ChemaVXandClaude Fable 5
93a506b636
Build & Deploy ResearchOwl / build-and-push (push) Successful in 9s
shortsmith ya compone la banda sonora (sonar); ahora publica una paleta (pulse, static) y este repo la consume en vivo: el prompt la ofrece con sus notas de mood, validate_spec la usa como fuente de verdad para audio.preset, y editar el preset en /short_spec es la manera gratis de escucharlas. Sin /audio (404 o caída) todo cae a la paleta base y nada deja de renderizar. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
251 lines
9.0 KiB
Python
251 lines
9.0 KiB
Python
"""Escritura del shot spec: prompt, bucle de reintento y fallback.
|
|
|
|
El LLM entra como un callable, así que aquí se prueba el bucle, no a Haiku:
|
|
respuesta buena, respuesta malformada, typo en una prop, y las tres seguidas.
|
|
"""
|
|
import json
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
|
|
from src.generator.shortspec import (
|
|
MAX_ATTEMPTS, ShortSpecWriter, SpecWriteFailed, extract_json,
|
|
)
|
|
from tests.test_spec_contract import TEMPLATES
|
|
|
|
EXAMPLE = Path(__file__).resolve().parents[1] / "src/generator/examples/jal1628.json"
|
|
|
|
GOOD = {
|
|
"version": 1,
|
|
"meta": {"id": "caso", "title": "Un caso"},
|
|
"shots": [
|
|
{"template": "radar_sweep", "duration": 15.0, "props": {"headline": "3 RADARS"}},
|
|
{"template": "scale_bars", "duration": 15.0, "props": {
|
|
"headline": "REPORTED SCALE",
|
|
"bars": [{"label": "BOEING 747", "value": 232, "unit": "FT"}]}},
|
|
],
|
|
}
|
|
|
|
|
|
class FakeLLM:
|
|
"""Devuelve respuestas de una cola y guarda los prompts que recibió."""
|
|
|
|
def __init__(self, *responses):
|
|
self.responses = list(responses)
|
|
self.prompts: list[str] = []
|
|
self.systems: list[str] = []
|
|
|
|
async def __call__(self, system, prompt):
|
|
self.systems.append(system)
|
|
self.prompts.append(prompt)
|
|
return self.responses.pop(0) if len(self.responses) > 1 else self.responses[0]
|
|
|
|
|
|
def writer(*responses, **kw):
|
|
llm = FakeLLM(*responses)
|
|
return ShortSpecWriter(llm, TEMPLATES, **kw), llm
|
|
|
|
|
|
# --- parseo -----------------------------------------------------------------
|
|
|
|
def test_extract_json_survives_markdown_fences():
|
|
assert extract_json('```json\n{"a": 1}\n```') == {"a": 1}
|
|
assert extract_json('Here you go:\n{"a": 1}\nHope that helps') == {"a": 1}
|
|
assert extract_json('{"a": 1}') == {"a": 1}
|
|
|
|
|
|
def test_extract_json_complains_when_there_is_no_object():
|
|
with pytest.raises(ValueError):
|
|
extract_json("I'm afraid I can't do that")
|
|
|
|
|
|
def test_straight_quotes_inside_a_string_are_repaired():
|
|
"""Cómo falla esto en la vida real (sesión de Bélgica, 2026-08-01): el
|
|
modelo escribe la cita con comillas rectas, que cierran la cadena JSON antes
|
|
de tiempo. Se arregla aquí porque además es lo que se quiere dibujar."""
|
|
broken = '{"quote_a": ""CREDIBLE PEOPLE. THEY TOLD WHAT THEY SAW."", "n": 1}'
|
|
assert extract_json(broken) == {
|
|
"quote_a": "“CREDIBLE PEOPLE. THEY TOLD WHAT THEY SAW.”", "n": 1}
|
|
|
|
|
|
def test_the_repair_leaves_correct_json_alone():
|
|
good = {"a": 'texto con “tipográficas” dentro', "b": [1, 2], "c": {"d": "e"}}
|
|
assert extract_json(json.dumps(good, ensure_ascii=False)) == good
|
|
|
|
|
|
def test_the_repair_does_not_eat_escaped_quotes():
|
|
assert extract_json(r'{"a": "dijo \"hola\" y se fue"}') == {"a": 'dijo "hola" y se fue'}
|
|
|
|
|
|
def test_an_unrepairable_response_reports_the_offending_fragment():
|
|
"""El modelo no ve su salida numerada: "line 189 column 22" no le sirve; el
|
|
trozo sí."""
|
|
with pytest.raises(ValueError) as exc:
|
|
extract_json('{"a": 1, "b": [1, 2,,,], "c": 3}')
|
|
assert "aquí:" in str(exc.value)
|
|
|
|
|
|
# --- el prompt --------------------------------------------------------------
|
|
|
|
def test_prompt_carries_the_fetched_contract_not_a_copy():
|
|
w, _ = writer("{}")
|
|
prompt = w.build_prompt("Caso X", "material", None, "THEEXCLUSIONZONE.COM")
|
|
assert "radar_sweep:" in prompt and "contact_bearing_deg" in prompt
|
|
assert "1-3 elementos" in prompt # los límites de longitud, del esquema
|
|
assert "ink, amber, amber_dark" in prompt # la paleta, también del esquema
|
|
|
|
|
|
def test_prompt_states_the_editorial_constraints():
|
|
w, _ = writer("{}")
|
|
prompt = w.build_prompt("Caso X", "material", "https://x.test/post", "X.TEST")
|
|
assert "20-45 seconds" in prompt
|
|
assert "never hex" in prompt
|
|
assert "https://x.test/post" in prompt
|
|
assert "X.TEST" in prompt
|
|
assert "material" in prompt
|
|
|
|
|
|
def test_prompt_includes_the_worked_example_in_full():
|
|
w, _ = writer("{}")
|
|
prompt = w.build_prompt("Caso X", "material", None, "X.TEST")
|
|
example = json.loads(EXAMPLE.read_text())
|
|
assert example["meta"]["title"] in prompt
|
|
assert "counter_close" in prompt
|
|
|
|
|
|
def test_prompt_says_out_loud_that_there_is_no_article_yet():
|
|
w, _ = writer("{}")
|
|
assert "No article URL yet" in w.build_prompt("X", "m", None, "X.TEST")
|
|
|
|
|
|
def test_prompt_offers_the_live_audio_palette():
|
|
"""La mitad de audio del contrato vivo: los presets y sus notas de mood
|
|
vienen de GET /audio, no de este repo."""
|
|
palette = {"sonar": "the case-file mood", "pulse": "tension, built for debunks"}
|
|
w, _ = writer("{}", presets=palette)
|
|
prompt = w.build_prompt("X", "m", None, "X.TEST")
|
|
assert '"pulse" — tension, built for debunks' in prompt
|
|
assert "whose mood fits the shape" in prompt
|
|
|
|
|
|
def test_prompt_without_a_palette_only_offers_the_baseline():
|
|
w, _ = writer("{}")
|
|
prompt = w.build_prompt("X", "m", None, "X.TEST")
|
|
assert '"sonar"' in prompt and '"none"' in prompt
|
|
assert '"pulse"' not in prompt
|
|
|
|
|
|
# --- bucle ------------------------------------------------------------------
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_a_good_response_validates_on_the_first_attempt():
|
|
w, llm = writer(json.dumps(GOOD))
|
|
result = await w.write("Caso X", "material")
|
|
assert result.attempts == 1
|
|
assert result.spec["meta"]["id"] == "caso"
|
|
assert len(llm.prompts) == 1
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_malformed_json_triggers_a_retry():
|
|
w, llm = writer("no soy JSON", json.dumps(GOOD))
|
|
result = await w.write("Caso X", "material")
|
|
assert result.attempts == 2
|
|
assert "no es un objeto JSON válido" in result.history[0][0]
|
|
assert "previous attempt was rejected" in llm.prompts[1]
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_the_exact_error_paths_are_fed_back_verbatim():
|
|
"""La ruta que devuelve la validación es lo más útil que se le puede dar al
|
|
modelo: se le pasa tal cual, sin parafrasear."""
|
|
bad = json.loads(json.dumps(GOOD))
|
|
bad["shots"][0]["props"]["sweeeps"] = 2
|
|
w, llm = writer(json.dumps(bad), json.dumps(GOOD))
|
|
|
|
result = await w.write("Caso X", "material")
|
|
|
|
assert result.attempts == 2
|
|
assert "shots.0.radar_sweep.props.sweeeps" in llm.prompts[1]
|
|
assert "sweeps" in llm.prompts[1] # y cuáles sí valen
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_three_failures_raise_but_keep_the_last_attempt():
|
|
"""La parte cara es la generación, no el render: el último intento viaja en
|
|
la excepción para poder editarlo a mano y reenviarlo."""
|
|
bad = json.loads(json.dumps(GOOD))
|
|
bad["meta"]["id"] = "Caso Con Espacios"
|
|
w, _ = writer(json.dumps(bad))
|
|
|
|
with pytest.raises(SpecWriteFailed) as exc:
|
|
await w.write("Caso X", "material")
|
|
|
|
assert exc.value.attempts == MAX_ATTEMPTS
|
|
assert exc.value.last_spec["meta"]["id"] == "Caso Con Espacios"
|
|
assert any("meta.id" in e for e in exc.value.errors)
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_an_off_target_duration_is_commented_once_then_accepted():
|
|
"""70 s cumple el contrato pero no el encargo: se comenta y, si el modelo
|
|
insiste, se renderiza igual antes que tirar la generación."""
|
|
long_spec = json.loads(json.dumps(GOOD))
|
|
long_spec["shots"][0]["duration"] = 55.0 # 70 s en total
|
|
w, llm = writer(json.dumps(long_spec))
|
|
|
|
result = await w.write("Caso X", "material")
|
|
|
|
assert result.attempts == MAX_ATTEMPTS
|
|
assert result.notes and "recorta" in result.notes[0]
|
|
assert "off-brief" in llm.prompts[1]
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_a_valid_attempt_is_not_thrown_away_by_a_worse_one():
|
|
long_spec = json.loads(json.dumps(GOOD))
|
|
long_spec["shots"][0]["duration"] = 55.0
|
|
w, _ = writer(json.dumps(long_spec), "esto ya no es JSON", "tampoco")
|
|
|
|
result = await w.write("Caso X", "material")
|
|
|
|
assert result.spec["shots"][0]["duration"] == 55.0
|
|
assert result.notes
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_the_contract_is_refetched_after_a_validation_failure():
|
|
"""Si el renderizador se actualizó a mitad de la run, la plantilla nueva
|
|
entra en el segundo intento."""
|
|
new_template = {"type": "object", "additionalProperties": False,
|
|
"required": ["title"],
|
|
"properties": {"title": {"type": "string", "minLength": 1}}}
|
|
refreshed = {**TEMPLATES, "holo_scan": new_template}
|
|
|
|
async def refresh():
|
|
return refreshed
|
|
|
|
with_new = json.loads(json.dumps(GOOD))
|
|
with_new["shots"][1] = {"template": "holo_scan", "duration": 15.0,
|
|
"props": {"title": "X"}}
|
|
w, llm = writer(json.dumps(with_new), json.dumps(with_new),
|
|
refresh_templates=refresh)
|
|
|
|
result = await w.write("Caso X", "material")
|
|
|
|
assert result.attempts == 2
|
|
assert "holo_scan" in llm.prompts[1]
|
|
assert result.spec["shots"][1]["template"] == "holo_scan"
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_progress_is_reported_only_when_it_retries():
|
|
seen = []
|
|
|
|
async def on_progress(text):
|
|
seen.append(text)
|
|
|
|
w, _ = writer("no JSON", json.dumps(GOOD))
|
|
await w.write("Caso X", "material", on_progress=on_progress)
|
|
assert len(seen) == 1 and "attempt 2/3" in seen[0]
|