[Bug] GPT-5.6 explicit max effort is downgraded to xhigh in Claude → Responses conversion

[xiaodream551-a11y]

Self Checks / 自检

• I have read the FAQ section in README. / 我已阅读 README 中的常见问题。
• I have searched existing issues, including closed ones. / 我已搜索已有 Issue（含已关闭）。

CC Switch Version / 版本号

3.17.0

Operating System / 操作系统

Linux

Related App / 涉及应用

Claude Code

Steps to Reproduce / 重现步骤

1. Configure a Claude provider using apiFormat = "openai_responses", with the Claude Fable/Opus role mapped to gpt-5.6-sol.
2. In Claude Code, select /effort max.
3. Send a request through the CC Switch local proxy.
4. Inspect the Claude → OpenAI Responses conversion in resolve_reasoning_effort() and anthropic_to_responses().

CC Switch v3.17.0 currently contains:

return match effort {
    "low" => Some("low"),
    "medium" => Some("medium"),
    "high" => Some("high"),
    "max" => Some("xhigh"),
    _ => None,
};

Source in the v3.17.0 release:

cc-switch/src-tauri/src/proxy/providers/transform.rs

Lines 71 to 94 in 3d176b9

	/// Resolve the appropriate OpenAI `reasoning_effort` from an Anthropic request body.
	///
	/// Priority:
	/// 1. Explicit `output_config.effort` — preserves the user's intent directly.
	/// `low`/`medium`/`high` map 1:1; `max` maps to `xhigh`
	/// (supported by mainstream GPT models). Unknown values are ignored.
	/// 2. Fallback: `thinking.type` + `budget_tokens`:
	/// - `adaptive` → `xhigh` (adaptive = maximum reasoning effort)
	/// - `enabled` with budget → `low` (<4 000) / `medium` (4 000–15 999) / `high` (≥16 000)
	/// - `enabled` without budget → `high` (conservative default)
	/// - `disabled` / absent → `None`
	pub fn resolve_reasoning_effort(body: &Value) -> Option<&'static str> {
	// --- Priority 1: explicit output_config.effort ---
	if let Some(effort) = body
	.pointer("/output_config/effort")
	.and_then(|v| v.as_str())
	{
	return match effort {
	"low" => Some("low"),
	"medium" => Some("medium"),
	"high" => Some("high"),
	"max" => Some("xhigh"), // OpenAI xhigh = maximum reasoning effort
	_ => None, // unknown value — do not inject
	};

anthropic_to_responses() then emits the resolved value as:

{
  "model": "gpt-5.6-sol",
  "reasoning": { "effort": "xhigh" }
}

Therefore an explicit Claude Code request with output_config.effort = "max" is downgraded to OpenAI reasoning.effort = "xhigh".

OpenAI's current GPT-5.6 documentation explicitly lists none, low, medium, high, xhigh, and max, and describes max as distinct from xhigh:

• https://developers.openai.com/api/docs/guides/latest-model
• https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6

This can be observed without logging request contents: CC Switch request records confirm claude-fable-5 is mapped to gpt-5.6-sol and succeeds, while the conversion code deterministically changes the explicit effort value before forwarding.

Expected Behavior / 期望行为

For models that support OpenAI reasoning.effort = "max" (notably GPT-5.6), preserve an explicit Claude output_config.effort = "max":

{
  "model": "gpt-5.6-sol",
  "reasoning": { "effort": "max" }
}

The mapping should be model- or capability-aware rather than globally changing max to xhigh, because older models/providers may still accept only up to xhigh.

Possible direction:

"max" if supports_max_reasoning_effort(model) => Some("max"),
"max" => Some("xhigh"),

Alternatively, provider/model-level supported effort metadata could drive the mapping.

Suggested tests:

1. GPT-5.6 Responses: Claude max → OpenAI reasoning.effort = "max".
2. Older GPT model supporting only through xhigh: Claude max → xhigh compatibility fallback.
3. Existing low/medium/high behavior remains unchanged.

Actual Behavior / 实际行为

CC Switch globally maps Claude max to OpenAI xhigh, despite GPT-5.6 supporting both as separate effort levels. The user's explicit highest-effort selection is therefore not preserved for gpt-5.6-sol.

Additional Context / 补充信息

Related but distinct issues:

• cc使用ultracode 同时本地路由到openai协议 effort失效 #3389 covers incoming Claude xhigh being ignored, not explicit max being downgraded.
• [回归] v3.16.5 的 NativeResponses 模型目录模板导致推理档位和图片能力退化 #5266 covers generated Codex model-catalog capabilities, not Claude → Responses request conversion.
• API Error: 400 The parameter output_config.effort specified in the request are not valid: expected low, medium, high or max, but got xhigh instead. #2851 demonstrates that global max/xhigh assumptions are already provider-dependent.
• Codex Desktop gpt-5.6-sol uses ultra reasoning effort, but cc-switch Codex /responses path only models up to xhigh/max #5209 discusses Codex Desktop's ultra UI behavior; this report concerns the official GPT-5.6 API-level max effort.

No credentials, prompts, or full request bodies are included in this report.

[xiaodream551-a11y]

I found an already-open PR that appears to address this same conversion bug: #5267 (fix(proxy): map Claude effort by GPT-5 generation). Its frontend and backend CI checks are currently passing, and GitHub reports it as mergeable, although it has not been merged yet.

The narrow behavior needed for this issue is:

• preserve an explicit output_config.effort = "max" as reasoning.effort = "max" when the mapped outbound model is in the GPT-5.6 family (gpt-5.6, gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna);
• retain the existing compatibility downgrade to xhigh for older models that do not accept max;
• leave the inferred thinking.type = "adaptive" fallback as xhigh, because that is not an explicit request for max;
• add a Responses integration test that checks the final reasoning.effort, plus regression tests for GPT-5.6 variants and an older model.

OpenAI's current documentation lists max as a distinct supported reasoning effort for GPT-5.6, rather than an alias for xhigh:
https://developers.openai.com/api/docs/guides/deployment-checklist#set-up-reasoningeffort

So this should be model-aware clamping, not a global max -> max replacement. Cross-linking #5267 here may also help avoid implementing the same fix twice.