forked from 1b-agent/fieldcraft-upstream
jyvikoandClaude Opus 5.5
a423a9ef49
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NmAttHhUZTokgPeEJia8sm
55 lines
3.6 KiB
Markdown
55 lines
3.6 KiB
Markdown
# Demo video
|
||
|
||
`make.py` regenerates the README's demo, `docs/media/ibeam-analysis.mp4` and
|
||
`docs/media/ibeam-analysis.gif`: an agent analyses an I-beam cantilever over MCP while the
|
||
FieldCraft desktop app shows each step.
|
||
|
||
```sh
|
||
python3 tools/demo-video/make.py # build, run, encode into docs/media
|
||
python3 tools/demo-video/make.py --no-build # with the release binaries already built
|
||
python3 tools/demo-video/make.py --frames /tmp/f # also keep the captured pictures
|
||
cd tools/demo-video && python3 -m unittest # test the caption drawing
|
||
```
|
||
|
||
Needs Python 3 (standard library only), a display for the FieldCraft window (the pictures are drawn
|
||
offscreen, so a locked screen is fine) and `ffmpeg` with libx264 on the `PATH` (or `FFMPEG`).
|
||
It honours `CARGO_TARGET_DIR`.
|
||
|
||
What it does:
|
||
|
||
1. Builds `fieldcraft`, `fieldcraft-cli` and the testkit's `corpus` tool (release).
|
||
2. Writes the corpus cantilever I-beam (`tests/corpus/cantilever-i-beam`: 40 x 40 mm, 3 mm
|
||
flanges, a 2 mm web upright, 200 mm long along x) as a `.solidcraft` design to a temporary
|
||
folder with `corpus --part cantilever-i-beam --design`.
|
||
3. Starts FieldCraft (`fieldcraft`) with `--control PORT` and `fieldcraft-cli mcp --connect 127.0.0.1:PORT`, and
|
||
speaks MCP (JSON-RPC over stdio) to it. Every step is an MCP tool call: `open` the design,
|
||
`mesh` (`hex_sweep`, 4 mm), `batch` (steel, a solid section, the root clamp, a kinematic
|
||
coupling of the free end at x = 200 mm to a reference point loaded with 1000 N down, −z), `check_model`, `solve`
|
||
(`static`), and `show_results` for the von Mises contours on the undeformed shape
|
||
(`scale: 0`) and then the deformed shape, its `scale` swung (eased) from 0 to 150, back to 0
|
||
and out to 150 again. Every picture is the MCP `screenshot` tool with `source: window` (the
|
||
whole window drawn offscreen, `ui.capture`): the setup in the `iso` view, the results through
|
||
its `camera` argument (the iso orientation drawn back to 420 mm so the contour legend leaves
|
||
the beam clear), and to close, a slow orbit of 60 pictures, one `camera` angle each, from the
|
||
iso view half way round the beam. Each stage holds for 2 to 3 seconds.
|
||
4. Captions each stage with the MCP tool it calls and what it does (`open · tests/corpus/…`,
|
||
`mesh · hex sweep 4 mm`, `solve · static`, …) in a band under the window. `caption.py`
|
||
draws the bars in the standard library: a minimal TrueType reader (`cmap` format 4, `loca`,
|
||
simple and composite `glyf` outlines, `hmtx` advances), a scanline rasteriser (non-zero
|
||
winding, anti-aliased) and a PNG writer, with the interface's own font
|
||
(`crates/ui-egui/fonts/Inter-SemiBold.ttf`, SIL OFL 1.1). ffmpeg pads each window picture
|
||
with the band and lays the bar over it with its `overlay` filter.
|
||
5. Encodes the pictures at 10 frames per second, each held for its stage, into the MP4 (H.264,
|
||
1280 x 848) and the GIF (960 px, its own palette), and prints the tip deflection and the peak
|
||
von Mises stress.
|
||
|
||
The theme is FieldCraft's default (dark).
|
||
|
||
The script checks the solution against beam theory and stops when it is off. Tip deflection:
|
||
P L³ / 3 E I = 0.1429 mm in bending plus P L / G A_web = 0.0335 mm in shear, 0.1763 mm; the
|
||
solution gives 0.1764 mm (within 3 %). Root bending stress: P L c / I = 45.0 MPa
|
||
(I = 88 871 mm⁴); the solution's stress_xx on the top flange, fitted along the beam between one
|
||
depth from either end and extrapolated to the root, gives 44.8 MPa (within 2 %). At the root
|
||
itself the clamp, which stops the section's Poisson contraction, adds a local peak (62 MPa
|
||
stress_xx, 56 MPa von Mises) that beam theory does not have.
|