修复裸 invoke 标签导致流式输出卡住的问题

当模型输出不带 <tool_calls> 包裹的裸 <invoke> 标签时（如文档示例或行内引用），之前的逻辑会一直等待关闭标签导致流式输出卡住。现在通过 shouldKeepBareInvokeCapture 判断是否为可修复的调用，不可修复的直接作为文本释放。同时更新 README 文档中工具调用解析说明，并移除 allow-list 过滤的过时描述。 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-02 07:25:26 +08:00 · 2026-04-26 11:14:26 +08:00
parent f6df01d3aa
commit 3627c7366d
6 changed files with 117 additions and 4 deletions
--- a/README.MD
+++ b/README.MD
@@ -4,11 +4,14 @@

 # DS2API

+<a href="https://trendshift.io/repositories/24508" target="_blank"><img src="https://trendshift.io/api/badge/repositories/24508" alt="CJackHwang%2Fds2api | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+
 [![License](https://img.shields.io/github/license/CJackHwang/ds2api.svg)](LICENSE)
 ![Stars](https://img.shields.io/github/stars/CJackHwang/ds2api.svg)
 ![Forks](https://img.shields.io/github/forks/CJackHwang/ds2api.svg)
 [![Release](https://img.shields.io/github/v/release/CJackHwang/ds2api?display_name=tag)](https://github.com/CJackHwang/ds2api/releases)
 [![Docker](https://img.shields.io/badge/docker-ready-blue.svg)](docs/DEPLOY.md)
+
 [![Deploy on Zeabur](https://zeabur.com/button.svg)](https://zeabur.com/templates/L4CFHP)
 [![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/CJackHwang/ds2api)

@@ -317,9 +320,9 @@ Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `
 4. `responses` 支持并执行 `tool_choice`（`auto`/`none`/`required`/强制函数）；`required` 违规时非流式返回 `422`，流式返回 `response.failed`
 5. 客户端请求哪种协议，就按该协议返回工具调用（OpenAI/Claude/Gemini 各自原生结构）；模型侧优先约束输出规范 XML，再由兼容层转译

-> 说明：当前版本在 parser 层仍以“尽量解析成功”为优先，未启用基于 allow-list 的工具名硬拒绝。
+> 说明：当前版本 parser 层以”尽量解析成功”为优先，所有格式合法的 XML 工具调用都会通过，不做工具名 allow-list 过滤。
 >
-> 想评估“把工具调用封装成 XML 再输入模型”的方案，可参考：`docs/toolcall-semantics.md`。
+> 想评估”把工具调用封装成 XML 再输入模型”的方案，可参考：`docs/toolcall-semantics.md`。

 ## 本地开发抓包工具

--- a/README.en.md
+++ b/README.en.md
@@ -4,6 +4,8 @@

 # DS2API

+<a href="https://trendshift.io/repositories/24508" target="_blank"><img src="https://trendshift.io/api/badge/repositories/24508" alt="CJackHwang%2Fds2api | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+
 [![License](https://img.shields.io/github/license/CJackHwang/ds2api.svg)](LICENSE)
 ![Stars](https://img.shields.io/github/stars/CJackHwang/ds2api.svg)
 ![Forks](https://img.shields.io/github/forks/CJackHwang/ds2api.svg)
--- a/2
+++ b/2
@@ -1 +1 @@
-4.0.0
+4.0.1
--- a/internal/toolstream/tool_sieve_core.go
+++ b/internal/toolstream/tool_sieve_core.go
@@ -193,5 +193,8 @@ func consumeToolCapture(state *State, toolNames []string) (prefix string, calls
 	if hasOpenXMLToolTag(captured) {
 		return "", nil, "", false
 	}
+	if shouldKeepBareInvokeCapture(captured) {
 		return "", nil, "", false
 	}
+	return captured, nil, "", true
+}
--- a/internal/toolstream/tool_sieve_xml.go
+++ b/internal/toolstream/tool_sieve_xml.go
@@ -90,6 +90,38 @@ func hasOpenXMLToolTag(captured string) bool {
 	return false
 }

+func shouldKeepBareInvokeCapture(captured string) bool {
+	lower := strings.ToLower(captured)
+	invokeIdx := strings.Index(lower, "<invoke")
+	if invokeIdx < 0 || strings.Contains(lower, "<tool_calls") {
+		return false
+	}
+	if findXMLCloseOutsideCDATA(captured, "</tool_calls>", invokeIdx) > invokeIdx {
+		return true
+	}
+
+	startEnd := findXMLTagEnd(captured, invokeIdx+len("<invoke"))
+	if startEnd < 0 {
+		return true
+	}
+	body := captured[startEnd+1:]
+	trimmedBody := strings.TrimLeft(body, " \t\r\n")
+	if trimmedBody == "" {
+		return true
+	}
+
+	invokeCloseIdx := findXMLCloseOutsideCDATA(captured, "</invoke>", startEnd+1)
+	if invokeCloseIdx >= 0 {
+		afterClose := captured[invokeCloseIdx+len("</invoke>"):]
+		return strings.TrimSpace(afterClose) == ""
+	}
+
+	trimmedLower := strings.ToLower(trimmedBody)
+	return strings.HasPrefix(trimmedLower, "<parameter") ||
+		strings.HasPrefix(trimmedLower, "{") ||
+		strings.HasPrefix(trimmedLower, "[")
+}
+
 func findXMLCloseOutsideCDATA(s, closeTag string, start int) int {
 	if s == "" || closeTag == "" {
 		return -1
@@ -122,6 +154,27 @@ func findXMLCloseOutsideCDATA(s, closeTag string, start int) int {
 	return -1
 }

+func findXMLTagEnd(s string, start int) int {
+	quote := byte(0)
+	for i := start; i < len(s); i++ {
+		ch := s[i]
+		if quote != 0 {
+			if ch == quote {
+				quote = 0
+			}
+			continue
+		}
+		if ch == '"' || ch == '\'' {
+			quote = ch
+			continue
+		}
+		if ch == '>' {
+			return i
+		}
+	}
+	return -1
+}
+
 // findPartialXMLToolTagStart checks if the string ends with a partial canonical
 // XML wrapper tag (e.g., "<too") and returns the position of the '<'.
 func findPartialXMLToolTagStart(s string) int {
--- a/internal/toolstream/tool_sieve_xml_test.go
+++ b/internal/toolstream/tool_sieve_xml_test.go
@@ -567,6 +567,58 @@ func TestProcessToolSievePassesThroughBareToolCallAsText(t *testing.T) {
 	}
 }

+func TestProcessToolSieveBareInvokeInlineProseDoesNotStall(t *testing.T) {
+	var state State
+	chunk := "Use `<invoke name=\"read_file\">` as plain documentation text."
+	events := ProcessChunk(&state, chunk, []string{"read_file"})
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if toolCalls != 0 {
+		t.Fatalf("expected inline invoke prose to remain text, got %d events=%#v", toolCalls, events)
+	}
+	if textContent.String() != chunk {
+		t.Fatalf("expected inline invoke prose to stream immediately, got %q", textContent.String())
+	}
+	if state.capturing {
+		t.Fatal("expected inline invoke prose not to leave stream capture open")
+	}
+}
+
+func TestProcessToolSieveBareInvokeExampleReleasesWhenNotRepairable(t *testing.T) {
+	var state State
+	chunks := []string{
+		`Example: <invoke name="read_file"><parameter name="path">README.md</parameter>`,
+		"</invoke> then continue.",
+	}
+	var events []Event
+	for _, c := range chunks {
+		events = append(events, ProcessChunk(&state, c, []string{"read_file"})...)
+	}
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if toolCalls != 0 {
+		t.Fatalf("expected non-repairable bare invoke to remain text, got %d events=%#v", toolCalls, events)
+	}
+	if textContent.String() != strings.Join(chunks, "") {
+		t.Fatalf("expected non-repairable bare invoke to pass through, got %q", textContent.String())
+	}
+	if state.capturing {
+		t.Fatal("expected non-repairable bare invoke not to leave stream capture open")
+	}
+}
+
 func TestProcessToolSieveRepairsMissingOpeningWrapperWithoutLeakingInvokeText(t *testing.T) {
 	var state State
 	chunks := []string{
@@ -1 +1 @@
 .0.0
 .0.1