3. Switch the OpenCode model / reasoning effort (config only)

Historical since 2026-09-07: every OpenCode worker is in RETIRED_WORKERS (ks4sf), so these knobs change nothing until an entry is deleted there. Kept because the adapter stays as the CLI-worker reference.

# Cheaper non-think model:
OPENCODE_MODEL=deepseek/deepseek-v4-flash npm run demo
# Max reasoning (DeepSeek Think Max - billed separately):
OPENCODE_VARIANT=max npm run demo
# A different provider OpenCode is authed for:
OPENCODE_MODEL=moonshotai/kimi-k2.7-code npm run demo
# Xiaomi MiMo through OpenCode (once `opencode providers login` has authed `xiaomi`):
OPENCODE_MODEL=xiaomi/mimo-v2.5-pro npm run demo

Canonical ids only: deepseek/deepseek-v4-pro (default), deepseek/deepseek-v4-flash, moonshotai/kimi-k2.7-code, xiaomi/mimo-v2.5-pro. deepseek-chat / deepseek-reasoner were removed from DeepSeek's API on 2026-07-24 (confirmed absent from GET /models 2026-08-24) — they are dead ids, not merely deprecated aliases, and every call using one fails at the provider.

To reach Xiaomi MiMo as a direct loop worker instead (no OpenCode), put XIAOMI_API_KEY in .env (or ~/.cyberine/.env): loop-mimo-direct then joins the compare fleet automatically, and loop-mimo-build becomes available for CYBERINE_BUILD_FLEET — e.g. CYBERINE_BUILD_FLEET="loop-deepseek-build,loop-mimo-build" npm run wf -- <def>.