AI Flow
Public Made by Adomby adom
Adom's AI Flow: a tool to help the AI follow all of the steps it takes to build a board.
main
8586980
18d ago
Native landing and component-library acceptance
These are maintenance acceptance tests, not a new ESC design benchmark. The original delivered run and clocks are unchanged.
One part and one trace at a time
Starting from the original unplaced ESC on an isolated copy, C1 and R1 landed as two native calls in 17.8 seconds. One GND plan entry containing two paths landed as two native calls in 12.4 seconds. Each was its own undo step and returned a distinct revision. Native readback confirmed both part positions and all six track segments. Preflight found zero new DRC errors and 13 inherited errors; the partially placed test intentionally retained 269 unconnected items. It cannot pass finish.
Placement recording, 10x | Routing recording, 10x | Native route replies
A first static take was rejected: the visible second editor did not own KiCad's single IPC server. After saving and gracefully closing only task-owned windows, opening the isolated editor first made it reachable. AI Flow 0.1.28 now checks this during capture setup and stops permanent identity errors immediately. The finished ESC was restored afterward.
Selected component variants through to the final cut
All 40 selected marked STEP hashes occur in the actual native board model bindings. KiCad's visible board was reviewed, including pin-one dots, J1 and lit LEDs. The STEP LED lens overlay is separated to avoid z-fighting; its GLB counterpart uses the existing emissive lens and light node. The library selection now includes the later MOSFET pin-one model rather than the old unmarked-corner version.
Short composed overview | Separate full component tour | Native board screenshot
The overview registered with board, clip, selection, model and native-evidence hashes, then composed into a four-second result. The full tour stays separate. Real MP4 playback was checked in Hydrogen webview: time advanced from 0 to 1.76 seconds, 51 decoded frames, no media error. Playback evidence. This verifies the video path and selected model bindings, not electrical qualification or every dimension of each component.
The tour originally stalled because a remote viewer module failed to import. AI Flow 0.1.29 serves the complete pinned canonical Adom 3D Viewer 0.7.1 with the tour and reports import failures visibly. It does not introduce a competing viewer.
# Native landing and component-library acceptance
These are maintenance acceptance tests, not a new ESC design benchmark. The original delivered run and clocks are unchanged.
## One part and one trace at a time
Starting from the original unplaced ESC on an isolated copy, C1 and R1 landed as two native calls in 17.8 seconds. One GND plan entry containing two paths landed as two native calls in 12.4 seconds. Each was its own undo step and returned a distinct revision. Native readback confirmed both part positions and all six track segments. Preflight found zero new DRC errors and 13 inherited errors; the partially placed test intentionally retained 269 unconnected items. It cannot pass finish.
[Placement recording, 10x](placement-10x.mp4) | [Routing recording, 10x](routing-10x.mp4) | [Native route replies](land-route-log.json)
A first static take was rejected: the visible second editor did not own KiCad's single IPC server. After saving and gracefully closing only task-owned windows, opening the isolated editor first made it reachable. AI Flow 0.1.28 now checks this during capture setup and stops permanent identity errors immediately. The finished ESC was restored afterward.
## Selected component variants through to the final cut
All 40 selected marked STEP hashes occur in the actual native board model bindings. KiCad's visible board was reviewed, including pin-one dots, J1 and lit LEDs. The STEP LED lens overlay is separated to avoid z-fighting; its GLB counterpart uses the existing emissive lens and light node. The library selection now includes the later MOSFET pin-one model rather than the old unmarked-corner version.
<video controls playsinline preload="metadata" src="library-overview.mp4"></video>
[Short composed overview](library-overview.mp4) | [Separate full component tour](library-detail.mp4) | [Native board screenshot](native-board.png)
The overview registered with board, clip, selection, model and native-evidence hashes, then composed into a four-second result. The full tour stays separate. Real MP4 playback was checked in Hydrogen webview: time advanced from 0 to 1.76 seconds, 51 decoded frames, no media error. [Playback evidence](library-playback-proof.json). This verifies the video path and selected model bindings, not electrical qualification or every dimension of each component.
The tour originally stalled because a remote viewer module failed to import. AI Flow 0.1.29 serves the complete pinned canonical Adom 3D Viewer 0.7.1 with the tour and reports import failures visibly. It does not introduce a competing viewer.