**Chat log** from Bead: Computer Vision Harness · [canonical source](https://redfish.acequia.io/guerin/.agents/b5221927-24ee-40f6-bb8b-94b3b9a88783/2026-06-29/chats/2026-06-29-computer-vision-harness.md) · session 2026-06-29 · discussion: Talk: Computer Vision Harness *Verbatim chat log: immutable provenance. The bead's copy is ground truth; wiki edits here are a fork, never the record.*
Bead: `b5221927-24ee-40f6-bb8b-94b3b9a88783` · opened 2026-06-29 · animator: Claude (CV) · parciante: guerin Verbatim user prompts in full; assistant turns summarized with pointers to the durable artifacts.
### Turn 1 — open the bead + the whole ask (verbatim) > start-bead computer-vision-harness > let's write an html harness for selecting different computer vision models using transformer.js > let's start with segment anything, depth and scene descrtipion, and yolo. also optical flow and brief feature tracking. > > allow selecting front and backfacing cameras and be responsive for laptop and mobile. sync bead at end and deploy in bead. use git devops **Assistant.** Read `.ai/START-HERE.md`, `.ai/beads.md`, `.ai/conventions.md`, and the `prepare-new-bead` skill to load the bead protocol + git/deploy conventions. Surveyed the deploy precedent (agentscript apps bead `b6fcda63`: git repo living in the deploy dir under `guerin/`, `.acequia-access.json` read:anonymous; camera-grid `plume.html` for house style). Generated GUID `b5221927-24ee-40f6-bb8b-94b3b9a88783`. Checked jsdelivr for the transformers.js version: latest is 4.2.0, latest stable 3.x is **3.8.1**. Pinned to `@3.8.1` (avoid the v4 major; all needed APIs present) after first drafting against 3.5.2. Built `repo/index.html` (v0.1.0) — the single-file harness. Six model pills in two families: - transformers.js on a captured still: **Segment** (SlimSAM, point-prompted), **Depth** (depth-anything-small), **Describe** (vit-gpt2 captioning), **YOLO** (yolos-tiny detection); - classical CV live per frame: **Optical flow** (Lucas-Kanade), **BRIEF tracks** (Shi-Tomasi + BRIEF + Hamming). Front/back camera via `facingMode` flip + `enumerateDevices` dropdown; front preview mirrored; responsive `100dvh` layout with scrollable pills. WebGPU/wasm backend toggle. Default mode optical flow so something runs the instant the camera is granted. Design rationale + architecture map captured in [`notes/00-cv-harness-orientation.md`](../notes/00-cv-harness-orientation.md). App + model table in `repo/README.md`. DevOps: git-init'd `repo/`, committed; deployed a copy to `guerin/apps/cv-harness/` with a provenance README + anonymous-read; MKCOL'd the bead root on redfish and synced both the bead and the deploy dir. Noted for the user: "YOLO" is served by YOLOS-tiny (canonical transformers.js detector), labelled as such; optical flow + BRIEF are classical, not transformers. App built and deployed; not yet field-tested on a phone.
*(append further turns below as the session continues)*