fix(claude): clamp max_tokens to model limit in normalizeClaudeBudget

When adjustedBudget < minBudget, the previous fix blindly set
max_tokens = budgetTokens+1 which could exceed MaxCompletionTokens.

Now: cap max_tokens at MaxCompletionTokens, recalculate budget, and
disable thinking entirely if constraints are unsatisfiable.

Add unit tests covering raise, clamp, disable, and no-op scenarios.

internal/thinking/provider/claude/apply.go

@@ -174,7 +174,8 @@ func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo
 	// Ensure the request satisfies Claude constraints:
 	//  1) Determine effective max_tokens (request overrides model default)
 	//  2) If budget_tokens >= max_tokens, reduce budget_tokens to max_tokens-1
-	//  3) If the adjusted budget falls below the model minimum, leave the request unchanged
+	//  3) If the adjusted budget falls below the model minimum, try raising max_tokens
+	//     (clamped to MaxCompletionTokens); disable thinking if constraints are unsatisfiable
 	//  4) If max_tokens came from model default, write it back into the request
 
 	effectiveMax, setDefaultMax := a.effectiveMaxTokens(body, modelInfo)
@@ -193,10 +194,28 @@ func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo
 		minBudget = modelInfo.Thinking.Min
 	}
 	if minBudget > 0 && adjustedBudget > 0 && adjustedBudget < minBudget {
-		// If enforcing the max_tokens constraint would push the budget below the model minimum,
-		// increase max_tokens to accommodate the original budget instead of leaving the
-		// request unchanged (which would cause a 400 error from the API).
-		body, _ = sjson.SetBytes(body, "max_tokens", budgetTokens+1)
+		// Enforcing budget_tokens < max_tokens pushed the budget below the model minimum.
+		// Try raising max_tokens to fit the original budget.
+		needed := budgetTokens + 1
+		maxAllowed := 0
+		if modelInfo != nil {
+			maxAllowed = modelInfo.MaxCompletionTokens
+		}
+		if maxAllowed > 0 && needed > maxAllowed {
+			// Cannot use original budget; cap max_tokens at model limit.
+			needed = maxAllowed
+		}
+		cappedBudget := needed - 1
+		if cappedBudget < minBudget {
+			// Impossible to satisfy both budget >= minBudget and budget < max_tokens
+			// within the model's completion limit. Disable thinking entirely.
+			body, _ = sjson.DeleteBytes(body, "thinking")
+			return body
+		}
+		body, _ = sjson.SetBytes(body, "max_tokens", needed)
+		if cappedBudget != budgetTokens {
+			body, _ = sjson.SetBytes(body, "thinking.budget_tokens", cappedBudget)
+		}
 		return body
 	}
 

internal/thinking/provider/claude/apply_test.go

@@ -0,0 +1,99 @@
+package claude
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/tidwall/gjson"
+)
+
+func TestNormalizeClaudeBudget_RaisesMaxTokens(t *testing.T) {
+	a := &Applier{}
+	modelInfo := &registry.ModelInfo{
+		MaxCompletionTokens: 64000,
+		Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000},
+	}
+	body := []byte(`{"max_tokens":1000,"thinking":{"type":"enabled","budget_tokens":5000}}`)
+
+	out := a.normalizeClaudeBudget(body, 5000, modelInfo)
+
+	maxTok := gjson.GetBytes(out, "max_tokens").Int()
+	if maxTok != 5001 {
+		t.Fatalf("max_tokens = %d, want 5001, body=%s", maxTok, string(out))
+	}
+}
+
+func TestNormalizeClaudeBudget_ClampsToModelMax(t *testing.T) {
+	a := &Applier{}
+	modelInfo := &registry.ModelInfo{
+		MaxCompletionTokens: 64000,
+		Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000},
+	}
+	body := []byte(`{"max_tokens":500,"thinking":{"type":"enabled","budget_tokens":200000}}`)
+
+	out := a.normalizeClaudeBudget(body, 200000, modelInfo)
+
+	maxTok := gjson.GetBytes(out, "max_tokens").Int()
+	if maxTok != 64000 {
+		t.Fatalf("max_tokens = %d, want 64000 (capped to model limit), body=%s", maxTok, string(out))
+	}
+	budget := gjson.GetBytes(out, "thinking.budget_tokens").Int()
+	if budget != 63999 {
+		t.Fatalf("budget_tokens = %d, want 63999 (max_tokens-1), body=%s", budget, string(out))
+	}
+}
+
+func TestNormalizeClaudeBudget_DisablesThinkingWhenUnsatisfiable(t *testing.T) {
+	a := &Applier{}
+	modelInfo := &registry.ModelInfo{
+		MaxCompletionTokens: 1000,
+		Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000},
+	}
+	body := []byte(`{"max_tokens":500,"thinking":{"type":"enabled","budget_tokens":2000}}`)
+
+	out := a.normalizeClaudeBudget(body, 2000, modelInfo)
+
+	if gjson.GetBytes(out, "thinking").Exists() {
+		t.Fatalf("thinking should be removed when constraints are unsatisfiable, body=%s", string(out))
+	}
+}
+
+func TestNormalizeClaudeBudget_NoClamping(t *testing.T) {
+	a := &Applier{}
+	modelInfo := &registry.ModelInfo{
+		MaxCompletionTokens: 64000,
+		Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000},
+	}
+	body := []byte(`{"max_tokens":32000,"thinking":{"type":"enabled","budget_tokens":16000}}`)
+
+	out := a.normalizeClaudeBudget(body, 16000, modelInfo)
+
+	maxTok := gjson.GetBytes(out, "max_tokens").Int()
+	if maxTok != 32000 {
+		t.Fatalf("max_tokens should remain 32000, got %d, body=%s", maxTok, string(out))
+	}
+	budget := gjson.GetBytes(out, "thinking.budget_tokens").Int()
+	if budget != 16000 {
+		t.Fatalf("budget_tokens should remain 16000, got %d, body=%s", budget, string(out))
+	}
+}
+
+func TestNormalizeClaudeBudget_AdjustsBudgetToMaxMinus1(t *testing.T) {
+	a := &Applier{}
+	modelInfo := &registry.ModelInfo{
+		MaxCompletionTokens: 8192,
+		Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000},
+	}
+	body := []byte(`{"max_tokens":8192,"thinking":{"type":"enabled","budget_tokens":10000}}`)
+
+	out := a.normalizeClaudeBudget(body, 10000, modelInfo)
+
+	maxTok := gjson.GetBytes(out, "max_tokens").Int()
+	if maxTok != 8192 {
+		t.Fatalf("max_tokens = %d, want 8192 (unchanged), body=%s", maxTok, string(out))
+	}
+	budget := gjson.GetBytes(out, "thinking.budget_tokens").Int()
+	if budget != 8191 {
+		t.Fatalf("budget_tokens = %d, want 8191 (max_tokens-1), body=%s", budget, string(out))
+	}
+}