Markdown-to-video pipeline: reveal.js slides rendered to PDF, per-slide TTS narration, combined into a narrated mp4 with intro and thumbnail
md2video exposes 7 tools with inconsistent quality. Tools 1-5 (Python scripts) have detailed parameter schemas with types and descriptions, but lack structured output documentation. Tools 6-7 (Node.js scripts) have minimal schemas, only html_path parameter visible, no output specs. Naming is verb-forward (combine_video, tts_from_script, upload_bilibili) but some are compound actions (combine_video does render+segment+concat). Descriptions range from 50-200 chars, meeting baseline, but lack error recovery guidance. No tool declares permissions, error categories, or idempotency. Parameter defaults are reasonable (no destructive defaults), but several tools accept file paths without validation hints. Output schemas are entirely undocumented, agents cannot plan downstream calls.
Validates chart configurations in a reveal.js presentation. Checks for valid JSON syntax and required properties in canvas[data-chart] elements.
Checks reveal.js slides for content overflow using Puppeteer. Detects vertical and horizontal overflow on each slide.
Build a narrated video from output.pdf and per-slide MP3 files. Renders PDF pages to images, generates video segments with audio, and combines them into a final MP4.
Live read-only dashboard for the md2video pipeline. Scans output/<slug>/ on every request and reports per-deck stage status (slides, audio, video, intro, thumb) derived entirely from the filesystem.
Generate a video thumbnail PNG from a text-to-image prompt using DashScope Qwen-Image.
Generate per-slide TTS narration from a script.json file. Supports both DashScope and OpenAI TTS providers with optional subtitle generation.
Upload a finished presentation video to Bilibili. Automatically extracts login credentials from the local browser and drives the upload process.
Tools 6-7 (check_charts, check_overflow) have minimal schemas. Only html_path parameter visible; no output schema documented. Agents cannot infer what these tools return or when to call them.
No tool documents output schema. Agents cannot plan downstream calls or extract required fields (e.g., does tts_from_script return file paths? timestamps?). This violates the response-shaper pattern.
No error handling guidance. Tools accept file paths but do not document validation rules, missing-file recovery, or permission errors. Agents have no recovery path on failure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
combine_video and tts_from_script perform multiple sequential operations (render→segment→concat, script→audio→subtitles) but are not split into composable tools. Agents cannot retry individual steps or reuse intermediate outputs.
upload_bilibili auto-discovers credentials from browser and performs irreversible upload without explicit confirmation. No dry-run or confirmation-request pattern. Agents risk unintended uploads.