Live event logs (session 908) showed a planner-scheduled render task whose
tools were hidden by the HITL visibility rule, a story sub-agent denied the
add_scene tools the planner told it to use, deep-linked chats orphaning as
duplicate empty sessions, and new project assets missing from the canvas
until re-open. Render-intent now recognizes clip/shot/storyboard; sub-agent
payloads keep approval-gated tools visible (permission stays at the
execute-time guard, whose denial teaches ask_user); the story role owns
scene structure; the canvas deep-link session param is trusted; GET
/api/canvas/{id} auto-seeds new entity cards.
55 lines
2.4 KiB
Python
55 lines
2.4 KiB
Python
"""Regression tests for render-intent detection (policy.is_render_request).
|
|
|
|
The failure that motivated this: "all 20 scenes use this text2video workflow.
|
|
Do not overthinking, no image need." read as NON-render (compound word
|
|
"text2video" unmatched + a message-global negation veto from "Do not"/"no"),
|
|
so the registry hid enqueue_video_jobs/run_workflow, and the agent — unable
|
|
to say why — fabricated job ids instead of admitting the tools were hidden.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from calliope.agent.harness.policy import is_render_request
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("text", "want"),
|
|
[
|
|
# The exact messages from the canvas/62 incident
|
|
("all 20 scenes use this text2video workflow. Do not overthinking, no image need.", True),
|
|
("you are not doing the job go away", False),
|
|
# Compound / model-name cues
|
|
("make a text2video clip of the establishing shot", True),
|
|
("txt2vid for scene 3 please", True),
|
|
("use fastvideoH3_t2v-API for all 20 scenes", True),
|
|
("image-to-video on scene 2", True),
|
|
# Negation must stay authoritative when it governs the render cue
|
|
("do not generate videos", False),
|
|
("no images please", False),
|
|
("don't render anything", False),
|
|
("no render, just fix the script", False),
|
|
("i said do not render", False),
|
|
# …but a negated clause must not veto an unrelated render clause
|
|
("fix the script but render afterwards", True),
|
|
("no image need, render all scenes", True),
|
|
# Plain positive cues
|
|
("render all scenes now", True),
|
|
("yes, generate the images", True),
|
|
# Clip/shot are render-unit nouns (session 908: "shot clips" left the
|
|
# video sub-agent's enqueue tools hidden → "generation tool is not
|
|
# exposed")
|
|
("Break the beats into the 8-scene script + shot clips", True),
|
|
("break scene 3 into clips", True),
|
|
("render the shots for scene 2", True),
|
|
# Non-render text must not flip positive
|
|
("generate the story", False),
|
|
("provide the lyrics", False),
|
|
# 'clip' as a noun of speech ("clip the beard") is not a cue — but the
|
|
# verb is rare in film asks; keep the broad match and rely on
|
|
# clause-scoped negation for safety.
|
|
],
|
|
)
|
|
def test_render_intent(text: str, want: bool):
|
|
assert is_render_request(text) is want, text
|