Merge pull request #179 from kael-odin/fix/anthropic-thinking-disable

fix(anthropic): apply thinking config on override path, symmetric with openai

internal/backend/agent/model/anthropic.go

@@ -274,18 +274,17 @@ func (adapter *AnthropicAdapter) Stream(ctx context.Context, req StreamRequest,
 			body["tools"] = tools
 		}
 		body["system"] = anthropicProviderSystemBlocks(systemParts)
-		if thinkingConfig != nil {
-			body["thinking"] = thinkingConfig
-			if normalizeRuntimeThinkingEffort(req.ThinkingEffort) != "disabled" {
-				body["output_config"] = buildAnthropicOutputConfig(req)
-			}
-		}
 		frontier := buildAnthropicCacheFrontier(body, stableMessageCount)
 		req.RequestKnobs = annotateAnthropicRequestKnobs(req.RequestKnobs, body, frontier)
 		body = cloneRequestBodyOverride(body)
 		applyAnthropicCacheBreakpoints(body, frontier.BreakpointPositions)
 		frontier.BreakpointCount = len(frontier.BreakpointPositions)
 	}
+	// applyAnthropicThinkingConfig 在 override 块之外无条件调用，确保 RequestBodyOverride
+	// 路径与正常构造路径行为一致：disabled 时强制 thinking:{type:disabled} 并清理冲突字段，
+	// 非 disabled 时按 AnthropicThinkingEffort 写 adaptive 配置。与 openai.go 的
+	// applyOpenAIThinkingDisable 对称——后者也是无条件在两条路径之后调用。
+	applyAnthropicThinkingConfig(body, req)
 	if err := ApplyAnthropicExtraParams(body, req.AnthropicExtraParamsEnabled, req.AnthropicExtraParamsJSON); err != nil {
 		finishedAt = time.Now().UTC()
 		recordLLMSummaryArtifact(req, buildLLMSummaryPayload(req, "anthropic", modelID, startedAt, time.Time{}, finishedAt, "", 0, 0, 0, 0, err))
@@ -1441,6 +1440,34 @@ func buildAnthropicThinkingConfig(req StreamRequest) map[string]any {
 	}
 }
 
+// applyAnthropicThinkingConfig 在请求体构造完成后（含 RequestBodyOverride 路径）无条件调用，
+// 与 openai.go 的 applyOpenAIThinkingDisable 对称。它把 thinking 配置写入 body 并在 disabled
+// 时清理与之冲突的字段，确保两条构造路径行为一致：
+//   - disabled: 强制 thinking:{type:"disabled"}，删除 output_config / 残留 thinking adaptive 配置，
+//     记录 thinking_disabled_provider_param=thinking.type knob
+//   - adaptive: 按 AnthropicThinkingEffort 写 thinking:{type:adaptive,display:summarized} + output_config
+//
+// 在 override 路径下，上层若已在 override body 里塞了 thinking/output_config，disabled 时会被正确覆盖。
+func applyAnthropicThinkingConfig(body map[string]any, req StreamRequest) {
+	if len(body) == 0 {
+		return
+	}
+	if normalizeRuntimeThinkingEffort(req.ThinkingEffort) != "disabled" {
+		if strings.TrimSpace(req.AnthropicThinkingEffort) == "" {
+			return
+		}
+		body["thinking"] = map[string]any{
+			"type":    "adaptive",
+			"display": "summarized",
+		}
+		body["output_config"] = buildAnthropicOutputConfig(req)
+		return
+	}
+	body["thinking"] = map[string]any{"type": "disabled"}
+	delete(body, "output_config")
+	setRequestKnob(req, "thinking_disabled_provider_param", "thinking.type")
+}
+
 func buildAnthropicOutputConfig(req StreamRequest) map[string]any {
 	return map[string]any{
 		"effort": anthropicThinkingEffort(req),

internal/backend/agent/model/anthropic_thinking_disable_test.go

@@ -0,0 +1,141 @@
+package modeladapter
+
+import (
+	"testing"
+)
+
+// TestApplyAnthropicThinkingConfig covers the three effort branches on a body
+// built by the normal construction path, plus the override path symmetry that
+// was the original bug (thinking config was only applied inside the
+// `if len(body)==0` block, so RequestBodyOverride skipped it entirely).
+func TestApplyAnthropicThinkingConfig(t *testing.T) {
+	tests := []struct {
+		name           string
+		body           map[string]any
+		thinkingEffort string
+		adaptiveEffort string
+		wantThinking   any
+		wantOutputCfg  bool // expect output_config key present
+	}{
+		{
+			name:           "disabled_writes_disabled_and_drops_output_config",
+			body:           map[string]any{"model": "m", "output_config": map[string]any{"effort": "high"}},
+			thinkingEffort: "disabled",
+			wantThinking:   map[string]any{"type": "disabled"},
+			wantOutputCfg:  false,
+		},
+		{
+			name:           "adaptive_writes_adaptive_and_output_config",
+			body:           map[string]any{"model": "m"},
+			adaptiveEffort: "high",
+			wantThinking:   map[string]any{"type": "adaptive", "display": "summarized"},
+			wantOutputCfg:  true,
+		},
+		{
+			name:           "empty_effort_no_op",
+			body:           map[string]any{"model": "m"},
+			wantThinking:   nil,
+			wantOutputCfg:  false,
+		},
+		{
+			name:           "disabled_overrides_existing_adaptive_thinking",
+			body:           map[string]any{"thinking": map[string]any{"type": "adaptive", "display": "summarized"}, "output_config": map[string]any{"effort": "high"}},
+			thinkingEffort: "disabled",
+			wantThinking:   map[string]any{"type": "disabled"},
+			wantOutputCfg:  false,
+		},
+		{
+			name:           "disabled_normalizes_aliases",
+			body:           map[string]any{"model": "m"},
+			thinkingEffort: "off",
+			wantThinking:   map[string]any{"type": "disabled"},
+			wantOutputCfg:  false,
+		},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			req := StreamRequest{
+				ThinkingEffort:         tc.thinkingEffort,
+				AnthropicThinkingEffort: tc.adaptiveEffort,
+				RequestKnobs:           map[string]any{},
+			}
+			applyAnthropicThinkingConfig(tc.body, req)
+
+			gotThinking, hasThinking := tc.body["thinking"]
+			if tc.wantThinking == nil {
+				if hasThinking {
+					t.Fatalf("expected no thinking key, got %v", gotThinking)
+				}
+			} else {
+				if !hasThinking {
+					t.Fatalf("expected thinking=%v, key absent", tc.wantThinking)
+				}
+				if !mapEqual(gotThinking, tc.wantThinking) {
+					t.Fatalf("thinking mismatch\nwant: %v\ngot:  %v", tc.wantThinking, gotThinking)
+				}
+			}
+
+			_, hasOutputCfg := tc.body["output_config"]
+			if hasOutputCfg != tc.wantOutputCfg {
+				t.Fatalf("output_config presence=%v, want %v", hasOutputCfg, tc.wantOutputCfg)
+			}
+
+			if tc.thinkingEffort == "disabled" || tc.thinkingEffort == "off" {
+				gotKnob := req.RequestKnobs["thinking_disabled_provider_param"]
+				if gotKnob != "thinking.type" {
+					t.Fatalf("knob thinking_disabled_provider_param=%v, want thinking.type", gotKnob)
+				}
+			}
+		})
+	}
+}
+
+// TestApplyAnthropicThinkingConfigOverridePath simulates the RequestBodyOverride
+// branch: the adapter receives a body that did NOT go through the normal
+// construction (no tools/messages/system written by the adapter). Before the
+// fix, thinking config was skipped entirely on this path. Now
+// applyAnthropicThinkingConfig runs unconditionally and disables thinking.
+func TestApplyAnthropicThinkingConfigOverridePath(t *testing.T) {
+	overrideBody := map[string]any{
+		"model":         "claude-x",
+		"messages":      []any{map[string]any{"role": "user", "content": "hi"}},
+		"thinking":      map[string]any{"type": "adaptive", "display": "summarized"},
+		"output_config": map[string]any{"effort": "high"},
+		"stream":        true,
+	}
+	req := StreamRequest{
+		ThinkingEffort: "disabled",
+		RequestKnobs:   map[string]any{},
+	}
+
+	applyAnthropicThinkingConfig(overrideBody, req)
+
+	gotThinking := overrideBody["thinking"]
+	if !mapEqual(gotThinking, map[string]any{"type": "disabled"}) {
+		t.Fatalf("override path: thinking not forced to disabled, got %v", gotThinking)
+	}
+	if _, hasOutputCfg := overrideBody["output_config"]; hasOutputCfg {
+		t.Fatalf("override path: output_config should be dropped on disabled, still present")
+	}
+	if gotKnob := req.RequestKnobs["thinking_disabled_provider_param"]; gotKnob != "thinking.type" {
+		t.Fatalf("override path: knob=%v, want thinking.type", gotKnob)
+	}
+}
+
+func mapEqual(a, b any) bool {
+	ma, okA := a.(map[string]any)
+	mb, okB := b.(map[string]any)
+	if !okA || !okB {
+		return a == b
+	}
+	if len(ma) != len(mb) {
+		return false
+	}
+	for k, va := range ma {
+		vb, ok := mb[k]
+		if !ok || !mapEqual(va, vb) {
+			return false
+		}
+	}
+	return true
+}