Mythosia.AI.Providers.Alibaba - Release Notes

v3.0.2

Changed

• Completion resolves the prepared request message snapshot before request execution instead of continuing with the original caller input. This keeps captured request context and message ownership consistent with the core request pipeline.
• Rebuilds against Mythosia.AI 8.2.0 and transitively Mythosia.AI.Abstractions 4.2.0.

Compatibility

• Existing Qwen public APIs, endpoint defaults and thinking controls remain unchanged. No additional source migration is required from 3.0.x; this patch does not add provider-specific priority processing.
• The inherited core SSE path retains a known cancellation/timeout limitation with custom HTTP content that buffers a successful response during stream acquisition. A canceled Run can remain pending and block service reuse until acquisition finishes; default transport checks passed. See the streaming limitation.

v3.0.1

Changed

• Rebuilds the Qwen adapter against Mythosia.AI 8.1.0 and transitively Mythosia.AI.Abstractions 4.1.0. Inherits core request-context and stateless-summary corrections.

Compatibility

• Existing Qwen public APIs, endpoint defaults and thinking controls remain unchanged. No additional source migration is required from 3.0.0. The shared speed API does not add Qwen-specific priority processing; unsupported explicit speed selections remain rejected.

v3.0.0

This coordinated major release changes public contracts. See the v8 migration guide before upgrading the package family.

Added

• Inherits pure GetCapabilities() inspection on Qwen services and request builders. Definitions reflect endpoint mode, captured provider settings, model overrides and Ollama wire-ID mapping; custom deployment support stays Unknown when not known. UI controls and invocation validation use the shared capability definitions, with existing request validation retained. See capability inspection.

• Inherits the major AIRun.Result migration to Task<AIRunResult>, with accumulated .Text, reported usage/sources, captured provider/requested model, actual model when reported, library rounds, and finish details. Stream observation is optional; ordinary GetCompletionAsync remains a string API. Existing Run string callers must read (await run.Result).Text and rebuild. See migration.

• AIRunResult.RequestedModel records the model ID actually sent: the captured ModelIdOverride when set, or the endpoint-specific model ID, including Ollama mapping such as qwen3-32b to qwen3:32b. A builder retains its captured override after service defaults change; Model remains the separate server-reported model.

• Ordinary completion cancellation: Qwen forwards the caller token through DashScope, vLLM and Ollama HTTP requests, local tools and later rounds. The inherited completion, typed, builder and message-chain APIs accept the token. Caller cancellation is distinct from request-policy timeouts; remote generation or billing cancellation is not guaranteed. See the common contract.

Internal

• Inherits the shared streaming terminal guard: later content or changed finish reasons fail the round before saving its response or executing tools. A final delta in the first terminal event and trailing usage-only events remain supported.

• Qwen profile inspection shares only native mode flags with execution through the pure capability-profile hook; queries do not invoke execution preparation or serialize tool defaults.

• Inherits common tool return normalization, cooperative cancellation, and error-result handling from the core service; Qwen uses the same execution contract without requiring provider-native async tool support.

• Inherits CreateRequest(...) and immutable request builders from the core service. Common and Qwen provider defaults are captured per request; existing Qwen reasoning/search capability limits and shared-conversation rules remain. See the request guide.

• Rebuilt against Mythosia.AI v8.0.0 and its provider request-option capture hooks, with Mythosia.AI.Abstractions v4.0.0 as an indirect dependency.

Compatibility

• Existing source callers can omit the new completion token. Qwen endpoint behavior and Run controls remain; rebuild callers for the changed completion signatures. Claude-specific options are not added to Qwen.

---

v2.0.1

Fixed

• Qwen's non-streaming completion override now enters the common request-feature scope. Unsupported common reasoning or hosted-search options are validated and consumed before an HTTP request, rather than bypassing validation or leaking into a later call.

Changed

• Targets Mythosia.AI v7.1.0 and its request-scope implementation. Qwen inherits the core StartRunAsync controls for output, final results, cancellation, and disposal; existing tools continue through the common round policy.

Compatibility

• Requires Mythosia.AI v7.1.0, which depends on Mythosia.AI.Abstractions v3.1.0. No existing Qwen public member is removed or changed.
• DashScope, vLLM, and Ollama retain their provider-specific ThinkingMode and endpoint settings. This adapter does not gain native steering, hosted search, or native asynchronous-tool support; Qwen runs report CanSteer = false. Functions with AllowAsync set still use ordinary execution.
• Qwen remains a chat-completion provider and does not implement IImageGenerationService. The v2.0.0 migration notes below still apply when upgrading from v1.x.

---

v2.0.0

This is the Alibaba provider release paired with Mythosia.AI v7. Follow the v7 migration guide before upgrading.

Changed

• Ordered function-call batches — Qwen non-streaming and streaming paths now preserve every tool call returned in one assistant turn and send the matching ordered result batch back through the common Mythosia.AI v7 continuation contract.
• Common handler scheduling — FunctionCallingPolicy.ExecutionMode selects sequential compatibility behavior or bounded-parallel local handler execution; MaxConcurrency limits parallel work while provider call order remains stable.

Removed

• Legacy image-generation overrides — QwenService.GenerateImageAsync and GenerateImageUrlAsync were unsupported stubs inherited from the old core abstraction and have been removed with the Mythosia.AI v7 API surface. Qwen remains a chat-completion provider and does not implement IImageGenerationService.

Compatibility

• Recompiled against and requires Mythosia.AI v7.0.0.
• Breaking release for callers that referenced the removed public overrides or derived from the former single-function extraction contract.

---

v1.2.8

Fixed

• Context-overflow rejections reach the core's recovery. QwenService builds and throws its own HTTP failure, so it did not produce the ContextLengthExceededException that Mythosia.AI v6.8.0 reacts to — a Qwen model, or any vLLM deployment served through this provider, would have been refused for exceeding the context window and never compacted or re-sent, while every other provider recovered. The rejection now goes through AIHttpErrorFactory, which is also where vLLM's wording is recognised.

Compatibility

• Requires Mythosia.AI v6.8.0. No API changes.

---

v1.2.7

Fixed

• Thinking-off was silently dropped for models whose id does not literally contain qwen3. The request builder gated the "thinking off" signal behind a model-name check (modelId.Contains("qwen3")), while "thinking on" was always sent. Because a served model name is chosen freely by the operator (vLLM --served-model-name, aliases), a Qwen 3 model served under any other name never received enable_thinking = false — the caller believed reasoning was disabled while the server kept its default (reasoning on). This surfaced as summarization requests emitting long reasoning traces and hitting request timeouts.
• enable_thinking was sent in the wrong shape on vLLM for models outside the qwen3.5 name path. It was emitted as a top-level parameter instead of chat_template_kwargs.enable_thinking, so vLLM never applied it. Both the on and off signals are now sent in the documented per-platform format for every model.

Changed

• Thinking parameters are now derived solely from the configured ThinkingMode and translated per platform (DashScope / vLLM / Ollama). The provider no longer inspects the model id to infer capability — an unsupported model is expected to ignore the parameter or surface an error, which is preferable to a directive disappearing silently.

Internal

• Removed the duplicated Qwen 3.5-specific request path and the IsQwen35 / IsQwen3ThinkingCapable name heuristics; both request paths are unified into a single ApplyThinkingParameters step. No public API change.

Compatibility

• No API changes. Callers that set ThinkingMode (directly or via AIRequestProfile.DisableReasoning) will now actually have that setting reach the server; this can change model behavior where the directive was previously being dropped.

---

v1.2.6

Compatibility

• Recompiled for the Mythosia.AI v6.4.0 release line. No API changes.

---

v1.2.5

Compatibility

• Recompiled for the Mythosia.AI v6.3.0 release line. No API changes.

---

v1.2.4

Compatibility

• Recompiled for the Mythosia.AI v6.2.0 release line. No API changes.

---

v1.2.3

Compatibility

• Recompiled for the Mythosia.AI v6.1.0 release line. No API changes.

---

v1.2.2

Compatibility

• Recompiled against Mythosia.AI v6.0.0. No API changes.

---

v1.2.1

Compatibility

• Recompiled against Mythosia.AI v5.3.0. No API changes.

---

v1.2.0 - Mythosia.AI v5.2.0 Binary Compatibility

✅ Compatibility

• Recompiled against Mythosia.AI v5.2.0 (Abstractions split: AIService now implements IAIService)
• No API changes — fixes TypeLoadException when used alongside Mythosia.AI.Abstractions v1.0.0

---

🚀 v1.1.0 - Mythosia.AI v5.1.0 Compatibility & Token Usage Support

Token Usage in Streaming

QwenService streaming now reports token usage (input, output, cached, reasoning tokens) on Completion events via StreamingContent.Usage, inherited from the core package.

✅ Compatibility

• Compatible with Mythosia.AI v5.1.0
• Breaking: StreamOptions.IncludeTokenInfo / WithTokenInfo() removed in core package (see Mythosia.AI v5.1.0 release notes for migration guide)

---

🔧 v1.0.2 - Mythosia.AI v5.0.1 Compatibility

Streaming Architecture Alignment

• Aligned with Mythosia.AI v5.0.1 Template Method streaming refactor: QwenService now overrides StreamRoundAsync instead of StreamAsync, inheriting base class round-loop management, StatelessMode handling, and automatic conversation summary policy.

✅ Compatibility

• Compatible with Mythosia.AI v5.0.1
• No breaking changes

---

🐛 v1.0.1 - Thinking Request Handling Fix

DashScope Qwen 3.5 파라미터 포맷 수정

DashScope 엔드포인트에서 Qwen 3.5 thinking 파라미터가 chat_template_kwargs.enable_thinking으로 잘못 전송되던 문제를 수정했습니다. DashScope는 top-level enable_thinking 파라미터를 사용합니다.

vLLM / DashScope 요청 경로 분리

vLLM과 DashScope가 동일한 chat_template_kwargs 경로를 공유하던 문제를 수정했습니다.

Platform	Thinking On	Thinking Off
DashScope	enable_thinking = true	enable_thinking = false
vLLM	chat_template_kwargs.enable_thinking = true	chat_template_kwargs.enable_thinking = false
Ollama	reasoning.effort = "high"	(파라미터 생략)

Qwen3 모델 thinking-off 명시 전송

Qwen3 thinking-capable 모델에서 ThinkingMode가 off일 때 DashScope / vLLM에 enable_thinking = false를 명시적으로 전송하도록 수정했습니다. 이전에는 파라미터가 생략되어 서버 기본값으로 thinking이 의도치 않게 활성화될 수 있었습니다.

✅ Compatibility

• Compatible with Mythosia.AI v5.0.0
• No breaking changes

---

🚀 v1.0.0 - Package Documentation, Qwen 3.5 Request Handling, and Request Profile Integration

NuGet Packaging Metadata and Package Docs

This release also includes the package-level documentation and NuGet metadata alignment that had previously been tracked separately.

• Added package README.md
• Added package RELEASE_NOTES.md
• Added NuGet readme metadata to the project file
• Added package tags, description, and project URL metadata
• Added packaging entries so package documentation files are included properly

Expanded AlibabaModels Catalog

The package now exposes a broader built-in Qwen model catalog through AlibabaModels.

Added coverage includes Qwen 3 and Qwen 3.5 families such as:

• AlibabaModels.Qwen3_235B
• AlibabaModels.Qwen3_32B
• AlibabaModels.Qwen3_5_397B
• AlibabaModels.Qwen3_5_27B
• AlibabaModels.Qwen3_5_0_8B

This makes it easier to target newer Alibaba model variants without hardcoding IDs in application code.

Qwen 3.5 Thinking Request Handling

QwenService now applies Qwen 3.5-specific request shaping when thinking mode is enabled.

• vLLM and DashScope-style requests use chat_template_kwargs.enable_thinking
• Ollama requests continue to map thinking mode through reasoning parameters

This keeps thinking-mode behavior aligned with how different Qwen 3.5 endpoints expect the request payload.

AIRequestProfile.DisableReasoning Integration

With the core Mythosia.AI v5.0.0 request-profile APIs, QwenService now respects per-request reasoning disablement.

When AIRequestProfile.DisableReasoning is set, the provider temporarily turns ThinkingMode off for that call and restores the previous state afterward.

var answer = await service.GetCompletionAsync(
    "Summarize this policy without reasoning output.",
    new AIRequestProfile
    {
        DisableReasoning = true
    });

✅ Compatibility

• Package version advanced to v1.0.0
• Compatible with Mythosia.AI v5.0.0
• No breaking changes