[srivtx]

Stacked on #20 (base ai-tools). Merge #20 first, then retarget this to main.

I made a poster with nothing but morpho and fixed each place I got stuck. The first round could tweak a design but not make one, couldn't show timing, didn't judge layout, and had no video export.

New commands

Command	What it does
new	A blank design. --canvas 1080x1350 or a preset name, --bg as a colour, gradient or preset, --duration.
add	text, shape, line or path, with --name --at --size --font --color --material or --preset, --recipe, and --above, --below or --into for placement. Also add modifier <kind> --to L [--fields], add group A B [--merge], add font f.ttf (embeds it and prints the id for --font), and add image p.png [--to L].
remove, duplicate, arrange	duplicate gives the copy new ids and moves its modifier keyframes to the new uids.
set Layer.preset="…", Layer.recipe=…	Presets and recipes in any assignment.
align	Lines layers up by what they draw, not their anchors: a wavy word that looks 6 px off centre gets centred by its ink.
timing	Entrances and exits as bars on the clock; loops and keyframes as their cycles, with hold keyframes shown as rest, go, at To, back. Notes on rhythm: entrances all starting at once, long stretches where nothing moves, moves cut off by the clock.
layout	Every layer's box, centre and margins as text.
video	MP4, MOV, WebM or GIF through ffmpeg, or PNG frames. Every frame is drawn at its exact time, so a 4 s loop exports in about 3 s and joins seamlessly.
fit	Tunes settings until the design matches a reference picture. On a test where the title was secretly resized to 265 and moved to y −190, it found 265.2 and −188.9.

New layout notes, in check and layout:

• edges and centres that nearly line up (0.2–1.2% of the canvas apart)
• layers that are nearly centred
• things crowding an edge
• a design sitting a little off centre
• no clear lead between headlines
• too many text sizes
• where the visual weight sits, and how many hues the palette uses

Fixes found by using it

• Changing a loop or entrance kind now also sets that kind's default amount. Before, switching the loop to spin kept the generic 16, which meant 16 turns.
• Motion pops are judged against the frame steps either side, so a fast but smooth entrance no longer reads as a jump. The loop seam is judged against the step before it, so an entrance that replays every loop is caught.
• Embedded fonts survive the reloads that images, align and fit do.
• Visual weight is measured from pixel centres. An off-by-half-pixel error had made centred things read as "upper".
• The layout notes ignore 1–2 px differences between text boxes (from glyph shapes) and single letters that make up a word (the "2" and "K" in Y2K).

The poster, made with the tools

docs/AGENT-TOOLS.md has the full sequence; it runs as written.

• The first check said the title was 10% off both edges and that the loop jumped, because the entrance replayed.
• After set (smaller title, fade out at 3.3 s) and align, check reported no problems and a smooth loop join.
• video then exported a 120-frame MP4 in 2.6 s.

Tests

• 12 new unit tests (126 in all).
• New end-to-end checks for layout, the palette, video frame batches and align (112 checks in all).
• npx tsc --noEmit -p ui, npx vitest run and npm run build && npm run test:e2e pass locally. I ran the e2e test with CHROMIUM_PATH set to Chrome.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated
morphotype	Ready	Preview	Oct 23, 2026 8:20pm UTC