Files
Netcatty/components/settings
陈大猫andArcadia 2368fc8232 feat(ai): support extended reasoning effort with GPT-5.5 xhigh (#4111)
* feat(ai): support xhigh and max reasoning effort levels

The provider-level reasoning default only offered low/medium/high, so models
that document the full graded ladder could not be configured to their top
tiers. DeepSeek V4/V4.1 grade `reasoning_effort` continuously up to `max`, and
Anthropic exposes `max` on 4.6+ plus `xhigh` from 4.7 onward.

Beyond the missing select options, three separate layers silently discarded
the higher values:

- REASONING_RANK stopped at `high`, so the visibility resolver treated
  `xhigh`/`max` as stale and fell back to `low`.
- The OpenAI capability check only matched `deepseek-r1`, which left the
  DeepSeek V4 family unrecognised and hid the thinking chip entirely.
- The pinned `@ai-sdk/openai@4.0.8` validates provider options against a zod
  enum capped at `xhigh`, so `max` on Chat Completions threw
  "invalid openai provider options" before any request was made.

Add a shared PROVIDER_REASONING_EFFORT_OPTIONS list used by both the settings
select and its save guard, extend the rank ladder, and teach the per-model
derivation about DeepSeek V4 grading plus the Anthropic effort generations.
Bump `@ai-sdk/openai` to 4.0.11, the first release whose enum accepts `max`;
it keeps 4.0.8's data-URL image fix.

Models whose vocabulary stops at `high` now clamp these levels onto their own
ceiling instead of dropping the setting, so an explicit user choice is never
silently ignored and no unsupported value reaches a provider.

* fix(ai): keep extended effort for explicit custom relay aliases

A custom OpenAI-compatible provider with an explicit reasoning default is the
escape hatch for relay model IDs that match no capability pattern. That branch
still advertised only low/medium/high, so resolveVisibleCattyThinkingLevel
treated a configured `xhigh` or `max` as stale and downgraded it to `high`
before the request was built.

Advertise the full ladder there: the explicit provider default is already the
user's assertion that the alias accepts reasoning_effort, so it should also
carry the ceiling they picked. Unrecognized aliases without a default keep
advertising nothing.

* test(ai): verify max effort with the pinned OpenAI SDK

---------

Co-authored-by: Arcadia <57169358+RayYiHang@users.noreply.github.com>
2026-11-06 00:10:09 +08:00
..