[jasp402]

What

One new on-demand tool in the existing connector (no second connector or endpoint; the 4 publishing tools are untouched): get_motion_recipe(name?), plus a self-review step in validate_json.

• get_motion_recipe() returns an index (~240 tokens). With a name it returns a recipe built from a composition fragment that is the same JSON shown to the model: riso, choreography, audio, camera, code-texture, ui-mockup, scene-bg, looks (13 looks in one table) and look:<id> ×13 (ported from Studio lookSample, contract c75a0b3). Plain-text result, no outputSchema. Useful content arrives in tool results, not in initialize.instructions, so it works with the connector alone on claude.ai Free.
• Validator: kf-text TEXT_DOES_NOT_FIT false positive fixed (measures every value); TEXT_OVERLAP; UNKNOWN_EASE (kf ease, kf spring mode, animEase); TEXT_FIT_UNEVALUATED for expression text.
• A short checklist is added to validate_json / create_hcmotion results only when the composition is valid / created (never next to errors).
• get_capabilities START has one line pointing to get_motion_recipe; SKILL.md has one added line (header unchanged).

Tokens (tiktoken cl100k, an estimate; Claude's tokenizer counts a bit higher)

• Fixed load (tool descriptions + schemas + initialize.instructions + SKILL.md header): 3706 → 3844 (+138, ceiling +150).
• On demand: index 239; recipes 537–1191 each (audio 537, camera 595, scene-bg 602, code-texture 641, choreography 712, ui-mockup 807, looks 836, riso 1191); look: 638–945; checklist 134 per valid result. Most recipes exceed the 150–600 target: they carry real JSON/HTML, and I did not cut tested content to hit a number.

Validation

• tests/recipes.test.js: every recipe fragment passes validateJson (0 errors, 0 warnings other than CONTRACT_NOT_CHECKED) and the text contains the tested JSON. npm test: 122 pass, 0 fail.
• Rendered in the Studio (dev server): 20 compositions, no page errors. Studio-side checkComposition clean on all; normalizeAudio on the audio recipe: 3 tracks and 4 sfx kept, 0 warnings.
• Gallery sweep (39 templates): 0 TEXT_OVERLAP, 0 UNKNOWN_EASE.

Not verified / risks

• Audio was checked structurally only (nobody listened).
• TEXT_OVERLAP does not measure flow/flex/CSS-class html pieces: it missed the overlaps in the A experiment.
• src/looks.js is a copy of the Studio looks; it needs a refresh when they change.
• Real claude.ai behaviour was not tested.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
hypercodemotion-mcp-service	Ready	Preview	Oct 25, 2026 4:34am UTC