Independent confirmation from Windows, plus one theory ruled out.
Setup: T3 0.1.8-nightly.20261104.4165, OpenCode 2.2.5 (@opencode/cli) after an in-place opencode upgrade --major from 1.19.5. Same banner: "OpenCode 2 is running, but it did not list any models yet.", empty picker, old threads failing to resume (those were fixed by the T3 update itself, #27779 — this comment is only about the empty catalog).
Ruled out: OPENCODE_CONFIG_CONTENT="{}" is NOT the cause. I suspected T3's spawned-server override (it resolves to "{}" when no env is configured) because my first probe of a freshly spawned server returned empty/HTML for /api/model. Controlled retests disprove it: servers started with the exact same OPENCODE_CONFIG_CONTENT="{}" env T3 passes return the full catalog (/api/model → 43 KB JSON, ~40 models across opencode + opencode-go, MCP servers connected) from ~3 s after boot, stable across 110 s of polling. The one empty observation was just the first-boot window. Old issue #4239's ghost can stay buried — no need to chase the override.
What the first-boot window looked like here: immediately after the major upgrade, API routes served the SPA index.html / empty 200 application/json bodies for the first seconds — consistent with this issue's ServiceUnavailableError: Location plugins did not finish activating mechanism, except the likely bottleneck on my box was the v2 first-run conversion of a 5.2 GB v1 opencode.db (sidetrack: that file is enormous — mostly tool-output, might be worth a look on its own). So: major-upgrade first boot plausibly stretches the cold window far beyond the 5 s discovery budget, the empty result gets cached as a warning, and the picker stays empty even though the server is healthy seconds later.
Supporting the suggested fix direction (bounded retries on the same borrower / distinguishing "loading" from "completed empty"). Happy to pull T3-side traces (server.trace.ndjson shows the checkOpenCode2 empty-list path clearly) if useful.