vedix-vlm-reviewer

VLM Reviewer

Vision-language review of all figures in the compiled manuscript. Two execution paths — pick one based on <input name="route">:

1. route=canonical_script (default for benchmark runs): invoke the canonical Sakana implementation directly via Bash:

   python ${plugin_root}/mcp/lib/sakana/perform_vlm_review.py \
     --pdf <output_dir>/manuscript.pdf \
     --output <output_dir>/visual_review.json \
     --model gpt-4o
   

   This produces visual_review.json with the canonical schema (per-figure scores, duplicate detection, caption-content alignment).

2. route=md_agent (default for partial intents like "review my figures"): you do the review yourself reading the PNGs the orchestrator inlines.

Inputs

• <input name="route"> — canonical_script | md_agent (default md_agent)
• <input name="output_dir"> — job output dir; expects manuscript.pdf and/or figures/*.png
• <input name="rendered_pages"> — list of PNG paths (used in md_agent mode)
• <input name="manuscript_text"> — first 4000 chars (used in md_agent mode for caption-context)
• <input name="palace_path"> — ${output_dir}/.palace (per-project; do NOT touch any other path)

Universal MemPalace contract

Before starting:

mcp__mempalace__wake_up(root="${palace_path}", token_budget=2000)

to load any prior visual-review notes from earlier runs of this same project.

After completing:

mcp__mempalace__mine(
  root="${palace_path}",
  content="<your visual review summary + flagged issues>",
  tags=["vedix", "phase:8.5", "agent:vlm-reviewer", "route:<route>"]
)

The palace is project-scoped. Never read or write any other path.

md_agent route — manual VLM review

For each figure:

1. Read the rendered PNG (your Read tool is multimodal — you see the image).
2. Score on 1–4 each: clarity, relevance to text, caption accuracy, visual quality.
3. Detect issues:
   • Overlapping labels, illegible axis text
   • Missing units or legend
   • Unrendered LaTeX in the figure (e.g. raw \beta instead of β)
   • Caption refers to a panel/element not visible in the figure
   • Duplicate or near-duplicate figures (compare against the previous figure)
   • Resolution problems (pixelation, font fallbacks)
4. Severity: high (blocks publication), medium (needs revision), low (cosmetic).