S09 · Phone away from home

Situation. I’m away from home, Tailscale is on, and I open Open WebUI on my Android phone.

What happens. It’s served over HTTPS through tailscale serve and installs as an app. Image questions go to the always-loaded vision model; voice input goes to speech-to-text. (Corrected 2026-09-28, from Phase 1’s council: in Phase 1 the vision model is resident, not preloaded. After a restart it loads on its first request, then stays loaded; Phase 2 preloads it. That it installs as an app is not yet recorded as checked.)

What I see. The real model name on every reply. (Not yet recorded as checked on the phone, 2026-09-28.)

How to override. None needed. A chat can run to the model’s full context, 262,144 tokens on Gemma, and a photo costs up to about 1,120 of them. (2026-10-08: the registry’s coder, now Qwen3.8-27B, is held at 163,840 of its 262,144 tokens until Phase 2a’s Task 42, Dan’s decision.) (Changed the same day, Dan’s decision: with render’s corrected budget check on main, the registry gives the coder all 262,144 tokens; the Spark serves them once that is deployed.) The phone can’t change the context: Open WebUI’s num_ctx parameter is for Ollama, and v0.11.4 never sends it to llama-server. What it can change, per chat under Controls → Advanced Params or as a model’s default, includes max_tokens and Reasoning Effort. max_tokens counts the thinking too, so set too low it leaves the answer empty. llama-server b11146 hands Reasoning Effort to the model’s chat template, and none turns thinking off (its source). In Task 13’s one probe each, the coder (then Qwen3.6-35B-A3B; Qwen3.8-27B since 2026-10-08, not yet probed) followed low and high, and Gemma barely did. Open WebUI can also compact a long chat itself, summarizing its older messages once it passes a token threshold, 80,000 by default. That is off unless the admin turns it on, under Admin Panel → Settings → Interface → Context Compaction, and then a chat’s Context Compaction Threshold can lower the threshold. Dan leaves it off (2026-09-28): at 262,144 tokens a chat rarely needs it, and a chat that ever outgrows the context is refused with an error, so a new chat is the way on. (Added 2026-09-28, Dan’s decision: every model at its full context. The labels are read from Open WebUI v0.11.4’s source, and Dan found each of them in the web UI in the Mac’s browser the same day; not yet checked on the phone.)

Status: verified, 2026-09-28 (Phase 1, Task 14): on the phone, at home and on mobile data, the web UI answered chats, a photo question, dictation and voice mode. The settings under How to override* were checked in the web UI in the Mac’s browser the same day (Dan’s decision); the claims marked above as not yet checked stay marked.*