calderbuild
516324fe93
- Add Pydantic input validation (prompt length, duration, num_shots) - Per-job output directory for concurrent job isolation - LLM JSON parse retry (1 retry with explicit JSON instruction) - Music failure surfaced as amber "Skipped" state in frontend - Agent card states: completed (green), failed (red), skipped (amber) - Error banner with "Try Again" button on pipeline failure - Elapsed timer (MM:SS) during production - Script preview section after Script agent completes - Form inputs disabled during production - Style presets (5 visual cards replacing plain text chips) - Demo gallery section with placeholder videos - Production summary (time, shots, agents, cost) on completion - Structured logging replacing bare print calls - Background job cleanup (1 hour TTL, 10 min interval) - Per-stage error propagation with tracebacks in logs
92 lines
3.4 KiB
Python
92 lines
3.4 KiB
Python
"""Script Agent: refines the Director's outline into production-ready prompts."""
|
|
import json
|
|
from backend.config import MINIMAX_BASE_URL, api_session
|
|
|
|
|
|
SYSTEM_PROMPT = """You are the Scriptwriter of an AI film production team. You take the Director's shot outline and refine each shot into production-ready specifications.
|
|
|
|
For each shot, enhance:
|
|
1. video_prompt: a highly detailed, specific prompt optimized for AI video generation (Hailuo). Include: subject, action, camera movement, lighting, color palette, style. Be vivid and specific. Max 200 words.
|
|
2. narration: polished voiceover text, timed to fit the shot duration. Keep it concise and impactful.
|
|
3. subtitle: clean subtitle text (shorter than narration if needed)
|
|
|
|
Respond ONLY with valid JSON. Format:
|
|
{
|
|
"shots": [
|
|
{
|
|
"shot_number": 1,
|
|
"duration": 10,
|
|
"video_prompt": "Detailed AI video generation prompt...",
|
|
"narration": "Polished voiceover text...",
|
|
"subtitle": "Clean subtitle text...",
|
|
"mood": "inspiring"
|
|
}
|
|
]
|
|
}"""
|
|
|
|
|
|
def _strip_markdown_fences(text: str) -> str:
|
|
"""Strip markdown code fences from LLM response."""
|
|
text = text.strip()
|
|
if text.startswith("```"):
|
|
text = text.split("\n", 1)[1] if "\n" in text else text[3:]
|
|
if text.endswith("```"):
|
|
text = text[:-3]
|
|
text = text.strip()
|
|
return text
|
|
|
|
|
|
def _call_llm(messages: list, max_tokens: int = 3000) -> str:
|
|
"""Call MiniMax LLM and return the content string."""
|
|
url = f"{MINIMAX_BASE_URL}/text/chatcompletion_v2"
|
|
payload = {
|
|
"model": "MiniMax-M1",
|
|
"messages": messages,
|
|
"temperature": 0.7,
|
|
"max_tokens": max_tokens,
|
|
}
|
|
resp = api_session.post(url, json=payload)
|
|
resp.raise_for_status()
|
|
data = resp.json()
|
|
|
|
base_resp = data.get("base_resp", {})
|
|
if base_resp.get("status_code", 0) != 0:
|
|
raise RuntimeError(f"MiniMax API error: {base_resp.get('status_msg', 'unknown')} (code {base_resp.get('status_code')})")
|
|
|
|
content = data.get("choices", [{}])[0].get("message", {}).get("content", "")
|
|
if not content:
|
|
raise RuntimeError("MiniMax API returned empty response")
|
|
return content
|
|
|
|
|
|
def run(outline: dict) -> dict:
|
|
"""Refine Director's outline into production-ready shot specifications."""
|
|
user_message = (
|
|
f"Video title: {outline['title']}\n"
|
|
f"Visual style: {outline['style']}\n"
|
|
f"Total duration: {outline['total_duration']}s\n\n"
|
|
f"Director's outline:\n{json.dumps(outline['shots'], indent=2)}\n\n"
|
|
f"Refine each shot with detailed video prompts optimized for AI video generation."
|
|
)
|
|
|
|
messages = [
|
|
{"role": "system", "content": SYSTEM_PROMPT},
|
|
{"role": "user", "content": user_message},
|
|
]
|
|
|
|
content = _call_llm(messages)
|
|
try:
|
|
result = json.loads(_strip_markdown_fences(content))
|
|
except json.JSONDecodeError:
|
|
# Retry once with explicit JSON instruction
|
|
messages.append({"role": "assistant", "content": content})
|
|
messages.append({"role": "user", "content": "Your response was not valid JSON. Please return ONLY valid JSON with no markdown or extra text."})
|
|
content = _call_llm(messages)
|
|
result = json.loads(_strip_markdown_fences(content))
|
|
|
|
# Carry over metadata from outline
|
|
result["title"] = outline["title"]
|
|
result["style"] = outline["style"]
|
|
result["total_duration"] = outline["total_duration"]
|
|
return result
|