Skip to content

generation-model-strings.md

Latest commit

 

History

History
46 lines (36 loc) · 1.64 KB

generation-model-strings.md

File metadata and controls

46 lines (36 loc) · 1.64 KB
Raw

Milestone: narration / scene-generation model and prompts must be strings

Status: Shipped
PR: #107
Depends on: milestones/manim-font-quality.md (PR #106), milestones/wizard-prompt-strings.md (PR #100)

Problem

wizard.llm_model / wizard.system_prompt and tts.model are already typed at config load. These LLM keys were not:

  1. narration_from_source.model — str() turned a YAML list into "['gpt-4o']" and sent it to the chat API.
  2. narration_from_source.system_prompt — a list became a bracketed string used as the narration system prompt.
  3. manim_scene_generation.model / system_prompt / scene_spec_system_prompt — same str() coercion into scene-spec generation.

Goal

Fail closed at Config.from_yaml. Missing keys still use code defaults (gpt-4o / built-in prompts). Empty system_prompt strings remain allowed (same as wizard.system_prompt).

Done when

  • Present narration_from_source.model must be a non-empty YAML string.
  • Present narration_from_source.system_prompt must be a YAML string (empty allowed).
  • Present manim_scene_generation.model must be a non-empty YAML string.
  • Present manim_scene_generation.system_prompt / scene_spec_system_prompt must be YAML strings (empty allowed).
  • Tests for list values of those keys.
  • ruff check src/ tests/
  • pytest tests/
  • docgen benchmark (no clock change)

Out of scope

  • Per-segment system_prompt / class_name under those blocks.
  • wizard.default_guidance type gating is separate.
  • env_file / repo_root / dirs.* path strings are separate.