, or None for unknown). The form
+# consults the policy and capability tags on every menu render, so each
+# family's spec is read at most once per process.
+_FAMILY_SPECS: Dict[str, Optional[dict]] = {}
 
 
-def audiocpp_family_spec_tasks(family: str) -> Optional[Set[str]]:
-    """FAMILY's task set from the local audio.cpp checkout's model_specs.
+def _family_spec(family: str) -> Optional[dict]:
+    """FAMILY's parsed model spec from the local audio.cpp checkout.
 
     Reads ``<checkout>/model_specs/<family>.json`` (the checkout the setup
-    wizard manages, which also ships the specs for remote servers) and
-    returns its "tasks" list as a set, or None when the checkout is
-    missing, the family is not described, or the spec is unparsable.
-    Results are cached per process.
+    wizard manages, which also ships the specs for remote servers), or
+    None when the checkout is missing, the family is not described, or the
+    spec is unparsable. Results are cached per process.
     """
-    if family in _FAMILY_SPEC_TASKS:
-        return _FAMILY_SPEC_TASKS[family]
-    tasks: Optional[Set[str]] = None
+    if family in _FAMILY_SPECS:
+        return _FAMILY_SPECS[family]
+    spec: Optional[dict] = None
     try:
         # Imported lazily: backends.audiocpp imports this package (its
         # voices module), so a module-level import would cycle.
@@ -129,14 +128,56 @@ def audiocpp_family_spec_tasks(family: str) -> Optional[Set[str]]:
         checkout = None
     if checkout is not None:
         try:
-            spec = json.loads((checkout / "model_specs" / f"{family}.json")
-                              .read_text(encoding="utf-8"))
+            parsed = json.loads((checkout / "model_specs" / f"{family}.json")
+                                .read_text(encoding="utf-8"))
         except (OSError, ValueError):
-            spec = None
-        if isinstance(spec, dict) and isinstance(spec.get("tasks"), list):
-            tasks = {str(task) for task in spec["tasks"]}
-    _FAMILY_SPEC_TASKS[family] = tasks
-    return tasks
+            parsed = None
+        if isinstance(parsed, dict):
+            spec = parsed
+    _FAMILY_SPECS[family] = spec
+    return spec
+
+
+def audiocpp_family_spec_tasks(family: str) -> Optional[Set[str]]:
+    """FAMILY's task set from the local audio.cpp checkout's model_specs.
+
+    Returns the spec's "tasks" list as a set, or None when the family is
+    not described (see _family_spec).
+    """
+    spec = _family_spec(family)
+    if spec is None or not isinstance(spec.get("tasks"), list):
+        return None
+    return {str(task) for task in spec["tasks"]}
+
+
+def audiocpp_entry_supports_design(family: str, task: str,
+                                   model_id: str) -> bool:
+    """Whether a server model entry can design a voice from a description.
+
+    True for task-"vdes" entries (the model *is* a voice-design model) and
+    for entries of families whose spec advertises a design task — those
+    families design on the regular entry from the request's instructions
+    text (e.g. OmniVoice, VoxCPM2). Qwen3-TTS is the exception: its
+    design support lives only in a separate VoiceDesign model entry (also
+    task "vdes"), while its Base/CustomVoice entries cannot design. Model
+    IDs play no role today but stay in the signature for parity with
+    audiocpp_entry_voice_capability. Unknown families (no local specs)
+    conservatively report no design support.
+    """
+    if task == AUDIOCPP_TASK_VDES:
+        return True
+    if family == AUDIOCPP_FAMILY_QWEN3_TTS:
+        return False
+    spec = _family_spec(family)
+    if spec is None:
+        return False
+    design_markers = {"design", AUDIOCPP_TASK_VDES}
+    tasks = audiocpp_family_spec_tasks(family)
+    if tasks and tasks & design_markers:
+        return True
+    capabilities = spec.get("capabilities")
+    return isinstance(capabilities, dict) \
+        and bool(set(map(str, capabilities)) & design_markers)
 
 
 def audiocpp_family_voice_policy(family: str) -> str: