Fix Anthropic thinking budgets

Normalize Anthropic thinking budgets while building message parameters so budget_tokens stays below max_tokens with response headroom. When the configured max token budget is too small to support Anthropic's minimum thinking budget, omit thinking for that request instead of sending an invalid payload.

Signed-off-by: Cursor Agent <cursoragent@cursor.com>

Co-authored-by: Sacha Al Himdani <SachaProbo@users.noreply.github.com>

pkg/llm/anthropic/provider.go

@@ -47,6 +47,11 @@ type (
 	}
 )
 
+const (
+	anthropicThinkingMinBudgetTokens    = 1024
+	anthropicThinkingOutputTokenReserve = 1024
+)
+
 func WithHTTPClient(c *http.Client) Option {
 	return func(cfg *config) { cfg.httpClient = c }
 }
@@ -163,8 +168,8 @@ func buildParams(req *llm.ChatCompletionRequest) (anthropic.MessageNewParams, er
 		params.ToolChoice = buildToolChoice(req.ToolChoice)
 	}
 
-	if req.Thinking != nil && req.Thinking.Enabled {
-		params.Thinking = anthropic.ThinkingConfigParamOfEnabled(int64(req.Thinking.BudgetTokens))
+	if budgetTokens, ok := resolveThinkingBudget(*req.MaxTokens, req.Thinking); ok {
+		params.Thinking = anthropic.ThinkingConfigParamOfEnabled(int64(budgetTokens))
 	}
 
 	if req.ResponseFormat != nil {
@@ -192,6 +197,28 @@ func buildParams(req *llm.ChatCompletionRequest) (anthropic.MessageNewParams, er
 	return params, nil
 }
 
+func resolveThinkingBudget(maxTokens int, thinking *llm.ThinkingConfig) (int, bool) {
+	if thinking == nil || !thinking.Enabled || thinking.BudgetTokens <= 0 {
+		return 0, false
+	}
+
+	if maxTokens <= anthropicThinkingMinBudgetTokens {
+		return 0, false
+	}
+
+	maxBudgetTokens := maxTokens - anthropicThinkingOutputTokenReserve
+	if maxBudgetTokens < anthropicThinkingMinBudgetTokens {
+		maxBudgetTokens = maxTokens - 1
+	}
+
+	budgetTokens := max(min(thinking.BudgetTokens, maxBudgetTokens), anthropicThinkingMinBudgetTokens)
+	if budgetTokens >= maxTokens {
+		return 0, false
+	}
+
+	return budgetTokens, true
+}
+
 func extractSystem(messages []llm.Message) (system []string, rest []llm.Message) {
 	for _, msg := range messages {
 		if msg.Role == llm.RoleSystem {

pkg/llm/anthropic/provider_test.go

@@ -0,0 +1,127 @@
+// Copyright (c) 2026 Probo Inc <hello@probo.com>.
+//
+// Permission to use, copy, modify, and/or distribute this software for any
+// purpose with or without fee is hereby granted, provided that the above
+// copyright notice and this permission notice appear in all copies.
+//
+// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+// REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+// AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+// INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+// LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+// OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+// PERFORMANCE OF THIS SOFTWARE.
+
+package anthropic
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.probo.inc/probo/pkg/llm"
+)
+
+func TestResolveThinkingBudget(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name      string
+		maxTokens int
+		thinking  *llm.ThinkingConfig
+		want      int
+		wantOK    bool
+	}{
+		{
+			name:      "disabled",
+			maxTokens: 4096,
+			thinking:  &llm.ThinkingConfig{Enabled: false, BudgetTokens: 4000},
+		},
+		{
+			name:      "leaves valid budget unchanged",
+			maxTokens: 49152,
+			thinking:  &llm.ThinkingConfig{Enabled: true, BudgetTokens: 40000},
+			want:      40000,
+			wantOK:    true,
+		},
+		{
+			name:      "caps oversized default budget",
+			maxTokens: 4096,
+			thinking:  &llm.ThinkingConfig{Enabled: true, BudgetTokens: 40000},
+			want:      3072,
+			wantOK:    true,
+		},
+		{
+			name:      "caps oversized raised budget",
+			maxTokens: 32768,
+			thinking:  &llm.ThinkingConfig{Enabled: true, BudgetTokens: 40000},
+			want:      31744,
+			wantOK:    true,
+		},
+		{
+			name:      "omits thinking when max tokens cannot hold minimum budget",
+			maxTokens: 1024,
+			thinking:  &llm.ThinkingConfig{Enabled: true, BudgetTokens: 4000},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			got, ok := resolveThinkingBudget(tt.maxTokens, tt.thinking)
+
+			assert.Equal(t, tt.wantOK, ok)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestBuildParamsNormalizesThinkingBudget(t *testing.T) {
+	t.Parallel()
+
+	maxTokens := 32768
+	params, err := buildParams(&llm.ChatCompletionRequest{
+		Model: "claude-sonnet-test",
+		Messages: []llm.Message{
+			{
+				Role:  llm.RoleUser,
+				Parts: []llm.Part{llm.TextPart{Text: "assess this vendor"}},
+			},
+		},
+		MaxTokens: &maxTokens,
+		Thinking: &llm.ThinkingConfig{
+			Enabled:      true,
+			BudgetTokens: 40000,
+		},
+	})
+	require.NoError(t, err)
+
+	budgetTokens := params.Thinking.GetBudgetTokens()
+	require.NotNil(t, budgetTokens)
+	assert.Equal(t, int64(31744), *budgetTokens)
+	assert.Less(t, *budgetTokens, params.MaxTokens)
+}
+
+func TestBuildParamsOmitsInvalidThinkingBudget(t *testing.T) {
+	t.Parallel()
+
+	maxTokens := 1024
+	params, err := buildParams(&llm.ChatCompletionRequest{
+		Model: "claude-sonnet-test",
+		Messages: []llm.Message{
+			{
+				Role:  llm.RoleUser,
+				Parts: []llm.Part{llm.TextPart{Text: "assess this vendor"}},
+			},
+		},
+		MaxTokens: &maxTokens,
+		Thinking: &llm.ThinkingConfig{
+			Enabled:      true,
+			BudgetTokens: 4000,
+		},
+	})
+	require.NoError(t, err)
+
+	assert.Nil(t, params.Thinking.GetBudgetTokens())
+}