MechanismFigures v0.4.0 — stronger references and global installation

MechanismFigures v0.4.0

Updated gallery · Start here · Merged pull request

Eight more tangible published constructions

Replaced the eight user-retired entries with Mechanical Characters, Computational Caustics, reconfigurable photonic braids, RFdiffusion, Biomimetic 4D Printing, Diffusive Mechanical Machines, Generative Omnimatte, and Scutoids. The twelve retained references keep their image bytes unchanged. Twenty references and twenty-nine real published image assets remain, with no generated substitutes. Paired physical examples accompany the mechanical and optical construction diagrams.

Install once across supported local projects

git clone --branch v0.4.0 https://github.com/KumarNavish/MechanismFigures.git
cd MechanismFigures
python3 install.py --global --agents all
python3 install.py --doctor

For an existing managed installation: python3 install.py --global --agents all --update.

One canonical user skill, documented routes for eight local agent hosts, verified adapters, dry-run conflict detection, previous-version preservation, and journaled recovery. Global means the current user/machine across projects—not automatic installation into every web/mobile account, remote worker, or every message. No agents or paid model APIs are invoked and no unrelated global instructions are rewritten.

Clear agent navigation and distribution

START_HERE.md, separate install/host/maintainer documentation, llms.txt, a progressive skill index, explicit artifact contracts, and a portable plugin manifest/local marketplace package. The standalone skill ZIP and plugin ZIP are separate. GitHub publication and an installable plugin package do not assert universal directory listing or account-wide activation.

Evidence

110 regression tests pass locally and in GitHub Actions on Python 3.9/3.12. The 15 gallery and 30 motion browser checks pass locally; all twenty-nine source images decode. Both archives pass CRC and file-manifest hash verification. The actual local user-global update and doctor succeeded for the documented file routes; runtime activation in each agent was not tested by launching model runners.

The scoring rubric remains review policy, not a claim of universal scientific/aesthetic quality or cross-agent efficacy. All figures retain their original source-specific rights; the code's MIT license does not relicense them.

MechanismFigures v0.3.0 — the motion walkthrough

See how MechanismFigures works

Watch the 60-second walkthrough · Complete gallery

An authored motion sequence follows question → real reference inspection → visual construction → critique → evidence-linked delivery → the range of scientific constructions the skill teaches.

All scientific imagery comes from the unchanged approved reference canon. Published references remain attributed to their authors and are not presented as skill-generated outputs. The original 20 references and 29 image assets are preserved.

Controls and accessibility

Play/pause, replay, continuous seeking, six chapter controls, selectable final references, and direct source inspection. Reduced-motion users get instant manual chapters and a complete text transcript. Playback stops after one pass and pauses when hidden, offscreen, or inspecting a source. No model call, fake live-agent progress, audio autoplay, external script or third-party image hotlink.

Verification

83 local regression tests; 30 motion browser checks; all 15 existing gallery checks, including all 29 actual images. Desktop and mobile renders, real playback, pausing, seeking, source inspection and reduced-motion first load were checked. These verify implementation, not cross-agent scientific/aesthetic efficacy.

The attached ZIP includes the motion walkthrough, complete approved gallery, skill, provenance and helpers. Third-party figure rights remain separate from the code license.

MechanismFigures v0.2.0 — approved real-image canon

The complete approved published-image gallery

Restores all 20 approved references and all 29 real figure assets. The original six anchors—AlphaFold, CellRank, GraphCast, DreamFusion, AlphaDev, and Aardvark—appear first. Every reference displays its actual image; no text-only substitute cards or self-generated calibration examples remain in the current tree.

The installed skill now requires two actual reference-image inspections with registered paths and SHA-256 hashes, specific composition/encoding/style observations, a project-specific transfer plan, and final composition/encoding/style comparisons against the same references. A single viewed image, a self-generated reference, missing hashes, or missing comparisons cannot satisfy the recorded gate.

Self-generated demonstrations were removed from the current public documentation and source tree. Existing v0.1.0 history is preserved; it is not the active visual standard.

The gallery renders all primary images without JavaScript and includes full/focused/context inspection, native-size viewing, search and mechanism filters. All images are served with the package; no third-party image hotlinks are required.

Verification

73 local regression tests pass. The local reader passed 15 browser checks, including every one of the 29 image assets, desktop/mobile layout, search, filters, viewer behavior and offline asset loading. These are engineering and recorded-workflow checks, not a claim of universally guaranteed figure quality.

Rights

Original code/commentary is MIT licensed. Published images retain their original author/publisher rights; image inclusion is not a general license or endorsement. Each source, figure treatment and recorded permission status is preserved in THIRD_PARTY.md.

The ZIP includes the complete installed skill and all image files. Existing modified installations are preserved by the installer; review an update in a separate skills root before replacing locally customized files.

MechanismFigures v0.1.0

MechanismFigures v0.1.0

An installable Agent Skills-format workflow for mechanism-revealing scientific figures, not a plotting theme or autonomous model.

Install

git clone https://github.com/KumarNavish/MechanismFigures.git
python3 MechanismFigures/tools/install.py

Alternatively unpack the skill ZIP into your host's supported skills directory. The package contains mechanism-figures/SKILL.md, helpers, templates, calibration recipes and licensed images. Python helpers use only the Python 3.9+ standard library. No model API, paid service, browser installation, or telemetry is invoked.

Included

• A progressive-disclosure, mechanism-to-figure execution workflow with persistent contracts and revision records.
• Twenty publication-derived calibration recipes; twelve real figure assets from seven references with verified CC BY 4.0 reuse permission. Other figures are linked to their originals, not silently relicensed.
• A ten-dimensional anchored rubric: weighted score at least 90/100, all dimensions at least 4/5, fidelity 5/5, all ten hard gates passed.
• Hash-bound evidence/source/output reviews; separate self-review and independent-review statuses.
• Two original analytical demonstrations with actual first-draft inspection and repair histories.
• Eight diagnostic four-input task packets, matched evaluation protocol and descriptive benchmark reporter.

Verification

61 engineering regression tests pass locally. The same suite, repository integrity audit, reproducible gallery build, and clean installation checks pass on GitHub Actions with Python 3.9 and 3.12. Ten desktop/mobile/offline browser checks pass for the calibration reader.

These tests verify tooling and recorded review policy, not general scientific/aesthetic quality. Example reviews are creator self-reviews. No cross-agent efficacy study, independently established reference-quality pass rate, or universal publication-quality guarantee is claimed.

The ZIP and its SHA-256 file are attached. Original code/editorial content is MIT licensed; third-party images retain the separate licenses and credits in THIRD_PARTY.md.