Skills-first workflow redesign — proposal + code, with 4 fixes already PR'd (#45–48)

[nobullryder]

Hi maintainers — I've built a sizeable, "skills-first" workflow redesign of LeLab on top of main, and I'd like to hand it over so you can take what's useful and ignore what isn't. I know a big drop is a lot to ask of a review, so I've already split out the small, self-contained, IA-independent pieces as normal PRs, and I'm posting this issue for the larger redesign rather than dumping it on you.

Already opened as focused PRs (each stands alone, off main)

• Prompt to install policy extras (smolvla/pi0/diffusion) before training starts #45 — Policy-extra preflight. Install lerobot[smolvla|pi|diffusion] before training starts (instead of a buried ImportError mid-run). Builds on your existing install-extra pattern.
• Reliable shutdown on Windows + lelab --stop #46 — Reliable Windows shutdown + lelab --stop. Process-tree teardown so uvicorn/Vite children and camera handles are released; pre-flight port checks.
• Force-release camera/serial resources when a device disconnect fails #47 — Force-release camera/serial on disconnect failure. Stops a flaky camera/port staying held and blocking the next run.
• Surface the real inference error; don't report a gripper-overload shutdown as failed #48 — Surface the real inference error. Pull the actual exception from the rollout log; classify a gripper-overload-on-shutdown as ran_with_warning, not failed.

Each is small, tested, and reviewable independently of everything below.

The larger redesign — a visual tour

Full branch: nobullryder:redesign/machined-console → https://github.com/nobullryder/leLab/tree/redesign/machined-console
(diff vs main: ~118 files, ~11k insertions, ~3.3k deletions). Screenshots below are from a live local instance with real data.

Run the whole thing locally (the built UI is committed, so there's no npm step):

git clone -b redesign/machined-console https://github.com/nobullryder/leLab.git
cd leLab && pip install -e .
lelab            # serves the UI + API on http://localhost:8000

Browsing datasets / training / the UI needs no hardware; teleoperate / record / inference need an SO-101.

Home — a skills-first journey

Record → teleoperate → train → run, in plain words with the robotics terms glossed.

Robot — pick a robot, set ports, calibrate

Datasets — a library of your recordings

Pretty names + source badges (local / Hub / both), grid or list view.

Dataset detail + episode player

Per-episode listing, sync status, add-takes, publish, per-episode delete — and a custom player that clips playback to each episode (not the whole multi-episode video file).

Train — config + live job monitoring

Skills — trained policies with version history

Models grouped by their source dataset into versions: retrain after editing a dataset and the old version stays here to run. (Shown: one dataset trained four times — two ACT, one failed SmolVLA, one training live.)

Chat — a skill-aware assistant

Ask in plain words; it picks the right trained skill and (with your confirmation) runs it. Backed by an OpenAI-compatible API or a local claude / codex CLI.

Suggested review order (smallest / safest first)

1. Policy-extra preflight — done, Prompt to install policy extras (smolvla/pi0/diffusion) before training starts #45.
2. Runtime / platform reliability — done, Reliable shutdown on Windows + lelab --stop #46 / Force-release camera/serial resources when a device disconnect fails #47 / Surface the real inference error; don't report a gripper-overload shutdown as failed #48.
3. App shell + routes + design tokens — the product-direction piece; review on its own.
4. Recording / camera workflow — take/redo loop, resume fixes, camera config, recording prefs.
5. Dataset management — ⚠️ see flag below.
6. Training + Skills — Skills page, version history, training/monitoring polish.
7. Chat surface — ⚠️ see flag below.

I'm happy to open 3–7 as a stacked PR series of focused diffs if you want to review them that way — just say the word.

Please review these in isolation

• Dataset management (5) includes per-episode delete that rewrites the local dataset's parquet/video/metadata files and re-indexes episodes. It's behind an explicit confirm and has tests on synthetic datasets, but it's the riskiest piece — review it on its own.
• Chat surface (7) shells out to locally-installed claude / codex CLIs (or an OpenAI-compatible API) and produces assistant-driven actions; anything touching hardware is confirmed by the user. It's optional/local, not required infrastructure.
• This is a wholesale visual + IA redesign. If you don't want that direction, the reliability fixes (Prompt to install policy extras (smolvla/pi0/diffusion) before training starts #45–48) stand entirely on their own.

Testing

Backend pytest green; npm run build + tsc + eslint clean; core flows dogfooded (record/resume, episode delete on real + synthetic datasets, Skills run/delete, inference launch). The redesign is opinionated — treat the visuals as a proposal, not a fait accompli.