V2 /api/generate returns 503 while standalone succeeds

Summary

The experimental V2 sessionless POST /api/generate operation returns HTTP 503 for models that complete through opencode2 run --standalone in the same isolated environment.

Versions tested

• opencode2 v0.0.0-beta-18387
• opencode2 v0.0.0-beta-19242
• Linux x86_64

Minimal reproduction

Start the app-owned server with an isolated XDG home/config/data/state/database and autoupdate disabled:

export OPENCODE_PASSWORD="$(openssl rand -hex 32)"
opencode2 serve --stdio --hostname 127.0.0.1 --port 0 --log-level none

Using the announced loopback URL, submit a harmless canary with Basic authentication:

curl --user "opencode:$OPENCODE_PASSWORD" \
  --header 'content-type: application/json' \
  --data '{"prompt":"Return one short greeting.","model":{"providerID":"opencode","id":"nemotron-3-ultra-free"}}' \
  "$OPENCODE_URL/api/generate"

Observed: HTTP 503. On beta-18387 the same occurred with opencode/big-pickle. Beta-19242 required its current base model/plugin initialization before reaching the same 503 result.

Control using the same harmless canary and isolated environment:

printf '%s' 'Return one short greeting.' |
  opencode2 run --standalone \
    --model opencode/nemotron-3-ultra-free \
    --format json \
    --log-level none

Observed: generation completes successfully on both tested builds.

Expected

POST /api/generate should generate successfully when the same explicit model and environment succeed through run --standalone, or return a stable diagnostic that identifies the missing initialization/contract requirement.

No credentials, prompts beyond the harmless canary, response bodies, or user content are included in this report.