Qwen3.8 Flash: variant=xhigh silently disables reasoning by sending output_config.effort without thinking

Description

On opencode-go/qwen3.8-flash, selecting the reasoning variant xhigh makes OpenCode send a bare output_config.effort to POST /zen/go/v1/messages with no thinking block. The request succeeds (HTTP 200) and no error is shown, but reasoning is never enabled: tokens.reasoning is always 0 for these sessions.

So variant=xhigh looks active but does nothing. This is deterministic, not intermittent.

Wire symptom:

variant=xhigh
  -> POST /messages
  -> output_config: {"effort":"xhigh"}
  -> no "thinking" field
  -> HTTP 200, no thinking block
  -> tokens.reasoning = 0

OpenCode version

1.18.31 (Linux). Relevant UA: opencode/1.18.31 ai-sdk/provider-utils/4.0.46 runtime/bun/1.3.14. Bundled @ai-sdk/anthropic is 3.0.111.

Steps to reproduce

opencode run --model opencode-go/qwen3.8-flash --variant xhigh "Reply with exactly: OK"

Check the session usage:

DB="$HOME/.local/share/opencode/opencode.db"
sqlite3 "$DB" "
  SELECT json_extract(model,'\$.variant'), tokens_reasoning, tokens_output
  FROM session
  WHERE model LIKE '%qwen3.8-flash%' AND model LIKE '%xhigh%'
  ORDER BY time_created DESC LIMIT 5;"

Result: tokens_reasoning = 0 on every row.

To see the body OpenCode actually sends, point the provider at a local recording endpoint (provider.opencode-go.options.baseURL) and run the same command. Captured body for the xhigh request (message/system content redacted):

{
  "model": "qwen3.8-flash",
  "stream": true,
  "max_tokens": 32000,
  "output_config": { "effort": "xhigh" },
  "tool_choice": { "type": "auto" },
  "tools": ["<14 tools>"]
}

There is no thinking key. (For comparison, the auxiliary/title request is sent the same way with effort: "low".)

Expected behavior

variant=xhigh should enable the model's reasoning mode and produce reasoning tokens. If the requested mode cannot be honoured, OpenCode should warn instead of silently sending a no-op parameter.

Root cause (source)

packages/opencode/src/provider/transform.ts:

reasoningVariants()  -> model has { type:"effort", values:["low","medium","xhigh"] }
  -> effortVariants() -> reasoningEffort(model, "xhigh")
       case "@ai-sdk/anthropic":
         return anthropicEffort(model, effort) ?? { effort }   // fallback taken
           anthropicEffort(): anthropicAdaptiveEfforts("qwen3.8-flash") -> null  // needs claude-/kimi
                              -> returns undefined
       => { effort: "xhigh" }

@ai-sdk/anthropic 3.0.111 then faithfully serialises effort into output_config.effort; since no thinking was provided, none is sent. The catalog defines qwen3.8-flash with a model-level provider.npm: "@ai-sdk/anthropic" (provider default is @ai-sdk/openai-compatible).

Evidence

Local opencode.db, qwen3.8-flash on opencode-go:

variant	sessions	reasoning tokens
xhigh	42	0
default	16	0
high	3	0
max	2	0

42/42 xhigh sessions (100%) have reasoning = 0; message-level cross-check: 1,656 messages, 0 with reasoning.

Controls on the same provider:

model	reasoning tokens
qwen3.7-plus	29,577
deepseek-v4.1-flash	242,466
qwen3.8-flash	0

So the provider/adapter path does produce reasoning for other models; only the qwen3.8-flash effort-variant path yields zero.

Related (separate) compatibility note

On the same endpoint for this model, thinking: {"type":"adaptive"} returns HTTP 200 with a thinking block, while thinking: {"type":"enabled","budget_tokens":N} is rejected (HTTP 500 before ~2026-09-14 23:13 UTC, later HTTP 400 {"model":"qwen3.8-flash"}). The enabled form is accepted for qwen3.7-plus and minimax-m3. This looks like a provider/model compatibility limitation, and it means replacing the bare effort with enabled + budget_tokens is not safe — adaptive is what this model accepts.

Proposed fix

Guard the Anthropic effort mapping for Qwen models so the variant produces a real thinking config, e.g. in anthropicEffort():

if (isQwenAnthropicFamily(model)) // family === "qwen" on @ai-sdk/anthropic
  return { thinking: { type: "adaptive", display: "summarized" }, effort }

adaptive is confirmed accepted by this endpoint. If mapping xhigh -> adaptive changes the intended semantics of xhigh, the alternative is to not expose effort variants for this model on this adapter and surface that the capability is unavailable — rather than emitting a parameter that does nothing.

Regression test

Given opencode-go/qwen3.8-flash + variant=xhigh, assert the generated body contains a thinking block and does not rely on bare output_config.effort alone; then assert the response contains a thinking block and tokens.reasoning > 0. Keep default qwen3.8-flash, qwen3.7-plus, deepseek-v4.1-flash, and non-reasoning models covered as controls.

Related issues

• #40182 — --variant silently ignored when the value is not in reasoning_options. Different cause: here the value is valid and applied, but has no effect.
• #17876 (closed, not planned) — opposite symptom on native Anthropic (adaptive + effort sent, causes 400).

Note

Separately, opencode-go/qwen3.8-flash also has intermittent Internal server error / Type validation failed failures (26 of 63 local sessions). That is a different failure domain and is not related to this reasoning-variant bug.