fix(proxy): strip empty optional fields from streamed tool inputs

gpt-5.5 occasionally emits unused optional fields (e.g. "pages":"" on
the Read tool) when generating function calls. The proxy was forwarding
each input_json_delta as-is, so the polluted input reached claudecode
and the model in subsequent turns kept rationalizing "the tool layer
brought in empty pages" — issue james-6-23#95.

Buffer function_call_arguments deltas until output_item.done, drop
top-level keys whose value is "" or null, then emit the cleaned JSON
as a single input_json_delta before content_block_stop. Also apply the
same cleanup when reconstructing tool_use blocks from non-streaming
response.completed payloads and when re-encoding historical assistant
tool_use blocks back into upstream function_call items.

Empty objects/arrays and whitespace strings are preserved (some tools
treat those as meaningful). Trade-off: claudecode no longer sees tool
arguments stream char-by-char, but tool JSON is small enough that the
UX delta is imperceptible.

Author: james-6-23

proxy/anthropic.go

@@ -359,7 +359,9 @@ func appendAssistantBlocks(input []any, blocks []anthropicContentBlock) []any {
 			}
 			args := "{}"
 			if len(b.Input) > 0 {
-				args = string(b.Input)
+				if cleaned := sanitizeToolInputJSON(string(b.Input)); cleaned != "" {
+					args = cleaned
+				}
 			}
 			input = append(input, map[string]any{
 				"type":      "function_call",
@@ -489,18 +491,19 @@ func convertAnthropicToolChoice(raw json.RawMessage) any {
 
 // anthropicStreamTranslator 有状态的流式响应翻译器（Codex → Anthropic）
 type anthropicStreamTranslator struct {
-	model              string
-	responseID         string
-	messageStartSent   bool
-	contentBlockIndex  int
-	contentBlockOpen   bool
-	currentBlockType   string // "text" | "thinking" | "tool_use"
-	currentToolUseID   string
-	currentToolUseName string
-	hasToolUse         bool
-	inputTokens        int
-	outputTokens       int
-	cachedTokens       int
+	model                   string
+	responseID              string
+	messageStartSent        bool
+	contentBlockIndex       int
+	contentBlockOpen        bool
+	currentBlockType        string // "text" | "thinking" | "tool_use"
+	currentToolUseID        string
+	currentToolUseName      string
+	currentToolInputBuffer  strings.Builder
+	hasToolUse              bool
+	inputTokens             int
+	outputTokens            int
+	cachedTokens            int
 }
 
 // newAnthropicStreamTranslator 创建流式翻译器
@@ -708,22 +711,17 @@ func (t *anthropicStreamTranslator) handleThinkingDelta(data []byte) []anthropic
 	return events
 }
 
-// handleToolInputDelta 处理工具调用参数增量
+// handleToolInputDelta 缓冲工具调用参数增量。
+// 不直接转发为 input_json_delta：上游模型偶尔会塞入空可选字段（如 gpt-5.5
+// 给 Read 工具加 "pages":""），逐片透传后下游会看到污染后的入参。统一在
+// closeCurrentBlock 时整段清洗后一次性下发。
 func (t *anthropicStreamTranslator) handleToolInputDelta(data []byte) []anthropicStreamEvent {
 	delta := gjson.GetBytes(data, "delta").String()
 	if delta == "" {
 		return nil
 	}
-
-	idx := t.contentBlockIndex - 1
-	return []anthropicStreamEvent{{
-		Type:  "content_block_delta",
-		Index: &idx,
-		Delta: &anthropicDelta{
-			Type:        "input_json_delta",
-			PartialJSON: delta,
-		},
-	}}
+	t.currentToolInputBuffer.WriteString(delta)
+	return nil
 }
 
 // handleContentDone 处理内容完成（文本/推理块）
@@ -803,17 +801,70 @@ func (t *anthropicStreamTranslator) handleFailed() []anthropicStreamEvent {
 	return events
 }
 
-// closeCurrentBlock 关闭当前打开的 content block
+// closeCurrentBlock 关闭当前打开的 content block。
+// 关闭 tool_use 块时会先把累积的 arguments JSON 整段清洗（删除空字符串/null
+// 的可选字段），再作为单次 input_json_delta 下发。
 func (t *anthropicStreamTranslator) closeCurrentBlock() []anthropicStreamEvent {
 	if !t.contentBlockOpen {
 		return nil
 	}
 	t.contentBlockOpen = false
 	idx := t.contentBlockIndex - 1
-	return []anthropicStreamEvent{{
+
+	var events []anthropicStreamEvent
+	if t.currentBlockType == "tool_use" && t.currentToolInputBuffer.Len() > 0 {
+		cleaned := sanitizeToolInputJSON(t.currentToolInputBuffer.String())
+		if cleaned != "" {
+			events = append(events, anthropicStreamEvent{
+				Type:  "content_block_delta",
+				Index: &idx,
+				Delta: &anthropicDelta{
+					Type:        "input_json_delta",
+					PartialJSON: cleaned,
+				},
+			})
+		}
+		t.currentToolInputBuffer.Reset()
+	}
+
+	events = append(events, anthropicStreamEvent{
 		Type:  "content_block_stop",
 		Index: &idx,
-	}}
+	})
+	return events
+}
+
+// sanitizeToolInputJSON 清洗工具调用 arguments JSON：
+// 仅删除顶层值为空字符串("")或 null 的字段。
+// 不动空对象 {} / 空数组 []（部分工具语义上允许这两者）。
+// 上游 gpt-5.5 偶尔会给 Read 工具加 "pages":""，导致 claudecode 看到的入参
+// 带上无效空字段，模型在后续轮次里反复纠结"工具层带入了空 pages"。在代理
+// 层统一清掉，比让客户端去兼容更简单。
+func sanitizeToolInputJSON(raw string) string {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return raw
+	}
+	var obj map[string]json.RawMessage
+	if err := json.Unmarshal([]byte(raw), &obj); err != nil {
+		return raw
+	}
+	changed := false
+	for k, v := range obj {
+		s := strings.TrimSpace(string(v))
+		if s == `""` || s == "null" {
+			delete(obj, k)
+			changed = true
+		}
+	}
+	if !changed {
+		return raw
+	}
+	out, err := json.Marshal(obj)
+	if err != nil {
+		return raw
+	}
+	return string(out)
 }
 
 // finalize 在流结束时补齐缺失的事件
@@ -907,7 +958,9 @@ func buildAnthropicResponseFromCompleted(completedData []byte, model string) *an
 			callID := fromCodexCallID(item.Get("call_id").String())
 			name := item.Get("name").String()
 			args := item.Get("arguments").String()
-			if args == "" {
+			if cleaned := sanitizeToolInputJSON(args); cleaned != "" {
+				args = cleaned
+			} else {
 				args = "{}"
 			}
 			content = append(content, anthropicContentBlock{