feat(providers): complete Qwen3.8 Max support for Alibaba Token Plan

[MoonOld]

Problem

Alibaba Token Plan has promoted Qwen3.8 to the formal model id qwen3.8-max, but Maka does not include that id in the curated Token Plan fallback list. The list is authoritative for fresh connections when model discovery is unavailable or fails.

Alibaba now documents qwen3.8-max-preview as retired. The old id still routes to qwen3.8-max temporarily, with Credits and usage attributed to the formal model, and users are explicitly advised to update their configuration. Maka should therefore expose the formal id and migrate a persisted preview selection when a live inventory refresh supplies it, rather than offering the retired alias as a new choice.

A live China Token Plan /models check confirmed the current inventory contains qwen3.8-max and not the preview id.

Core support required

1. Put qwen3.8-max first in the curated China and global Token Plan fallback catalogs.
2. Treat qwen3.8-max-preview as a provider-scoped compatibility alias of qwen3.8-max for stored-selection reconciliation only.
3. Preserve the documented model contract:
   • 1M context window and 131072 maximum output;
   • vision, reasoning, structured output, and function calling;
   • reasoning choices off, low, medium, and xhigh; off maps to reasoning_effort: none.
4. Verify exact model-id preservation, Chat reasoning replay, and a two-step Maka tool loop.

Live findings

The formal model was verified on the China Token Plan service for:

• streamed text and low / medium / xhigh;
• two-step Function Calling with reasoning preserved across the tool result;
• vision input;
• thinking disabled with zero reasoning tokens;
• JSON structured output;
• a 34,960-input-token context probe;
• Responses API native code interpreter.

Remaining protocol work

The service supports Responses and native Harness tools, but Maka currently uses the OpenAI-compatible Chat wire for Alibaba Token Plan. Responses reasoning replay and provider-executed web_search, web_extractor, code_interpreter, text/image search event mappings need separate implementation and durable tests before Maka advertises them.

The dashed OpenAI-compatible provider-options warning observed during the live run is already tracked by #1430 and should not be duplicated here.

Acceptance criteria

• Fresh China and global Token Plan connections offer qwen3.8-max without relying on /models.
• The retired preview id is not offered, while stored preview selections reconcile to the formal id on inventory refresh.
• Model metadata and off / low / medium / xhigh choices match the current provider contract.
• The formal id completes a streamed, multi-step Maka function loop with reasoning replay.
• Responses and native Harness capabilities remain explicit follow-up work until implemented and tested.

References

• Token Plan personal overview: https://help.aliyun.com/zh/model-studio/token-plan-personal-overview
• Token Plan team overview: https://help.aliyun.com/zh/model-studio/token-plan-team-overview
• Chat parameters and reasoning effort: https://help.aliyun.com/zh/model-studio/qwen-api-via-openai-chat-completions
• Deep thinking: https://help.aliyun.com/zh/model-studio/deep-thinking
• Previous closed preview-only attempt: feat(core): add qwen3.8-max-preview to Alibaba Token Plan #1315

简体中文

Token Plan 当前正式模型 ID 是 qwen3.8-max。qwen3.8-max-preview 已结束预览并下线，只作为服务端临时兼容别名继续路由到正式版，Credits 与用量也按正式版统计。Maka 应在新连接中只提供正式 ID，并在模型目录刷新时把已有 Preview 选择迁移到正式 ID。

本 Issue 的核心范围包括中国区/国际区目录、兼容别名迁移、视觉与 Function Calling 元数据、off/low/medium/xhigh 推理档位，以及真实两步 Maka 工具循环。Responses API 和百炼原生 Harness 工具需要单独完成推理续接与 provider-executed 事件映射后才能宣称支持。

Disclosure: this issue was drafted with OpenAI Codex assistance and reviewed and approved by the human contributor.

[MoonOld]

I’m taking the required first slice now: add the documented qwen3.8-max-preview ID to both Alibaba Token Plan catalogs, pin its distinct preview contract, and add focused catalog, reasoning-wire, and generated two-step tool-loop coverage. I will keep the Responses/native Harness expansion out of this first PR unless the implementation proves it can be added without broadening the provider contract.

Disclosure: OpenAI Codex is assisting with implementation and verification; I will review the changes and remain the human contributor of record.

[MoonOld]

The required first slice is implemented in #3157.

It adds qwen3.8-max-preview to both Token Plan fallback catalogs and verifies the distinct preview metadata, reasoning efforts, current Chat reasoning replay, exact model id, and generated two-step tool loop.

I kept Responses and the native Harness tools as the remaining scope here: Maka currently has no live-verified Alibaba Responses reasoning replay, and enabling that wire speculatively could drop the reasoning state that must survive tool turns. The native tools likewise need provider-executed event mappings rather than a capability flag alone.

[MoonOld]

Correction after current official-document and live-service verification: the supported primary id is qwen3.8-max. Alibaba explicitly documents the preview id as retired and temporarily routed to the formal model, with billing/usage attributed to qwen3.8-max. PR #3157 has been rewritten accordingly and now also migrates stored preview selections through the provider-scoped alias table.

Live Maka adapter verification passed Chat reasoning at all supported efforts, thinking off, a two-step function loop with reasoning replay, vision input, structured JSON, and a 34,960-input-token context probe. The service-side Responses code interpreter also completed, but the Maka Responses/native-tool integration remains follow-up scope.

[MoonOld]

Concrete follow-ups are now split by protocol boundary:

• feat(providers): add Alibaba Token Plan Responses wire with durable Qwen reasoning replay #3162 — Alibaba Token Plan Responses wire and durable Qwen reasoning replay
• feat(runtime): map Alibaba Token Plan native Harness tools into Maka #3163 — provider-native Harness tool mappings (search, extractor, code interpreter, image search)
• feat(runtime): materialize native PDF inputs for verified PDF-capable providers #3164 — native PDF attachment materialization for PDF-capable models

This keeps #3157 limited to the formal model catalog, preview-alias migration, verified Chat reasoning controls, vision, and Maka-owned tool loops.