Track four test files that were never committed

CLAUDE.md cites tests/test_model_config.py as the guard against reintroducing
retired OpenRouter model ids, but the file existed only on disk — the guard was
not in the repository. Same for the town geo, voice roster and voicemail
transcript tests, and the Reaper bleep-selection test.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_model_config.py

@@ -0,0 +1,90 @@
+"""Guards against the two model-config failures that have bitten us:
+a routed model with no pricing entry (silently costs $0.00 on the dashboard),
+and a stale model id that 404s at the OpenRouter API."""
+
+from backend.config import settings
+from backend.services.cost_tracker import OPENROUTER_PRICING
+from backend.services.llm import OPENROUTER_MODELS, LLMService
+
+_CALLER_DIALOG_MODEL_PARAMS = LLMService._CALLER_DIALOG_MODEL_PARAMS
+
+# Retired on OpenRouter — kept here so a reintroduction fails loudly.
+RETIRED_MODELS = {
+    "anthropic/claude-3.5-haiku",
+    "anthropic/claude-3.5-sonnet",
+    "anthropic/claude-sonnet-4-5",   # note the dash; the live id uses a dot
+    "google/gemini-flash-1.5",
+    "mistralai/mistral-small-creative",
+    "x-ai/grok-4",
+    "x-ai/grok-4-fast",
+    "x-ai/grok-4.1-fast",
+}
+
+
+def test_every_routed_model_has_pricing():
+    """A routed model missing from OPENROUTER_PRICING records cost as $0.00."""
+    unpriced = {
+        cat: model
+        for cat, model in settings.category_models.items()
+        if model not in OPENROUTER_PRICING
+    }
+    assert not unpriced, f"routed models with no pricing entry: {unpriced}"
+
+
+def test_no_retired_models_are_routed():
+    routed = {
+        cat: model
+        for cat, model in settings.category_models.items()
+        if model in RETIRED_MODELS
+    }
+    assert not routed, f"category routed to a retired model: {routed}"
+
+
+def test_no_retired_models_in_the_pool():
+    stale = sorted(set(OPENROUTER_MODELS) & RETIRED_MODELS)
+    assert not stale, f"retired models still listed in OPENROUTER_MODELS: {stale}"
+
+
+def test_no_retired_models_in_caller_dialog_params():
+    stale = sorted(set(_CALLER_DIALOG_MODEL_PARAMS) & RETIRED_MODELS)
+    assert not stale, f"per-model tuning keyed to retired models: {stale}"
+
+
+def test_pool_models_are_priced():
+    """Anything selectable should be costable."""
+    unpriced = [m for m in OPENROUTER_MODELS if m not in OPENROUTER_PRICING]
+    assert not unpriced, f"pool models with no pricing entry: {unpriced}"
+
+
+def test_no_retired_model_ids_anywhere_in_the_codebase():
+    """publish_episode.py, make_clips.py and relabel_transcripts.py each shipped
+    a retired model id and only failed when someone ran them. Scan every source
+    file so the next one fails here instead.
+
+    cost_tracker.py is exempt: it intentionally keeps retired ids as pricing
+    keys so historical cost records stay costable.
+    """
+    import pathlib
+
+    root = pathlib.Path(__file__).resolve().parent.parent
+    exempt = {root / "backend" / "services" / "cost_tracker.py",
+              pathlib.Path(__file__).resolve()}
+
+    offenders = {}
+    for path in root.rglob("*.py"):
+        if path in exempt:
+            continue
+        # Skip venvs, vendored models, and git worktrees (separate checkouts
+        # on other branches — not this tree's code).
+        if any(part in {"venv", "mlx_models", ".git", ".claude", ".worktrees",
+                        "node_modules", "remotion-demo"} for part in path.parts):
+            continue
+        try:
+            text = path.read_text(encoding="utf-8", errors="ignore")
+        except OSError:
+            continue
+        hits = sorted(m for m in RETIRED_MODELS if f'"{m}"' in text or f"'{m}'" in text)
+        if hits:
+            offenders[str(path.relative_to(root))] = hits
+
+    assert not offenders, f"retired model ids still referenced: {offenders}"

tests/test_town_geo.py

@@ -0,0 +1,29 @@
+from backend.main import _get_town_from_location, BIG_BEND_TOWNS
+
+
+def test_matches_simple_town():
+    assert _get_town_from_location("Alpine, Texas") == "alpine"
+    assert _get_town_from_location("Marfa") == "marfa"
+    assert _get_town_from_location("outside Terlingua") == "terlingua"
+
+
+def test_matches_two_word_town():
+    assert _get_town_from_location("Fort Stockton, TX") == "fort stockton"
+    assert _get_town_from_location("ft stockton") == "fort stockton"
+
+
+def test_case_insensitive():
+    assert _get_town_from_location("MARATHON, texas") == "marathon"
+
+
+def test_no_match_returns_none():
+    assert _get_town_from_location("Albuquerque, New Mexico") is None
+    assert _get_town_from_location("") is None
+
+
+def test_every_town_has_coords():
+    for town, coords in BIG_BEND_TOWNS.items():
+        assert town == town.lower()
+        lat, lon = coords
+        assert 28.0 < lat < 32.0
+        assert -105.0 < lon < -102.0

tests/test_voice_roster.py

@@ -0,0 +1,47 @@
+from backend.main import (
+    INWORLD_MALE_VOICES,
+    INWORLD_FEMALE_VOICES,
+    BLACKLISTED_VOICES,
+)
+
+NEWLY_ADDED = ["Arthur", "Daniel", "Brooke", "Joy", "Selene", "Zadie"]
+
+# Child / adolescent voices — deliberately kept off the roster for an
+# adult late-night call-in show.
+EXCLUDED_CHILD_VOICES = ["Abby", "Mia", "Riley"]
+
+
+def test_rosters_are_sorted_and_deduped():
+    for roster in (INWORLD_MALE_VOICES, INWORLD_FEMALE_VOICES):
+        assert roster == sorted(roster), "roster must stay alphabetical"
+        assert len(roster) == len(set(roster)), "duplicate voice in roster"
+
+
+def test_no_voice_appears_in_both_genders():
+    overlap = set(INWORLD_MALE_VOICES) & set(INWORLD_FEMALE_VOICES)
+    assert not overlap, f"voice in both rosters: {overlap}"
+
+
+def test_newly_added_voices_are_present():
+    roster = set(INWORLD_MALE_VOICES) | set(INWORLD_FEMALE_VOICES)
+    missing = [v for v in NEWLY_ADDED if v not in roster]
+    assert not missing, f"missing: {missing}"
+
+
+def test_child_voices_stay_off_the_roster():
+    roster = set(INWORLD_MALE_VOICES) | set(INWORLD_FEMALE_VOICES)
+    present = [v for v in EXCLUDED_CHILD_VOICES if v in roster]
+    assert not present, f"child/adolescent voice on roster: {present}"
+
+
+def test_blacklist_only_references_real_roster_voices():
+    roster = set(INWORLD_MALE_VOICES) | set(INWORLD_FEMALE_VOICES)
+    orphans = sorted(set(BLACKLISTED_VOICES) - roster)
+    assert not orphans, f"blacklist names voices not in the roster: {orphans}"
+
+
+def test_effective_pool_is_large_enough_for_session_sampling():
+    """main.py samples 25 voices per session from the non-blacklisted pool."""
+    roster = set(INWORLD_MALE_VOICES) | set(INWORLD_FEMALE_VOICES)
+    effective = roster - set(BLACKLISTED_VOICES)
+    assert len(effective) >= 25

tests/test_voicemail_transcript.py

@@ -0,0 +1,138 @@
+import asyncio
+import json
+
+import pytest
+
+from backend.main import (
+    Voicemail,
+    _load_voicemails,
+    _save_voicemails,
+    play_voicemail_on_air,
+)
+import backend.main as main
+
+
+@pytest.fixture
+def isolated_store(tmp_path, monkeypatch):
+    """Point the voicemail store at a temp dir and start from a clean slate."""
+    meta = tmp_path / "voicemails.json"
+    monkeypatch.setattr(main, "VOICEMAILS_META", meta)
+    monkeypatch.setattr(main, "_voicemails", [])
+    monkeypatch.setattr(main, "_deleted_vm_timestamps", set())
+    return meta
+
+
+def test_voicemail_defaults_to_empty_transcript():
+    vm = Voicemail(id="a1", phone="+15551234567", timestamp=1.0, duration=30, file_path="/x.wav")
+    assert vm.transcript == ""
+
+
+def test_transcript_round_trips_through_save_and_load(isolated_store):
+    main._voicemails.append(
+        Voicemail(
+            id="a1", phone="+15551234567", timestamp=1.0, duration=30,
+            file_path="/x.wav", transcript="Hey Luke, it's Susan out in Deming.",
+        )
+    )
+    _save_voicemails()
+
+    on_disk = json.loads(isolated_store.read_text())
+    assert on_disk["voicemails"][0]["transcript"] == "Hey Luke, it's Susan out in Deming."
+
+    main._voicemails.clear()
+    _load_voicemails()
+    assert main._voicemails[0].transcript == "Hey Luke, it's Susan out in Deming."
+
+
+def test_load_tolerates_legacy_entries_without_transcript(isolated_store):
+    isolated_store.write_text(json.dumps({
+        "voicemails": [{
+            "id": "old", "phone": "+15551234567", "timestamp": 1.0,
+            "duration": 30, "file_path": "/x.wav", "listened": True,
+        }],
+        "deleted_timestamps": [],
+    }))
+    _load_voicemails()
+    assert main._voicemails[0].transcript == ""
+
+
+def test_play_on_air_puts_transcript_into_conversation(isolated_store, tmp_path, monkeypatch):
+    """The actual fix: Devon reads session.conversation, so playing a
+    voicemail on air must write its transcript there."""
+    wav = tmp_path / "vm.wav"
+    wav.write_bytes(b"fake wav bytes")
+
+    main._voicemails.append(
+        Voicemail(
+            id="vm1", phone="+15753424105", timestamp=1.0, duration=77,
+            file_path=str(wav), transcript="My neighbor keeps moving my fence posts.",
+        )
+    )
+
+    monkeypatch.setattr(main.session, "conversation", [])
+    monkeypatch.setattr(main, "_save_voicemails", lambda: None)
+
+    played = {}
+    monkeypatch.setattr(
+        main.audio_service, "play_caller_audio",
+        lambda data, sr: played.update(bytes=len(data), rate=sr),
+    )
+
+    class FakeLibrosa:
+        @staticmethod
+        def load(path, sr=None, mono=True):
+            import numpy as np
+            return np.zeros(sr or 24000, dtype="float32"), sr or 24000
+
+    monkeypatch.setitem(__import__("sys").modules, "librosa", FakeLibrosa)
+
+    result = asyncio.run(play_voicemail_on_air("vm1"))
+    assert result["status"] == "playing"
+
+    roles = [m["role"] for m in main.session.conversation]
+    contents = [m["content"] for m in main.session.conversation]
+    assert any("voicemail" in r for r in roles), f"no voicemail role in {roles}"
+    assert any("fence posts" in c for c in contents), f"transcript missing from {contents}"
+
+
+def test_voicemail_role_is_normalized_to_a_valid_llm_role():
+    """A raw 'voicemail:+1555...' role would reach the LLM API as an invalid
+    role; llm.py swallows the resulting error and returns empty text, so
+    callers would silently go mute."""
+    out = main._normalize_messages_for_llm([
+        {"role": "voicemail:+15753424105", "content": "My name is Sandra."},
+    ])
+    assert out[0]["role"] in ("user", "assistant", "system"), out[0]["role"]
+    assert "Sandra" in out[0]["content"]
+    assert "+15753424105" in out[0]["content"], "caller identity should survive"
+
+
+def test_play_on_air_without_transcript_still_announces_the_voicemail(isolated_store, tmp_path, monkeypatch):
+    """An untranscribed voicemail should still tell Devon something happened,
+    rather than silently playing audio he can't perceive."""
+    wav = tmp_path / "vm.wav"
+    wav.write_bytes(b"fake wav bytes")
+
+    main._voicemails.append(
+        Voicemail(
+            id="vm2", phone="+15753424105", timestamp=1.0, duration=77,
+            file_path=str(wav), transcript="",
+        )
+    )
+
+    monkeypatch.setattr(main.session, "conversation", [])
+    monkeypatch.setattr(main, "_save_voicemails", lambda: None)
+    monkeypatch.setattr(main.audio_service, "play_caller_audio", lambda data, sr: None)
+
+    class FakeLibrosa:
+        @staticmethod
+        def load(path, sr=None, mono=True):
+            import numpy as np
+            return np.zeros(sr or 24000, dtype="float32"), sr or 24000
+
+    monkeypatch.setitem(__import__("sys").modules, "librosa", FakeLibrosa)
+
+    asyncio.run(play_voicemail_on_air("vm2"))
+
+    assert main.session.conversation, "nothing was added to the conversation"
+    assert any("voicemail" in m["role"] for m in main.session.conversation)