mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-08-26 00:53:48 +00:00
22006ebf81
* fix(core): cap the DashScope effort tier at what its ladder accepts `/effort max` writes the tier into config, and the DashScope provider emitted it as a flat `reasoning_effort` for the qwen3.8-max family without checking the endpoint's ladder, which stops at `xhigh`. The server rejected it with a 400, and because the tier lives in config every later request in the session rebuilt the same body and 400d too. The tier also persists to settings.json, so new sessions re-broke. `max` exists only as a DeepSeek extension. Declare the tiers DashScope accepts and clamp through the existing `clampReasoningEffort`, warning once, the same way the Anthropic generator caps tiers its model lacks. Only the configured `reasoning.effort` is clamped. An explicit `reasoning_effort` in `extra_body` or `samplingParams` is a documented verbatim override and still ships unchanged. Refs: #9459 Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * fix(core): cap the effort tier at what each endpoint accepts Follow-up on the same defect at the generic layer, per cross-model review. The unified ladder ends at `max`, but `max` is a vendor extension: DeepSeek and GLM-5.2+ take it, and a generic OpenAI-compatible endpoint stops at `xhigh`. The effort-ladder design already specifies OpenAI `max -> xhigh`, but nothing implemented it, so a configured `max` reached the wire raw and 400d every later request in the session. Declare the accepted tiers on the provider and clamp there. The base provider ceilings at `xhigh`; DeepSeek and Z.ai override to the full ladder. The override is on the provider class, not the hostname, because which tiers a model accepts is a property of the model while the flat-vs-nested wire shape is a property of the endpoint: a self-hosted deepseek-* or glm-* model reached through the model-name fallback still understands `max`. Also corrects a wrong claim from the first commit. `max` is not DeepSeek-only: Anthropic opus/sonnet 4.6+ and every 5.x family accept it natively, and the DashScope note now says only that this family does not. Adds the warn-once and alias coverage the review found missing. Refs: #9459 Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * fix(core): scope the effort ceiling to verified endpoints Closes four gaps the cross-model review found in the previous commit. The clamp rewrote a `reasoning` object the user set in `samplingParams`. The pipeline hands those keys straight to the wire and skips the injection entirely, so that object is the user's own value and documented to ship verbatim. Skip the clamp when it is present. DashScope overrides buildRequest without calling super, and only capped its own flat qwen field, so a non-qwen model on a DashScope host still shipped a raw nested `max`. Route that branch through the generic ceiling, carving out GLM-5.2+, which does accept `max`. The DeepSeek and Z.ai ladders were keyed on the provider class, but both classes also route on a model-name substring, so anything merely named `deepseek-*` or `glm-*` claimed a tier its endpoint may reject. Gate both on the verified hostname, which is the rule deepseek.ts already documents for decisions about DeepSeek's own wire shape (#3613). Z.ai additionally gates on GLM-5.2+ rather than every `glm-*`, matching what its comment claimed. One existing DeepSeek test asserted a self-hosted deepseek-* keeps `max`. That was the old no-clamp behavior; an unverified endpoint now gets the generic ceiling, and the test says so. Refs: #9459 Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * docs(core): document the effort ceiling per endpoint Adds the Z.ai/GLM row, notes that DeepSeek's `max` is hostname-gated, and says plainly that a `reasoning` object inside `samplingParams` is the user's own value and is not clamped. Also adds the end-to-end pipeline test for the generic provider, mirroring the DashScope one: it drives pipeline.execute and asserts on the body handed to the SDK. Refs: #9459 Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * fix(core): answer the effort ceiling for the wire model The capability check read the configured model, but the pipeline resolves `request.model || contentGeneratorConfig.model`, so a request-level model override was answered against the wrong model. Configuring Z.ai `glm-5.2` and requesting `glm-4.6` shipped a raw `max` again, which is the failure this change exists to prevent. `supportedReasoningEfforts` becomes `supportedReasoningEffortsFor(model)` and takes the wire model. Drops the GLM exception on DashScope. It contradicted this PR's own docs, which say a `glm-*` model reached on a non-Z.ai host keeps the generic ceiling, and there is no evidence DashScope's GLM deployment accepts `max`. A quiet downgrade is the safer side to be wrong on. This also removes the import of Z.ai's helper from the DashScope provider. Adds the warn-once test for the base provider and regression tests that cross the configured and request models in both directions. Refs: #9459 Co-authored-by: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| assets | ||
| design | ||
| developers | ||
| e2e-tests | ||
| images | ||
| plans | ||
| superpowers | ||
| users | ||
| verification/abort-controller-refactor | ||
| _meta.ts | ||
| index.md | ||