Add OpenWebUI + LMStudio integration (CORS, docker compose, env-var host)

[offbyonebit]

Summary

• CORS middleware added to server.py — without it, OpenWebUI (running in a browser) gets blocked by same-origin policy when calling arc-llama from a different port
• created field in /v1/models now uses the GGUF file's mtime instead of 0 — model pickers in OpenWebUI sort and display timestamps correctly
• ARC_LLAMA_HOST / ARC_LLAMA_PORT env vars on arc-llama serve — the container needs to bind to 0.0.0.0, now possible without editing config.toml
• docker-compose.yml — new file; docker compose up brings up arc-llama + Open WebUI together, with GPU device passthrough and a shared network; chat UI at http://localhost:3000
• README — new Open WebUI and LMStudio sections covering both the compose path and bare-metal configuration

Rebase note (2026-07-10)

Rebased onto current main (9ebb2b8) and resolved the server.py import conflict (kept agent/UploadFile/FileResponse imports from main; added os + CORSMiddleware). Follow-up commit: compose healthcheck uses python3 urllib instead of curl (runtime image has no curl — depends_on: service_healthy would otherwise hang forever). Unit tests in tests/test_openwebui_integration.py.

Test plan

• docker compose up starts both services; Open WebUI at :3000 shows arc-llama models in the model picker
  • Not run here: host has Docker but no docker compose plugin; GPU image build not exercised.
• Bare-metal: arc-llama serve, configure Open WebUI with base URL http://127.0.0.1:11437/v1, models appear
  • Not run end-to-end against a real Open WebUI instance.
• GET /v1/models returns created as a non-zero Unix timestamp for registered GGUFs
  • Verified via unit tests + live uvicorn probe (created from GGUF mtime).
• Browser fetch to http://127.0.0.1:11437/v1/models from a different origin succeeds (CORS headers present)
  • Verified: OPTIONS preflight + GET with Origin: http://localhost:3000 return Access-Control-Allow-Origin: *.
• ARC_LLAMA_HOST=0.0.0.0 arc-llama serve binds to all interfaces
  • Verified: serve options bind envvars; live bind on 0.0.0.0:18765 (LISTEN 00000000:… in /proc/net/tcp).
• arc-llama upstream add lmstudio http://127.0.0.1:1234 merges LMStudio models into the model list
  • CLI add verified; /v1/models merges owned_by: upstream:lmstudio when upstream is configured (mocked LMStudio list).

https://claude.ai/code/session_016pwPb8iA8Ch62118zCNfyS

---

Generated by Claude Code; rebased/verified by Grok

[offbyonebit]

Rebased onto current main (9ebb2b8) and marked ready for review.

Conflicts: only server.py (imports). Kept main’s agent/UploadFile/FileResponse stack; added os + CORSMiddleware. CLI env vars, README, and compose applied cleanly.

Extra fix: compose healthcheck was curl, which is not in the runtime image — switched to a python3 urllib probe so depends_on: service_healthy can actually succeed.

Verified (automated + live local bind):

• CORS preflight + GET → Access-Control-Allow-Origin: *
• /v1/models created = GGUF mtime (non-zero)
• ARC_LLAMA_HOST / ARC_LLAMA_PORT bound on serve; live listen on 0.0.0.0
• upstream add lmstudio + model list merge (owned_by: upstream:lmstudio)
• unit tests: tests/test_openwebui_integration.py (6 passed)

Not verified on this host: full docker compose up (no compose plugin installed) and a real Open WebUI browser session against a GPU-backed arc-llama. Those two checklist items remain for a machine with compose + Arc.