Sivia is a scientific-figure plugin for Codex and Claude Code. It reads your manuscript, finds relevant image-and-prompt examples, and develops a detailed ImageGen prompt around the method's actual contribution. You review the image first; editable PowerPoint, WPS or draw.io reconstruction is a separate, explicit choice.

Knowledge base

6 illustrated entries: 3 approved references and 3 review drafts. Every entry displays its own image below. Click an image for the full-size version, or open its complete production prompt and case notes.

Neuralangelo, ReAct and DiffDock are the approved historical references. D4RT, AutoTool and SigmaDock are recent-paper augmentation drafts with review notes. These are generated conceptual illustrations.

Browse the classified knowledge base

3D reconstruction and dynamic geometry

Neuralangelo Approved reference Full-size image · Full prompt · Case details

D4RT Review draft · not admitted Full-size image · Full prompt · Case details

Agents, tools and retrieval

ReAct Approved reference Full-size image · Full prompt · Case details

AutoTool Review draft · not admitted Full-size image · Full prompt · Case details

Molecular modeling and AI for Science

DiffDock Approved reference Full-size image · Full prompt · Case details

SigmaDock Review draft · not admitted Full-size image · Full prompt · Case details

Why Sivia?

• A scientific story before a layout. Ground the figure in the manuscript, equations and supplied implementation; identify what belongs in the overview and what belongs in the paper.
• Templates with their actual outputs. Learn composition and graphic language from paired images and complete prompts, without importing another method's scientific claims.
• Your approval controls the next stage. Generate → review → revise. Start editable reconstruction only when you ask for it.
• Edit the structure that matters. Text, modules, operators, arrows and reconstructable charts stay native; complex illustrations remain independent assets.
• Failures lead to a changed action. Recover from inadequate crops, false transparency, ineffective setters and stalled backends within the authorized scope.
• Evidence-backed delivery. Distinguish file validity, editability, exact-file preview, target-application rendering and final-size readability.

Installation

Codex

Run in a terminal:

codex plugin marketplace add exsinger-hub/Sivia --ref main
codex plugin add sivia@sivia

The first command registers the repository marketplace; the second installs Sivia. Start a new task after installation. See OpenAI's plugin packaging documentation; use codex plugin add --help to check your local CLI. Disable an older differently named copy if both expose the same skills.

Claude Code

Run inside Claude Code:

/plugin marketplace add exsinger-hub/Sivia
/plugin install sivia@sivia

Optional Star invitation

After installation, Sivia shows a one-time invitation at the first startup where its hook is enabled and trusted. If you want to support the project, say “Star Sivia”: the assistant can help star exsinger-hub/Sivia with your GitHub account after your explicit request. You can also click Star on GitHub yourself. Skipping the invitation does not affect any feature.

The startup hook only displays the invitation and stores a local notice marker; it never signs in or contacts GitHub. Set SIVIA_STAR_INVITE=0 to hide it. Codex requires hook review and trust before running plugin hooks; clients without enabled hooks can use the link directly. See the Star assistance rule.

What you need

Backend availability is detected at runtime: Windows PowerPoint COM, connected macOS PowerPoint Office.js, native OOXML working copies, and the draw.io adapter have different capabilities. An installed application does not prove a connected document or successful rendering. Image-only work does not open presentation software.

Quick start

1. Make a paper overview

Attach the paper. Code, real experiment assets, target figure width and style references are useful when available, not mandatory template-selection homework.

Use Sivia to design and generate a scientific method overview from my attached paper.
Ground the content in the manuscript, equations and supplied implementation. Explain the main
message and proposed regions first. Select a relevant knowledge-base image and its complete
prompt, then write and validate the full region-specific ImageGen prompt against that template.
Generate and inspect the image. Deliver the actual image, full prompt and short design notes.
Ask for my revisions and whether I want an editable PPT, then stop. Do not open PowerPoint/WPS.

Sivia expands this user request into the actual production prompt. Before generation, the exact submitted prompt must contain at least as many non-whitespace characters as its bound template. Detailed production instructions do not mean verbose labels inside the figure.

2. Revise only what you name

Use Sivia to revise the current visual draft according to my attached comments.
Keep the approved ratio, regions, palette and reading path outside the affected objects.
Integrate the changes into the complete prompt, validate its length, generate and inspect
the revised image, and retain the previous version. Show me the result and wait; no PPT yet.

Add your actual comments or annotated image. “This image looks good” approves the image; it does not authorize a deck.

3. Reconstruct the approved image

Use Sivia to faithfully reconstruct this explicitly approved visual draft as an editable PPTX.
Preserve its layout, wording and visible connections. Create a separate file; do not modify
other open documents. Rebuild text, geometry, diagrams and arrows as native objects; retain
only irreducible artwork as independent pictures. Keep work in the background.
If a backend fails, recover within this scope using an available isolated native-file route.
Inspect the actual saved PPTX and its rendered preview. Deliver the PPTX, preview and concise
notes about fonts, assets, editability and any pending target-application verification.

Specify WPS, Microsoft PowerPoint, draw.io, or a required live document when that choice matters. A constrained backend is not silently replaced. Real-data substitution is a separate adaptation request: a faithful-copy request preserves the approved visible content and identifies schematic values as schematic.

What stays editable?

No stock PPT icon? Build the geometry or choose an appropriate independent asset. Cropping alone does not blur pixels; enlarging a small crop exposes its limited sampling. SVG can scale cleanly without offering native subobject editing.

Inspect retained pixels and real alpha with the bundled read-only helper:

python skills/design-scientific-figure/scripts/inspect_raster_asset.py --image robot.png --width-mm 30 --require-transparent

Run from the plugin root. Add --crop-px LEFT TOP RIGHT BOTTOM to evaluate a crop. See asset production for source-resolution planning and authorized contour-mask recovery.

Limitations

• Image generation and native rendering depend on tools actually available in your session. Without ImageGen, Sivia can deliver a full prompt and explain what is missing; it will not skip approval and create a PPT instead.
• Faithful editable reconstruction preserves composition and semantics; it does not guarantee pixel-identical handwriting, textures or regenerated character designs. Low-resolution source art remains low-resolution.
• Generated illustrations are not experimental evidence. Missing measurements cannot be repaired by inventing plausible heatmaps or curves.
• A valid PPTX or an alternate-renderer preview is not PowerPoint/WPS verification. Unavailable target rendering stays pending, and dense source figures may need a separately approved layout change for smaller publication widths.
• Fonts may require installation. Editing notes identify approximations, native/raster boundaries and any non-attached routes.

Development

From the repository root, with Node.js, Python 3/Pillow and PowerShell 7 available:

node scripts/sync-plugin-metadata.mjs --check
node --test tests/*.test.mjs
python -m unittest discover -s tests -p "test_*.py"

These tests exercise packaging, focus policy, payload behavior, arrow mapping, prompts, raster inspection and review-copy preparation. They do not replace live-application or visual tests. See v1.1.1 changes.

Acknowledgments

Built on Scientific Illustrator by icebird1998（一个地质博士）, retaining its MIT license and copyright. Sivia extends its editable drawing backends and Designer–Drawer–Reviewer–Corrector process with manuscript understanding, template-led ImageGen drafts and approval-locked reconstruction.

The bilingual README navigation and product-oriented organization take inspiration from CC Switch; its product claims and promotional content are not reused.

Thank you for using Sivia. Created by gatina.