go: /v1/chat/completions returns 400 for reasoning_effort=max|minimal on mimo-v2.6-flash (low/medium/high succeed) #50818

Summary

POST https://opencode.ai/zen/go/v1/chat/completions rejects reasoning_effort=max (and minimal) for mimo-v2.6-flash with HTTP 400, while low, medium and high all succeed. The gateway surfaces the failure as:

{"error":{"param":"","type":"server_error","message":"Upstream request failed: [400] Invalid request parameters"}}

The Upstream request failed: [400] wrapper suggests the gateway forwards the effort value to its upstream unchanged and the upstream rejects the enum value.

This matters because Xiaomi's official channel does support a max thinking level for mimo-v2.6-flash (their Anthropic-style endpoint accepts it), so clients that expose a max level for this model work against the official endpoint but break whenever the request is routed through zen/go.

Reproduction

curl https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer $OPENCODE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "x-opencode-session: 3f1c9a2e-7b4d-4c8a-9e2f-1a2b3c4d5e6f" \
  -d '{"model":"mimo-v2.6-flash","messages":[{"role":"user","content":"say ok"}],"max_tokens":16,"reasoning_effort":"max"}'

Results (tested 2026-09-23)

Evidence this is model/upstream-specific, not gateway-wide

• On the same gateway with the same session header: deepseek-v4.1-flash with reasoning: {"effort": "max"} via POST /v1/responses returns 200 — the gateway clearly supports max for models whose upstream accepts it.
• Xiaomi's official endpoint accepts a max thinking level for the very same model (mimo-v2.6-flash).

So for mimo on zen/go, the accepted effort set is effectively {low, medium, high} (plus on/off), and anything else becomes a raw upstream 400.

Suggested fix

1. Clamp effort values the upstream does not accept to the nearest supported one (max → high, minimal → low) instead of surfacing a raw upstream 400 to the client; and/or
2. Advertise the supported effort values per model (e.g. in /v1/models metadata) so clients can render accurate controls instead of discovering the failure at request time.

Related

• Console Go relay: HTTP 400 (empty body) when request combines reasoning_content echo-back with reasoning_effort on deepseek-v4-flash #48180 — Console Go 400 involving reasoning_effort on deepseek (different root cause, same endpoint family)
• Opencode Zen Deepseek V4 Flash Free (new) reasoning_effort produces wrong behaviour #40777 — reasoning_effort misbehaviour on Zen
• HTTP 400 error trying to use mimo-v2.5 #45990 — HTTP 400 trying to use mimo-v2.5