weft-ui

Reference UI implementation for driving weft — the execution substrate for agent-driven scientific analysis. Three surfaces over one local server that embeds a single Weft controller per workspace:

1. Chat — a generic agent panel (Claude Agent SDK) whose weft-specific tool renderers show plans, structured errors, and event digests. (M3)
2. Compute — site setup wizard (probe-first, ~/.ssh/config-aware) and live capability/load/footprint status per site. (M2)
3. Jobs — monitor/controller for tasks, arrays, kernels, and services: live logs, structured failures with one-click remediation, provenance. (live)

Design principle (from weft): the user and the agent are peers — every button is an API call the agent could also make (each action shows its ⌁ tool_name), and both land in the same audit trail.

_{Compute (site capabilities, live storage) and Jobs (structured OOM failure with one-click remediation) — click for full resolution.}

Quick start

pixi install && pixi run web-install && pixi run web-build

# serve a workspace (mints a token, prints the URL to open)
pixi run serve                      # demo/workspace
# or: pixi shell; weft-ui serve --workspace /path/to/project

# populate the demo story (DFT survey: OOM pilot, qc array with failure
# buckets, memoized resubmit, live-log stream) — drives the running
# server over its own HTTP facade:
pixi run serve &   # with --token demo (see pixi.toml) …
pixi run seed

Frontend dev loop: pixi run web-dev (vite on :5173, proxies /api to :8999; open http://localhost:5173/?token=…).

Architecture (short form; misc/plan.md has the full plan)

• server/weft_ui/ — FastAPI, single process, single asyncio loop.
  • facade.py — POST /api/w/<tool> for every tool in weft's PUBLIC_TOOLS, generated by introspection; tool returns pass through verbatim (weft's returns-never-raises contract crosses HTTP intact: errors are payloads inside a 200; HTTP codes are transport only).
  • events.py — weft's store bus bridged to one SSE stream with cursor replay; stale cursors and slow clients get a _resync control event. The UI never polls.
  • uiapi.py — list endpoints not yet in PUBLIC_TOOLS + the live log sub-stream (task_logs cursor-poll at 1 s, labeled honestly).
  • lock.py — flock on .weft/ui.lock: one controller per workspace.
• web/ — Vite + React + TS; design tokens in web/src/tokens.css, layout in web/src/app.css.
• shared/types.ts — hand-mirrored weft payload shapes; server/tests/test_conformance.py captures real payloads into shared/samples/ and fails loudly when upstream weft reshapes them (shared/samples/BASELINE records the weft SHA; refresh with UPDATE_SAMPLES=1 pixi run test).

Security: binds 127.0.0.1 only; per-process bearer token (injected into the served page); Origin checked on every /api request.

Embedding & programmatic use

Every screen is deep-linkable (#/jobs/kernels/krn_x, #/compute/clip, …) and any of them can be iframed chrome-less into another app with ?embed=1 — framing is CSP-gated by an embed_origins allowlist in the workspace config. Python/ASGI host apps can instead mount weft-ui in-process for same-origin panels. The same HTTP surface the UI uses (tool facade, SSE event stream, chat) is callable directly. See docs/embedding.md.

pixi run test · pixi run lint; working notes in misc/ (untracked).