alibaba-token-plan: `reasoning_effort` never sent for Qwen 3.8 Flash; Qwen `reasoning_content` history not replayed

Environment

omp 18.1.10 (installed binary), provider alibaba-token-plan, model qwen3.8-flash, modelRoles.coder: alibaba-token-plan/qwen3.8-flash:minimal.

Defect 1: effort level dropped on the wire

Catalog rule providers/alibaba-token-plan.kdl:3 sets thinkingFormat: qwen for the whole provider. The chat-completions request builder then takes the qwen-enable-thinking-false branch, which sends enable_thinking: true and adds reasoning_effort only when qwenTemplateReasoningEffort is set (local backends only). Only qwen3.8-max has a whenThinking { thinkingFormat: openai, extraBody { enable_thinking: true } } rule (alibaba-token-plan.kdl:20), so effort reaches DashScope for max and for nothing else.

Captured with PI_REQ_DEBUG=1 omp -p --model alibaba-token-plan/qwen3.8-flash:minimal:

{"model":"qwen3.8-flash","enable_thinking":true,"stream":true,"max_completion_tokens":64000}

DashScope docs: Qwen 3.8 thinking is on by default; reasoning_effort accepts low|medium|xhigh (default xhigh, budget 262144 tokens; low = 4096; minimal maps to low; none maps to enable_thinking=false). reasoning_effort and thinking_budget must not be combined. https://www.alibabacloud.com/help/en/model-studio/qwen-api-via-openai-chat-completions

Direct A/B against DashScope, identical prompt:

Workaround that fixes it (models.yml):

providers:
  alibaba-token-plan:
    modelOverrides:
      qwen3.8-flash:
        compat:
          supportsReasoningEffort: true
          whenThinking:
            thinkingFormat: openai
            extraBody:
              enable_thinking: true

Suggested fix: apply the qwen3.8-max wire rule to the whole class qwen, revision >=3.8.0 <3.9.0 block (which already sets supportsReasoningEffort: true at alibaba-token-plan.kdl:11), or set qwenTemplateReasoningEffort-equivalent behaviour for the remote DashScope endpoint.

Defect 2: Qwen reasoning history is dropped

For remote Alibaba models the resolved compat has replayReasoningContent: false, so the history converter omits prior reasoning_content from assistant turns. DashScope: for qwen3.8-max/flash preserve_thinking defaults to true and "you must send back all historical reasoning_content in the reasoning_content field". The API accepts the omission silently, so the model re-derives earlier decisions each turn. replayReasoningContent / qwenPreserveThinking are not exposed in the models.yml compat schema, so there is no user-side workaround.

Reproduction

Offline: extract Bv/f5s from the bundle and call with the resolved qwen compat; PI_REQ_DEBUG=1 live capture as above.