feat: generate a book with all installed models to compare

2 files changed, 293 insertions, 29 deletions

diff --git a/app/ui/hub.py b/app/ui/hub.py
index de5701f..9bd136a 100644
--- a/app/ui/hub.py
+++ b/app/ui/hub.py
@@ -74,6 +74,11 @@ from ui import runview, taskview, tui
 
 _CANCEL = object()  # sentinel: a convert preflight confirm backed out
 
+# The Generate form's audio.cpp Model pick for "All (multiple generation)":
+# one conversion per configured model (model-major), with model-tagged
+# output names. A sentinel string, distinct from every real model id.
+AUDIOCPP_MODEL_ALL = "__all__"
+
 
 class _BackToForm(Exception):
     """Raised when Esc backs out of a preflight confirm (re-show the form)."""
@@ -941,8 +946,13 @@ def _preflight(stdscr, cmd: tuple) -> bool:
     kwargs (``book_files``/``planned``) for ``audiobook.convert``. Returns
     False when nothing would be converted (a flash explains why), so the
     user stays in the menu instead of entering an empty run.
+
+    An "All (multiple generation)" run (``model_ids`` in the kwargs) is
+    planned by _preflight_all instead: one plan per model.
     """
     _kind, backend, kwargs = cmd
+    if kwargs.get("model_ids"):
+        return _preflight_all(stdscr, backend, kwargs)
     voice_mode = voice_mode_for(backend, kwargs.get("voice"),
                                 kwargs.get("clone"),
                                 kwargs.get("instructions"))
@@ -975,6 +985,59 @@ def _preflight(stdscr, cmd: tuple) -> bool:
     return True
 
 
+def _preflight_all(stdscr, backend: str, kwargs: dict) -> bool:
+    """Run the overwrite checks for an "All (multiple generation)" run.
+
+    Plans one conversion per model: every model's output names carry its
+    model tag and its own adapted voice (the mapper's ``model_voices``),
+    so each model's overwrites are asked — and accepted — separately, all
+    up front (a cancel returns to the form). Records the union book list
+    as ``book_files`` and the per-model plans as ``planned_by_model`` on
+    the command kwargs for ``audiobook.convert``. Returns False when
+    nothing would be converted (a flash explains why), so the user stays
+    in the menu instead of entering an empty run.
+    """
+    model_ids = kwargs.get("model_ids") or []
+    model_voices = kwargs.get("model_voices") or {}
+    instructions = kwargs.get("instructions")
+
+    def confirm(message: str, default: bool) -> bool:
+        answer = tui.confirm(stdscr, message, default=default,
+                             cancel_value=_CANCEL)
+        if answer is _CANCEL:
+            raise _BackToForm()
+        return answer
+
+    book_files: list = []
+    planned_by_model: dict = {}
+    with contextlib.redirect_stdout(io.StringIO()):
+        for model_id in model_ids:
+            voice = model_voices.get(model_id)
+            voice_mode = voice_mode_for(backend, voice,
+                                        kwargs.get("clone"), instructions)
+            books, planned = AudiobookConverter.preflight_overwrites(
+                backend=backend, voice=voice, voice_mode=voice_mode,
+                voice_clone_ref_audio=kwargs.get("clone"),
+                output_format=kwargs.get("output_format")
+                or config.AUDIO_FORMAT,
+                instructions=instructions, confirm=confirm,
+                name_tag=AudiobookConverter.compute_model_tag(model_id))
+            if not book_files:
+                book_files = books
+            planned_by_model[model_id] = planned
+    if not book_files:
+        tui.flash(stdscr, "No books to convert. Add a .txt, .pdf or .epub "
+                  "file to the input folder first.")
+        return False
+    if not any(planned_by_model.values()):
+        tui.flash(stdscr, "Nothing to convert — every existing output was "
+                  "kept.")
+        return False
+    kwargs["book_files"] = book_files
+    kwargs["planned_by_model"] = planned_by_model
+    return True
+
+
 def _gate_backend(field: dict, key: str) -> Callable:
     """A visible() that shows FIELD only when the Backend field is KEY.
 
@@ -1055,6 +1118,18 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
     the field keys ("" for the managed entry) so two entries of this
     backend can share one form without overwriting each other.
 
+    With more than one model configured, the Model menu closes with
+    "All (multiple generation)" (AUDIOCPP_MODEL_ALL): the run then
+    generates every book once per model, with model-tagged output names.
+    The single Voice pick is sent to every model that accepts it (the
+    same server-side clone voice, or a built-in speaker name on
+    CustomVoice entries); models the pick cannot serve fall back to
+    their own default (first speaker / first server voice / no voice —
+    design entries take the Instructions text), and Generate! refuses
+    the combinations that cannot work (a design model without
+    Instructions; a clone-only model without server voices or an
+    instruction-defined voice). See the "All" helpers below.
+
     The Voice field tracks the selected entry's capability — built-in
     speakers on CustomVoice, the server's clone voices on every other
     entry. A clone-capable entry whose server lists no voices cannot be
@@ -1168,20 +1243,136 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         return next((m for m in models if m.get("id") == model_id),
                     models[0])
 
-    def model_capability(fields) -> str:
-        entry = model_entry(fields)
+    def entry_capability(entry: dict) -> str:
+        """How ENTRY's voice is supplied (speaker/clone/design)."""
         return audiocpp_entry_voice_capability(
             entry.get("family") or "", entry.get("task") or "tts",
             entry.get("id") or "")
 
+    def all_selected(fields) -> bool:
+        """True when the Model pick is "All (multiple generation)"."""
+        return _field_value(fields, prefix + "model_id") == AUDIOCPP_MODEL_ALL
+
+    def model_capability(fields) -> str:
+        return entry_capability(model_entry(fields))
+
     def model_voice_policy(fields) -> str:
         """The selected entry's family voice policy (required/optional/none)."""
         return audiocpp_family_voice_policy(
             model_entry(fields).get("family") or "")
 
+    # -- "All (multiple generation)" ------------------------------------
+    # One conversion per configured model: the single Voice pick is used
+    # by every model that accepts it — the same server-side clone voice
+    # flows into every clone-capable family, a built-in speaker name into
+    # every CustomVoice entry — and each remaining model falls back to
+    # its own sensible default (first built-in speaker, first server
+    # voice, or no voice at all: design entries and pure-TTS families
+    # take no voice, the client then designs from Instructions or
+    # synthesizes plainly).
+
+    def any_voice_model() -> bool:
+        """True when at least one configured model takes a voice pick."""
+        return any(
+            entry_capability(m) == AUDIOCPP_VOICE_SPEAKER
+            or (entry_capability(m) == AUDIOCPP_VOICE_CLONE
+                and audiocpp_family_voice_policy(
+                    m.get("family") or "") != AUDIOCPP_VOICE_NONE)
+            for m in models)
+
+    def any_design_model() -> bool:
+        """True when at least one configured model designs its voice."""
+        return any(entry_capability(m) == AUDIOCPP_VOICE_DESIGN
+                   for m in models)
+
+    def all_voice_union() -> list:
+        """Every voice an "All" run can offer.
+
+        Each clone-capable family's server voices first (shared by all of
+        them on the managed entry; per-model on a remote one), then the
+        built-in speakers when a speaker-capable model is configured.
+        Duplicates removed, order stable — a clone voice leads the list,
+        matching the pick-falls-back rules.
+        """
+        union = []
+        for m in models:
+            if entry_capability(m) != AUDIOCPP_VOICE_CLONE \
+                    or audiocpp_family_voice_policy(
+                        m.get("family") or "") == AUDIOCPP_VOICE_NONE:
+                continue
+            for voice in voices_for(m.get("id")):
+                if voice not in union:
+                    union.append(voice)
+        if any(entry_capability(m) == AUDIOCPP_VOICE_SPEAKER
+               for m in models):
+            for speaker in QWEN3_TTS_SPEAKERS:
+                if speaker not in union:
+                    union.append(speaker)
+        return union
+
+    def all_voice_for(model_id: str, picked: Optional[str]) -> Optional[str]:
+        """The voice to send for MODEL_ID in an "All" run.
+
+        The picked voice wins wherever the model accepts it; models the
+        pick cannot serve fall back to their own default: the first
+        built-in speaker (CustomVoice) or first server voice (cloning),
+        or no voice at all (design entries and voice-less clone families
+        — the client then designs the voice from Instructions or
+        synthesizes plainly).
+        """
+        entry = next((m for m in models if m.get("id") == model_id),
+                     models[0])
+        capability = entry_capability(entry)
+        if capability == AUDIOCPP_VOICE_DESIGN:
+            return None
+        if capability == AUDIOCPP_VOICE_SPEAKER:
+            if picked and picked in QWEN3_TTS_SPEAKERS:
+                return picked
+            return QWEN3_TTS_SPEAKERS[0]
+        # Clone capability; the family policy decides whether a voice
+        # exists at all.
+        if audiocpp_family_voice_policy(
+                entry.get("family") or "") == AUDIOCPP_VOICE_NONE:
+            return None
+        voices = voices_for(model_id)
+        if picked and picked in voices:
+            return picked
+        return voices[0] if voices else None
+
+    def all_voice_problem() -> Optional[str]:
+        """Why an "All" run cannot start with the current settings, or None.
+
+        Refuses when a voice design model is configured without the
+        Instructions text its voice comes from, and when a clone-only
+        model has neither server voices nor an instruction-defined voice.
+        Every other mismatch is resolved by all_voice_for's per-model
+        fallback instead.
+        """
+        instructions_value = str(_field_value(fields, prefix + "instructions")
+                                 or "").strip()
+        if any_design_model() and not instructions_value:
+            return ("The 'All' run includes a voice design model — "
+                    "describe the voice in Instructions")
+        for m in models:
+            if entry_capability(m) != AUDIOCPP_VOICE_CLONE:
+                continue
+            if audiocpp_family_voice_policy(
+                    m.get("family") or "") != AUDIOCPP_VOICE_REQUIRED:
+                continue
+            if voices_for(m.get("id")) or instructions_value:
+                continue
+            return (f"No voices are available to clone for "
+                    f"'{m.get('id')}' — configure voices on the server, "
+                    "describe one in Instructions, or pick a single model")
+        return None
+
     def reset_voice(fields) -> None:
         """Re-point the Voice field at the newly selected model's voice.
 
+        With "All" picked, the pick survives when the union of every
+        model's voices still offers it; otherwise it falls back to the
+        union's first entry (a server clone voice when one exists).
+
         A model switch that keeps the same voice list (two clone entries
         sharing one server's voices) keeps the current pick: only a value
         the new list cannot offer is re-pointed at its default. Families
@@ -1190,6 +1381,12 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         """
         voice_field = next(f for f in fields
                            if f.get("key") == prefix + "audiocpp_voice")
+        if all_selected(fields):
+            voices = all_voice_union()
+            if voice_field.get("value") in voices:
+                return
+            voice_field["value"] = voices[0] if voices else ""
+            return
         capability = model_capability(fields)
         if capability == AUDIOCPP_VOICE_DESIGN \
                 or model_voice_policy(fields) == AUDIOCPP_VOICE_NONE:
@@ -1213,6 +1410,8 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         voice_field["value"] = voices[0] if voices else ""
 
     def voice_choices(fields) -> list:
+        if all_selected(fields):
+            return [(v, v) for v in all_voice_union()]
         capability = model_capability(fields)
         if capability == AUDIOCPP_VOICE_SPEAKER:
             # Built-in Qwen3-TTS CustomVoice speakers; no server query needed.
@@ -1246,12 +1445,19 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
     def voice_validate(value):
         """Refuse Generate! when this entry's clone voice is unavailable.
 
+        With "All" picked the same check runs across every configured
+        model (design models need Instructions; clone-only models need
+        server voices or an instruction-defined voice) — see
+        all_voice_problem.
+
         A blank Voice is valid on mixed tts+clone families (plain TTS —
         the model's own default voice) and, on any clone-capable entry,
         when an Instructions text substitutes for the voice: on families
         that condition synthesis on instructions alone the client designs
         the voice from it (instruction-voice mode).
         """
+        if all_selected(fields):
+            return all_voice_problem()
         if model_capability(fields) != AUDIOCPP_VOICE_CLONE:
             return None
         has_instruction = bool(str(_field_value(
@@ -1336,10 +1542,18 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
     def entry_supports_options(fs) -> bool:
         """True when the selected entry's family defines request options.
 
-        Resolved strictly from this machine's model_specs: a family the
-        specs prove unable to read options, or cannot classify at all,
-        keeps the field hidden (unknown support is treated as no).
+        With "All" picked: any configured model's family with declared
+        request options shows the field (the server ignores keys a model
+        does not know). Resolved strictly from this machine's model_specs:
+        a family the specs prove unable to read options, or cannot
+        classify at all, keeps the field hidden (unknown support is
+        treated as no).
         """
+        if all_selected(fs):
+            return any(
+                audiocpp_backend.supports_request_options(
+                    option_families, m.get("family") or "") is True
+                for m in models)
         family = model_entry(fs).get("family") or ""
         return audiocpp_backend.supports_request_options(
             option_families, family) is True
@@ -1358,10 +1572,47 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         "Examples: emotion=neutral, speed=1.1, temperature=0.8",
     ]
 
+    def all_voice_visible(fs) -> bool:
+        """Whether the Voice field applies to the current Model pick.
+
+        With "All" picked: shown when at least one configured model takes
+        a voice (the pick feeds every model that accepts it); hidden when
+        every model designs or plainly synthesizes. A single model keeps
+        the per-entry rule: hidden on design entries (the voice is
+        described) and on pure-TTS families (no cloning, no voice).
+        """
+        if all_selected(fs):
+            return any_voice_model()
+        return not (
+            model_capability(fs) == AUDIOCPP_VOICE_DESIGN
+            or (model_capability(fs) == AUDIOCPP_VOICE_CLONE
+                and model_voice_policy(fs) == AUDIOCPP_VOICE_NONE))
+
+    def instructions_validate(value) -> Optional[str]:
+        """Refuse a blank Instructions when the run needs it for a voice.
+
+        Single-model: required on design entries (the voice comes from
+        it). "All": required when any configured model is a design model,
+        even though the text is optional style control for the others.
+        """
+        if all_selected(fields):
+            if any_design_model() and not str(value).strip():
+                return ("The 'All' run includes a voice design model — "
+                        "describe the voice in Instructions")
+            return None
+        if model_capability(fields) != AUDIOCPP_VOICE_DESIGN \
+                or str(value).strip():
+            return None
+        return "Describe the voice, e.g. 'A warm female narrator'"
+
     fields = [
         {"key": prefix + "model_id", "label": "Model", "kind": "choice",
          "value": default_model,
-         "choices": [(_label(m), m.get("id")) for m in models],
+         "choices": [(_label(m), m.get("id")) for m in models]
+         + ([("All (multiple generation)", AUDIOCPP_MODEL_ALL)]
+            # Offered with more than one model configured: a single-model
+            # server has nothing to compare.
+            if len(models) > 1 else []),
          # The pick menu shows the padded capability table; the form row
          # collapses its column padding back to the two-space gutter.
          "compact_label": True,
@@ -1377,10 +1628,7 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
          "kind": "choice",
          "value": initial_voice,
          "choices": lambda fs: voice_choices(fs),
-         "visible": lambda fs: not (
-             model_capability(fs) == AUDIOCPP_VOICE_DESIGN
-             or (model_capability(fs) == AUDIOCPP_VOICE_CLONE
-                 and model_voice_policy(fs) == AUDIOCPP_VOICE_NONE)),
+         "visible": lambda fs: all_voice_visible(fs),
          "on_empty_choices": no_voices_hint,
          "validate": voice_validate},
         # Style/voice-design instruction. Required for design entries (the
@@ -1390,9 +1638,7 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         {"key": prefix + "instructions", "label": "Instructions", "kind": "text",
          "value": "",
          "help": INSTRUCTIONS_HELP,
-         "validate": lambda value: None
-         if (model_capability(fields) != AUDIOCPP_VOICE_DESIGN or str(value).strip())
-         else "Describe the voice, e.g. 'A warm female narrator'"},
+         "validate": instructions_validate},
         # Free-form per-model controls (--option KEY=VALUE on the CLI).
         # Shown only for families whose audio.cpp spec declares request
         # options; unknown-support families keep it hidden.
@@ -1404,10 +1650,6 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
     ]
 
     def mapper(result) -> Optional[tuple]:
-        model_id = result[prefix + "model_id"]
-        # The picked voice (a built-in speaker name on a CustomVoice entry,
-        # a server-side preset otherwise); the client resolves which it is.
-        voice = result[prefix + "audiocpp_voice"] or None
         # The instruction is forwarded for every capability: required on
         # design entries, optional style/delivery control elsewhere. With
         # no voice it defines the voice on instruction-conditioned families.
@@ -1419,7 +1661,6 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
         except ValueError:
             request_options = {}  # submit-time validation already caught this
         kwargs = {
-            "model_id": model_id, "voice": voice,
             "instructions": instructions,
             "request_options": request_options,
             **_common_kwargs(result),
@@ -1432,6 +1673,23 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None,
                 kwargs["audiocpp_rehost"] = True
         else:
             kwargs["api_url"] = api_url
+        if result[prefix + "model_id"] == AUDIOCPP_MODEL_ALL:
+            # "All (multiple generation)": one conversion per configured
+            # model, each with the picked voice where the model accepts
+            # it and its per-model fallback where it does not
+            # (see all_voice_for). audiobook.convert unloads loaded
+            # models between the per-model conversions.
+            picked = result.get(prefix + "audiocpp_voice") or ""
+            kwargs["model_ids"] = [m.get("id") for m in models]
+            kwargs["model_voices"] = {
+                model_id: all_voice_for(model_id, picked)
+                for model_id in kwargs["model_ids"]}
+            return ("convert", BACKEND_AUDIOCPP, kwargs)
+        model_id = result[prefix + "model_id"]
+        # The picked voice (a built-in speaker name on a CustomVoice entry,
+        # a server-side preset otherwise); the client resolves which it is.
+        kwargs["model_id"] = model_id
+        kwargs["voice"] = result[prefix + "audiocpp_voice"] or None
         return ("convert", BACKEND_AUDIOCPP, kwargs)
 
     return fields, mapper
diff --git a/app/ui/runview.py b/app/ui/runview.py
index 19a8f64..759fcfb 100644
--- a/app/ui/runview.py
+++ b/app/ui/runview.py
@@ -155,11 +155,11 @@ class RunView(ScreenView):
         self.server = "starting"
         self.server_message = ""
         self.log_tail: List[str] = []
-        self.book: Optional[tuple] = None      # (index, total, name)
+        self.book: Optional[tuple] = None      # (index, total, name, model)
         self.chapter: Optional[tuple] = None   # (index, total)
         self.chunk_done = 0
         self.chunk_total = 0
-        self.book_results: List[tuple] = []    # (name, ok)
+        self.book_results: List[tuple] = []    # (name, ok, files, error, model)
         self.error_message = ""
         self.started_server = False   # cancelled/cancelling/finished_at: base
         self.boot_started: Optional[float] = None
@@ -211,7 +211,7 @@ class RunView(ScreenView):
         elif kind == "book":
             self.phase = "convert"
             self.book = (event.get("index"), event.get("total"),
-                         event.get("name") or "")
+                         event.get("name") or "", event.get("model"))
             self.chapter = None
             self.chunk_done = 0
             self.chunk_total = 0
@@ -239,11 +239,12 @@ class RunView(ScreenView):
             self.book_results.append((event.get("name") or "?",
                                       bool(event.get("ok")),
                                       list(event.get("files") or []),
-                                      ""))
+                                      "", event.get("model")))
         elif kind == "book_failed":
             self.book_results.append((event.get("name") or "?", False,
                                       list(event.get("files") or []),
-                                      event.get("error") or "conversion failed"))
+                                      event.get("error") or "conversion failed",
+                                      event.get("model")))
             self.error_message = self.error_message or \
                 (event.get("error") or "conversion failed")
         elif kind == "done":
@@ -486,9 +487,10 @@ class RunView(ScreenView):
         lines = ["Audiobook generation finished",
                  f"Output directory: {AUDIOBOOKS_FOLDER}"]
         ok_count = 0
-        for name, ok, files, error in self.book_results:
+        for name, ok, files, error, model in self.book_results:
             ok_count += 1 if ok else 0
-            lines.append(f"{'[OK]' if ok else '[FAIL]'} {name}"
+            label = f"{name} — {model}" if model else name
+            lines.append(f"{'[OK]' if ok else '[FAIL]'} {label}"
                          + (f": {', '.join(files)}" if files else ""))
             if not ok and error:
                 lines.append(f"       {error}")
@@ -583,10 +585,13 @@ class RunView(ScreenView):
     def _draw_progress(self, scr, theme, y, inner_x, label_w, value_x,
                        value_w, width) -> int:
         """The live panel: book, chapter, chunk bar, elapsed, message."""
-        # Book line
+        # Book line ("— model" appended while an "All" run generates with
+        # a specific model, e.g. "1/6  dune.epub — qwen3_tts_..._q8_0")
         if self.book is not None:
-            index, total, name = self.book
+            index, total, name, model = self.book
             book_text = f"{index}/{total}  {name}"
+            if model:
+                book_text += f" — {model}"
         else:
             book_text = "waiting..." if self.phase == "convert" else "-"
         _text(scr, theme, y, inner_x, "Book".ljust(label_w), theme["dim"])
@@ -660,10 +665,11 @@ class RunView(ScreenView):
         _text(scr, theme, y, value_x, _fit(result, value_w),
               theme.get(kind, theme["body"]))
         y += 1
-        for name, ok, _files, _error in self.book_results[:5]:
+        for name, ok, _files, _error, model in self.book_results[:5]:
             mark = "[OK] " if ok else "[FAIL]"
+            label = f"{name} — {model}" if model else name
             _text(scr, theme, y, value_x,
-                  _fit(f"{mark} {name}", value_w),
+                  _fit(f"{mark} {label}", value_w),
                   theme["ok"] if ok else theme["err"])
             y += 1
         if len(self.book_results) > 5: