feat(claude): add adaptive thinking + output effort for Opus 4.7+

[NicolasWalter]

What type of PR is this?

feat

Check the PR title.

• This PR title matches the format: <type>(optional scope): <description>
• The description is user-oriented and clear enough for others to understand.
• User documentation PR attached (not needed — the new API is documented in the adapter's own README / README_zh, both updated in this PR).

(Optional) Translate the PR title into Chinese.

feat(claude): 为 Opus 4.7+ 模型支持自适应思考（adaptive thinking）及输出精力等级（output effort）

(Optional) More detailed description for this PR

en:

Claude Opus 4.7 rejects the legacy {"type":"enabled","budget_tokens":N} thinking shape and requires {"type":"adaptive"} with an optional output_config.effort. The current adapter only emits the legacy shape, so calling claude.WithThinking(...) against Opus 4.7 returns:

invalid_request_error: "thinking.type.enabled" is not supported for this model.
Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.

This PR adds an opt-in adaptive path. Changes are additive and backward-compatible — every existing caller keeps its current behavior because the empty Mode defaults to ThinkingModeEnabled.

Changes

• Thinking.Mode (ThinkingModeEnabled default, ThinkingModeAdaptive for Opus 4.7+). Empty mode preserves the pre-existing behavior.
• Thinking.Effort ("low" | "medium" | "high" | "max") → populates params.OutputConfig.Effort using the SDK's typed enum. Trim + case-insensitive; unknown values are silently ignored to match the rest of the adapter's leniency.
• Dispatch in genMessageNewParams: emit OfAdaptive for adaptive mode, OfEnabled (with BudgetTokens) otherwise. Effort applies to both.
• normalizeOutputEffort helper at the bottom of claude.go (next to getEnvWithFallbacks). utils.go stays untouched (no new imports there).
• README + README_zh snippet added next to the existing WithThinking example.
• TestThinkingConfig with 7 subtests covering the relevant paths: legacy default, empty mode default, adaptive with/without Effort, Effort normalization (trim + case), unknown Effort ignored, Enable=false short-circuit. All pass under mockey with the same style as the existing tests.

No new module dependency: the adapter already imports github.com/anthropics/anthropic-sdk-go (5 packages from it, actually) and the SDK exposes ThinkingConfigAdaptiveParam and OutputConfigParam since v1.25+. Nothing added to go.mod.

zh(optional):
本 PR 为 Claude 适配器增加 adaptive thinking 和 output effort 支持，以兼容 Opus 4.7 及更新模型。向后兼容：Mode 为空时保留原有 OfEnabled + budget_tokens 行为。新增 Thinking.Mode、Thinking.Effort 字段和 ThinkingModeEnabled / ThinkingModeAdaptive 常量；同时更新 README 示例，新增 7 个单元测试覆盖主要路径；未引入新的模块依赖（anthropic-sdk-go 已是适配器的直接依赖）。

(Optional) Which issue(s) this PR fixes

Fixes #780

(optional) The PR that updates user documentation

N/A — adapter-local README + README_zh updated in this PR.

[CLAassistant]

All committers have signed the CLA.

[tangchaojun-bytedance]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

Nicolas Walter seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
You have signed the CLA already but the status is still pending? Let us recheck it.

@NicolasWalter hello，please sign the CLA.

[tangchaojun-bytedance]

@NicolasWalter please solve conflicts.

[tangchaojun-bytedance]

@NicolasWalter Thanks for the PR! The feature direction looks good to me, and I confirmed that the current Claude adapter does not yet expose adaptive thinking / output effort support.

A few small suggestions before merging:

1. Thinking.Mode currently falls back to legacy enabled for any unknown value because of the default branch. Could we make it explicit instead?

   • "" / ThinkingModeEnabled -> legacy enabled
   • ThinkingModeAdaptive -> adaptive
   • any other value -> return an error

   This avoids silently sending the wrong request shape when users mistype the mode.

2. Effort currently supports low | medium | high | max, but the latest Anthropic Go SDK also exposes OutputConfigEffortXhigh. Could we add xhigh as well?

3. The comment mentions top-level OutputEffort config/option, but this PR does not add a Config.OutputEffort or WithOutputEffort option. Could we either remove that sentence or add the corresponding API to avoid confusing users?

[NicolasWalter]

Thanks for the review. I addressed all three points:

1. Thinking.Mode is now handled explicitly:

   • empty mode and ThinkingModeEnabled use the legacy enabled + budget_tokens request shape
   • ThinkingModeAdaptive uses the adaptive request shape
   • any unsupported mode now returns an error instead of silently falling back

2. Added support for xhigh effort.

3. Removed the misleading comment mentioning a top-level OutputEffort config/option, since this PR only exposes effort through Thinking.Effort.

I also rebased the branch on the latest main and resolved the conflicts.

[tangchaojun-bytedance]

@NicolasWalter hello, I suggest exposing Effort as a package-level typed enum instead of string. Since this is a new public field, using constants like OutputEffortLow/Medium/High/Xhigh/Max is more consistent with ThinkingMode and other eino-ext adapters, and avoids silently ignoring typos through normalization.