Files
Ray XiandGitHub 784025c04a fix(daemon): resolve the AMR default model from the caller's catalog (#8519)
* fix(daemon): resolve the AMR default model from the caller's catalog

A run that asked for `default` before the caller's `vela model list` had
arrived was concretized from the bundled `vela model preset` seed. The
seed is the same for every plan and OD orders deepseek-v4-flash first, so
free-tier callers (entitled only to deepseek-v4.1-flash since 09-24) had
Vela reject `session/set_model` before any work began. On 0.24.1 that was
22 runs from 21 users in the last 7 days, 13 of 16 bundled runs showing
the user's choice as `default`.

- When `default` is requested and only the preset is cached, wait up to
  12s for the caller's remote catalog (`getAuthoritative`).
- Remember remote catalogs separately; a later preset seed never displaces
  one when a default model is chosen.
- Order deepseek-v4.1-flash first: it is the one chat model every plan,
  including free, is entitled to, so a flagless list still lands on it.

* fix(daemon): bound the AMR default-model catalog wait per refresh

CI follow-up. A `default` run looked the catalog up twice (the resume probe
and the run itself) and each lookup waited up to 12 s for the caller's
remote catalog, so a slow or broken `vela model list` held the run queued
for up to 24 s. The AMR e2e runs timed out at 20 s that way.

- The wait is counted from the start of the remote refresh, so one run, or
  several concurrent runs, share a single wait.
- When the last refresh failed, the preset is used at once; the next
  refresh still starts in the background.
- e2e fake vela: answer `model list --all --format json`, the form the
  daemon actually sends. It only matched the flag-less form and hung.

* test(e2e): let the fake vela answer the daemon's `model list --all` probe

The catalog probe spawns `vela model list --all --format json`. The matcher
stripped `--all` with its preceding space only, so `' --format json'` never
equalled `'--format json'`: the probe fell through to the login check, the
cache recorded a failed refresh, and default runs settled on the preset. Match
the subcommand and the `--format json` flags in any order, as fake-vela.mjs and
turn.test.ts already do, and cover the exact daemon argv for both the catalog
and the 401 path.
2026-10-15 03:57:51 +00:00
..