Three reported issues: #66: the Video stage workflow picker was keyed by scene id, so choosing a workflow for one clip overwrote every sibling clip. It is now keyed by clip id, hydrates each clip's persisted choice, autosaves both clips.workflow_id and form_workflow_id, and Generate One / Generate All resolve the workflow per clip (a scene can mix image->video and video->video extend across its clips). #64: a 30-second ask could produce a 1:51 script - prompts carried no per-scene scale and the duration rescale could only grow. Prompts now state the total seconds and a per-scene budget, and the rescale may shrink (floor 0.5) so over-written boards pull toward the target; content itself is never edited. #67: a hand-written directory path (e.g. ComfyUI's input dir) reached ComfyUI and died as an opaque Errno 21. prepare_media_inputs now rejects directory / missing local files before queueing with a node-naming error, and the agent's run_workflow validates media input_values at call time. Bare Comfy-side names pass untouched. 534 backend tests pass; svelte-check clean.
303 lines
10 KiB
Python
303 lines
10 KiB
Python
"""Chunked script generation: a big board is written in several modest LLM
|
||
calls (not one giant one that hangs), with progress events and per-chunk
|
||
renumbering."""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import tempfile
|
||
from pathlib import Path
|
||
|
||
import pytest
|
||
|
||
import calliope.agent.script_agent as sa
|
||
from calliope.agent.script_agent import _generate_scenes_chunked, generate_script
|
||
|
||
|
||
@pytest.fixture(autouse=True, scope="module")
|
||
def _never_touch_real_db():
|
||
"""Redirect data_dir before ANY write in this module (repo test rule)."""
|
||
import calliope.config as config_module
|
||
from calliope.db import migrate_db
|
||
|
||
prev = {
|
||
"data_dir": config_module.settings.data_dir,
|
||
"assets_dir": config_module.settings.assets_dir,
|
||
"dry_run": config_module.settings.dry_run,
|
||
}
|
||
with tempfile.TemporaryDirectory() as tmp:
|
||
config_module.settings.data_dir = Path(tmp)
|
||
config_module.settings.assets_dir = Path(tmp) / "assets"
|
||
config_module.settings.dry_run = True
|
||
asyncio.run(migrate_db(config_module.settings.db_path))
|
||
yield
|
||
for k, v in prev.items():
|
||
setattr(config_module.settings, k, v)
|
||
|
||
|
||
def _scenes(n: int, start: int = 1):
|
||
return {
|
||
"scenes": [
|
||
{
|
||
"order_index": start + i,
|
||
"heading": f"SCENE {start + i}",
|
||
"action": f"action {start + i}",
|
||
"dialog": "",
|
||
"duration_sec": 5,
|
||
"character_ids": [],
|
||
"location_id": None,
|
||
}
|
||
for i in range(n)
|
||
]
|
||
}
|
||
|
||
|
||
def test_big_board_splits_into_chunks(monkeypatch):
|
||
calls: list[int] = []
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
user = messages[1]["content"]
|
||
# THIS CHUNK: exactly N scenes
|
||
n = 0
|
||
for line in user.splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
break
|
||
calls.append(n)
|
||
return _scenes(n)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
out = asyncio.run(
|
||
_generate_scenes_chunked(
|
||
project_id=1,
|
||
p={"title": "T", "idea": "i", "target_duration": "2 minutes"},
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
required_scenes=9,
|
||
)
|
||
)
|
||
# 9 scenes at SCRIPT_CHUNK=4 → 4+4+1
|
||
assert calls == [4, 4, 1]
|
||
assert len(out) == 9
|
||
assert [s["order_index"] for s in out] == list(range(1, 10))
|
||
|
||
|
||
def test_small_board_single_call(monkeypatch):
|
||
calls: list[int] = []
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
n = 3
|
||
for line in messages[1]["content"].splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
calls.append(n)
|
||
return _scenes(n)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
out = asyncio.run(
|
||
_generate_scenes_chunked(
|
||
project_id=1,
|
||
p={"title": "T", "idea": "i", "target_duration": "30 seconds"},
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
required_scenes=3,
|
||
)
|
||
)
|
||
assert calls == [3]
|
||
assert len(out) == 3
|
||
|
||
|
||
def test_restarts_order_index_is_renumbered(monkeypatch):
|
||
"""A model that restarts order_index at 1 for every chunk must still
|
||
produce a 1..N board."""
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
# Always returns order_index 1..n regardless of the chunk requested
|
||
n = 0
|
||
for line in messages[1]["content"].splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
return _scenes(n, start=1)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
out = asyncio.run(
|
||
_generate_scenes_chunked(
|
||
project_id=1,
|
||
p={"title": "T", "idea": "i", "target_duration": "2 minutes"},
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
required_scenes=8,
|
||
)
|
||
)
|
||
assert [s["order_index"] for s in out] == list(range(1, 9))
|
||
|
||
|
||
def test_short_chunk_retries_only_that_chunk(monkeypatch):
|
||
calls: list[int] = []
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
n = 0
|
||
for line in messages[1]["content"].splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
calls.append(n)
|
||
# First attempt for the 4-scene chunk returns only 2; retry is full.
|
||
if len(calls) == 1:
|
||
return _scenes(2)
|
||
return _scenes(n)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
out = asyncio.run(
|
||
_generate_scenes_chunked(
|
||
project_id=1,
|
||
p={"title": "T", "idea": "i", "target_duration": "1 minute"},
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
required_scenes=6,
|
||
)
|
||
)
|
||
# chunk1 short (2) → retry chunk1 (4); then chunk2 (2)
|
||
assert calls == [4, 4, 2]
|
||
assert len(out) == 6
|
||
|
||
|
||
def test_continuity_tail_is_passed_to_next_chunk(monkeypatch):
|
||
seen_tails: list[bool] = []
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
user = messages[1]["content"]
|
||
seen_tails.append("Scenes already written" in user)
|
||
n = 0
|
||
for line in user.splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
return _scenes(n)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
asyncio.run(
|
||
_generate_scenes_chunked(
|
||
project_id=1,
|
||
p={"title": "T", "idea": "i", "target_duration": "1 minute"},
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
required_scenes=6,
|
||
)
|
||
)
|
||
# First chunk has no tail; the second continues from the first's scenes.
|
||
assert seen_tails == [False, True]
|
||
|
||
|
||
# ─────────────────────────────────────────────────────────────────────────
|
||
# Issue #64 — target runtime is a budget, not a floor
|
||
# ─────────────────────────────────────────────────────────────────────────
|
||
|
||
|
||
def test_chunk_prompt_carries_per_scene_budget():
|
||
from calliope.agent.prompts import build_script_chunk_messages
|
||
|
||
messages = build_script_chunk_messages(
|
||
title="T",
|
||
idea="i",
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
target_duration="30 seconds",
|
||
scene_count=5,
|
||
chunk_start=1,
|
||
chunk_scenes=5,
|
||
)
|
||
user = messages[1]["content"]
|
||
# 30s / 5 scenes = 6s per scene; both the total and the per-scene scale
|
||
# must be stated so the model writes content volume to the runtime.
|
||
assert "Per-scene budget: ~6 seconds" in user
|
||
assert "must play in about 30s" in user
|
||
|
||
|
||
def test_script_prompt_carries_per_scene_budget():
|
||
from calliope.agent.prompts import build_script_messages
|
||
|
||
messages = build_script_messages(
|
||
title="T",
|
||
idea="i",
|
||
beats=[],
|
||
characters=[],
|
||
locations=[],
|
||
target_duration="30 seconds",
|
||
scene_count=5,
|
||
)
|
||
user = messages[1]["content"]
|
||
assert "Per-scene budget: ~6 seconds" in user
|
||
assert "must play in about 30s" in user
|
||
|
||
|
||
def _mk_project() -> int:
|
||
from calliope.db import get_db
|
||
from calliope.config import settings
|
||
|
||
conn = get_db(settings.db_path)
|
||
try:
|
||
cur = conn.execute(
|
||
"INSERT INTO projects (title, idea, target_duration) VALUES ('Budget', 'i', '30 seconds')"
|
||
)
|
||
conn.commit()
|
||
return int(cur.lastrowid)
|
||
finally:
|
||
conn.close()
|
||
|
||
|
||
def _oversized_scenes(n: int) -> dict:
|
||
"""Scenes whose content genuinely runs ~110s total (issue #64 numbers)."""
|
||
long_dialog = "\n".join(f"SPEAKER: {'word ' * 40}" for _ in range(4)) # 160w ≈ 64s
|
||
long_action = " ".join(["moves carefully forward"] * 40) # 160w ≈ 19s
|
||
return {
|
||
"scenes": [
|
||
{
|
||
"order_index": i + 1,
|
||
"heading": f"SCENE {i + 1}",
|
||
"action": long_action,
|
||
"dialog": long_dialog,
|
||
"duration_sec": 8,
|
||
"character_ids": [],
|
||
"location_id": None,
|
||
}
|
||
for i in range(n)
|
||
]
|
||
}
|
||
|
||
|
||
def test_oversized_script_durations_shrink_to_target(monkeypatch):
|
||
"""Issue #64 regression: a 30s target with ~110s of written content must
|
||
rescale scene durations DOWN toward the target (old max(...,1.0) treated
|
||
the target as a floor and left the overrun untouched — 1:51 vs 0:30)."""
|
||
from calliope.config import settings
|
||
from calliope.db import get_db
|
||
|
||
async def fake_structured(messages, temperature=0.7):
|
||
n = 0
|
||
for line in messages[1]["content"].splitlines():
|
||
if line.startswith("THIS CHUNK:"):
|
||
n = int(line.split("exactly", 1)[1].split("scenes")[0].strip())
|
||
return _oversized_scenes(n)
|
||
|
||
monkeypatch.setattr(sa, "generate_structured", fake_structured)
|
||
pid = _mk_project()
|
||
asyncio.run(generate_script(pid, with_clips=False))
|
||
|
||
conn = get_db(settings.db_path)
|
||
try:
|
||
rows = conn.execute(
|
||
"SELECT duration_sec FROM scenes WHERE project_id = ? ORDER BY order_index",
|
||
(pid,),
|
||
).fetchall()
|
||
finally:
|
||
conn.close()
|
||
total = sum(r["duration_sec"] for r in rows)
|
||
# 5 scenes × ~83s estimated ≈ 415s → floor 0.5 caps the shrink at ~207s.
|
||
# The old code left 415s; the fix must pull meaningfully below it.
|
||
assert total <= 220, f"scene durations did not shrink toward the 30s target: {total}s"
|
||
assert all(r["duration_sec"] >= 3 for r in rows)
|