Merge pull request #90 from CJackHwang/dev

Merge pull request #89 from CJackHwang/codex/review-changes-in-pull-request-#88

Support text-kv `function.name`/`function.arguments` fallback and looser name matching

API.md

@@ -284,6 +284,11 @@ data: [DONE]
 
 **流式**：命中高置信特征后立即输出 `delta.tool_calls`（不等待完整 JSON 闭合），并持续发送 arguments 增量；已确认的 toolcall 原始 JSON 不会回流到 `delta.content`。
 
+补充说明：
+
+- **非代码块上下文**下，工具 JSON 即使与普通文本混合，也会按特征识别并产出可执行 tool call（前后普通文本仍可透传）。
+- Markdown fenced code block（例如 ```json ... ```）中的 `tool_calls` 仅视为示例文本，不会被执行。
+
 ---
 
 ### `GET /v1/models/{id}`
@@ -301,7 +306,7 @@ OpenAI Responses 风格接口，兼容 `input` 或 `messages`。
 | `messages` | array | ❌ | 与 `input` 二选一 |
 | `instructions` | string | ❌ | 自动前置为 system 消息 |
 | `stream` | boolean | ❌ | 默认 `false` |
-| `tools` | array | ❌ | 与 chat 同样的工具识别与转译策略 |
+| `tools` | array | ❌ | 与 chat 同样的工具识别与转译策略（含代码块示例豁免） |
 | `tool_choice` | string/object | ❌ | 支持 `auto`/`none`/`required` 与强制函数（`{"type":"function","name":"..."}`） |
 
 **非流式响应**：返回标准 `response` 对象，`id` 形如 `resp_xxx`，并写入内存 TTL 存储。

README.MD

@@ -106,6 +106,14 @@ flowchart LR
 可通过配置中的 `claude_mapping` 或 `claude_model_mapping` 覆盖映射关系。
 另外，`/anthropic/v1/models` 现已包含 Claude 1.x/2.x/3.x/4.x 历史模型 ID 与常见别名，便于旧客户端直接兼容。
 
+
+#### Claude Code 接入避坑（实测）
+
+- `ANTHROPIC_BASE_URL` 推荐直接指向 DS2API 根地址（例如 `http://127.0.0.1:5001`），Claude Code 会请求 `/v1/messages?beta=true`。
+- `ANTHROPIC_API_KEY` 需要与 `config.json` 中 `keys` 一致；建议同时保留常规 key 与 `sk-ant-*` 形态 key，兼容不同客户端校验习惯。
+- 若系统设置了代理，建议对 DS2API 地址配置 `NO_PROXY=127.0.0.1,localhost,<你的主机IP>`，避免本地回环请求被代理拦截。
+- 如遇“工具调用输出成文本、未执行”问题，请升级到包含 Claude 工具调用多格式解析（JSON/XML/ANTML/invoke）的版本。
+
 ### Gemini 接口
 
 Gemini 适配器将模型名通过 `model_aliases` 或内置规则映射到 DeepSeek 原生模型，支持 `generateContent` 和 `streamGenerateContent` 两种调用方式，并完整支持 Tool Calling（`functionDeclarations` → `functionCall` 输出）。

README.en.md

@@ -106,6 +106,14 @@ flowchart LR
 Override mapping via `claude_mapping` or `claude_model_mapping` in config.
 In addition, `/anthropic/v1/models` now includes historical Claude 1.x/2.x/3.x/4.x IDs and common aliases for legacy client compatibility.
 
+
+#### Claude Code integration pitfalls (validated)
+
+- Set `ANTHROPIC_BASE_URL` to the DS2API root URL (for example `http://127.0.0.1:5001`). Claude Code sends requests to `/v1/messages?beta=true`.
+- `ANTHROPIC_API_KEY` must match an entry in `keys` from `config.json`. Keeping both a regular key and an `sk-ant-*` style key improves client compatibility.
+- If your environment has proxy variables, set `NO_PROXY=127.0.0.1,localhost,<your_host_ip>` for DS2API to avoid proxy interception of local traffic.
+- If tool calls are rendered as plain text and not executed, upgrade to a build that includes multi-format Claude tool-call parsing (JSON/XML/ANTML/invoke).
+
 ### Gemini Endpoint
 
 The Gemini adapter maps model names to DeepSeek native models via `model_aliases` or built-in heuristics, supporting both `generateContent` and `streamGenerateContent` call patterns with full Tool Calling support (`functionDeclarations` → `functionCall` output).
@@ -351,6 +359,7 @@ Queue limit = DS2API_ACCOUNT_MAX_QUEUE (default = recommended concurrency)
 When `tools` is present in the request, DS2API performs anti-leak handling:
 
 1. Toolcall feature matching is enabled only in **non-code-block context** (fenced examples are ignored)
+   - In non-code-block context, tool JSON may still be recognized even when mixed with normal prose; surrounding prose can remain as text output.
 2. `responses` streaming strictly uses official item lifecycle events (`response.output_item.*`, `response.content_part.*`, `response.function_call_arguments.*`)
 3. Tool names not declared in the `tools` schema are strictly rejected and will not be emitted as valid tool calls
 4. `responses` supports and enforces `tool_choice` (`auto`/`none`/`required`/forced function); `required` violations return `422` for non-stream and `response.failed` for stream

TESTING.md

@@ -51,7 +51,7 @@ DS2API 提供两个层级的测试：
 1. **Preflight 检查**：
    - `go test ./... -count=1`（单元测试）
    - `./tests/scripts/check-node-split-syntax.sh`（Node 拆分模块语法门禁）
-   - `node --test api/helpers/stream-tool-sieve.test.js api/chat-stream.test.js api/compat/js_compat_test.js`（Node 流式拦截 + compat 单测）
+   - `node --test`（如仓库存在 Node 单测文件时执行；当前默认以 Go 测试 + Node 语法门禁为主）
    - `npm run build --prefix webui`（WebUI 构建检查）
 
 2. **隔离启动**：复制 `config.json` 到临时目录，启动独立服务进程

docs/toolcall-semantics.md

@@ -0,0 +1,41 @@
+# Tool call parsing semantics (Go canonical spec)
+
+This document defines the cross-runtime contract for `ParseToolCallsDetailed` / `parseToolCallsDetailed`.
+
+## Output contract
+
+- `calls`: accepted tool calls with normalized tool names.
+- `sawToolCallSyntax`: true when tool-call-like syntax is detected (`tool_calls`, `<tool_call>`, `<function_call>`, `<invoke>`) or a valid call is parsed.
+- `rejectedByPolicy`: true when parser extracted call syntax but all calls are rejected by allow-list policy.
+- `rejectedToolNames`: de-duplicated rejected tool names in first-seen order.
+
+## Parse pipeline
+
+1. Strip fenced code blocks for non-standalone parsing.
+2. Build candidates from:
+   - full text,
+   - fenced JSON snippets,
+   - extracted JSON objects around `tool_calls`,
+   - first `{` to last `}` object slice.
+3. Parse each candidate in order:
+   - JSON payload parser (`tool_calls`, list, single call object),
+   - XML/Markup parser (`<tool_call>`, `<function_call>`, `<invoke>`; supports attributes + nested fields),
+   - Text KV fallback parser (`function.name: <name>` ... `function.arguments: {json}`).
+4. Stop at first candidate that yields at least one call.
+
+## Name normalization policy
+
+When matching parsed names against configured tools:
+
+1. exact match,
+2. case-insensitive match,
+3. namespace tail match (`a.b.c` => `c`),
+4. loose alnum match (remove non `[a-z0-9]`, compare).
+
+## Standalone mode
+
+Standalone mode (`ParseStandaloneToolCallsDetailed`) parses the whole input directly (no candidate slicing), while still applying:
+
+- example-context guard,
+- JSON then markup fallback,
+- the same allow-list normalization policy.

internal/adapter/claude/handler_stream_test.go

@@ -315,3 +315,78 @@ func asString(v any) string {
 	s, _ := v.(string)
 	return s
 }
+
+func TestHandleClaudeStreamRealtimeToolSafetyAcrossStructuredFormats(t *testing.T) {
+	tests := []struct {
+		name    string
+		payload string
+	}{
+		{name: "xml_tool_call", payload: `<tool_call><tool_name>Bash</tool_name><parameters><command>pwd</command></parameters></tool_call>`},
+		{name: "xml_json_tool_call", payload: `<tool_call>{"tool":"Bash","params":{"command":"pwd"}}</tool_call>`},
+		{name: "nested_tool_tag_style", payload: `<tool_call><tool name="Bash"><command>pwd</command></tool></tool_call>`},
+		{name: "function_tag_style", payload: `<function_call>Bash</function_call><function parameter name="command">pwd</function parameter>`},
+		{name: "antml_argument_style", payload: `<antml:function_calls><antml:function_call id="1" name="Bash"><antml:argument name="command">pwd</antml:argument></antml:function_call></antml:function_calls>`},
+		{name: "antml_function_attr_parameters", payload: `<antml:function_calls><antml:function_call id="1" function="Bash"><antml:parameters>{"command":"pwd"}</antml:parameters></antml:function_call></antml:function_calls>`},
+		{name: "invoke_parameter_style", payload: `<function_calls><invoke name="Bash"><parameter name="command">pwd</parameter></invoke></function_calls>`},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			h := &Handler{}
+			resp := makeClaudeSSEHTTPResponse(
+				`data: {"p":"response/content","v":"`+strings.ReplaceAll(tc.payload, `"`, `\"`)+`"}`,
+				`data: [DONE]`,
+			)
+			rec := httptest.NewRecorder()
+			req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
+
+			h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "use tool"}}, false, false, []string{"Bash"})
+
+			frames := parseClaudeFrames(t, rec.Body.String())
+			foundToolUse := false
+			for _, f := range findClaudeFrames(frames, "content_block_start") {
+				contentBlock, _ := f.Payload["content_block"].(map[string]any)
+				if contentBlock["type"] == "tool_use" {
+					foundToolUse = true
+					break
+				}
+			}
+			if !foundToolUse {
+				t.Fatalf("expected tool_use block for format %s, body=%s", tc.name, rec.Body.String())
+			}
+		})
+	}
+}
+
+func TestHandleClaudeStreamRealtimeDoesNotStopOnUnclosedFencedToolExample(t *testing.T) {
+	h := &Handler{}
+	resp := makeClaudeSSEHTTPResponse(
+		"data: {\"p\":\"response/content\",\"v\":\"Here is an example:\\n```json\\n{\\\"tool_calls\\\":[{\\\"name\\\":\\\"Bash\\\",\\\"input\\\":{\\\"command\\\":\\\"pwd\\\"}}]}\"}",
+		"data: {\"p\":\"response/content\",\"v\":\"\\n```\\nDo not execute it.\"}",
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
+
+	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "show example only"}}, false, false, []string{"Bash"})
+
+	frames := parseClaudeFrames(t, rec.Body.String())
+	for _, f := range findClaudeFrames(frames, "content_block_start") {
+		contentBlock, _ := f.Payload["content_block"].(map[string]any)
+		if contentBlock["type"] == "tool_use" {
+			t.Fatalf("unexpected tool_use for fenced example, body=%s", rec.Body.String())
+		}
+	}
+
+	foundEndTurn := false
+	for _, f := range findClaudeFrames(frames, "message_delta") {
+		delta, _ := f.Payload["delta"].(map[string]any)
+		if delta["stop_reason"] == "end_turn" {
+			foundEndTurn = true
+			break
+		}
+	}
+	if !foundEndTurn {
+		t.Fatalf("expected stop_reason=end_turn, body=%s", rec.Body.String())
+	}
+}

internal/adapter/claude/handler_util_test.go

@@ -125,8 +125,11 @@ func TestBuildClaudeToolPromptSingleTool(t *testing.T) {
 	if !containsStr(prompt, "Search the web") {
 		t.Fatalf("expected description in prompt")
 	}
-	if !containsStr(prompt, "tool_calls") {
-		t.Fatalf("expected tool_calls instruction in prompt")
+	if !containsStr(prompt, "tool_use") {
+		t.Fatalf("expected tool_use instruction in prompt")
+	}
+	if containsStr(prompt, "tool_calls") {
+		t.Fatalf("expected prompt to avoid tool_calls JSON instruction")
 	}
 }
 

internal/adapter/claude/handler_utils.go

@@ -51,7 +51,7 @@ func buildClaudeToolPrompt(tools []any) string {
 		parts = append(parts, fmt.Sprintf("Tool: %s\nDescription: %s\nParameters: %s", name, desc, schema))
 	}
 	parts = append(parts,
-		"When you need to use tools, you can call multiple tools in one response. Output ONLY JSON like {\"tool_calls\":[{\"name\":\"tool\",\"input\":{}}]}",
+		"When you need a tool, respond with Claude-native tool use (tool_use) using the provided tool schema. Do not print tool-call JSON in text.",
 		"History markers in conversation: [TOOL_CALL_HISTORY]...[/TOOL_CALL_HISTORY] are your previous tool calls; [TOOL_RESULT_HISTORY]...[/TOOL_RESULT_HISTORY] are runtime tool outputs, not user input.",
 		"After a valid [TOOL_RESULT_HISTORY], continue with final answer instead of repeating the same call unless required fields are still missing.",
 	)

internal/adapter/claude/stream_runtime_core.go

@@ -8,6 +8,7 @@ import (
 
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
+	"ds2api/internal/util"
 )
 
 type claudeStreamRuntime struct {
@@ -116,6 +117,18 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 
 		s.text.WriteString(p.Text)
 		if s.bufferToolContent {
+			if hasUnclosedCodeFence(s.text.String()) {
+				continue
+			}
+			detected := util.ParseToolCalls(s.text.String(), s.toolNames)
+			if len(detected) > 0 {
+				s.finalize("tool_use")
+				return streamengine.ParsedDecision{
+					ContentSeen: true,
+					Stop:        true,
+					StopReason:  streamengine.StopReason("tool_use_detected"),
+				}
+			}
 			continue
 		}
 		s.closeThinkingBlock()
@@ -144,3 +157,7 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 
 	return streamengine.ParsedDecision{ContentSeen: contentSeen}
 }
+
+func hasUnclosedCodeFence(text string) bool {
+	return strings.Count(text, "```")%2 == 1
+}

internal/adapter/gemini/handler_test.go

@@ -99,7 +99,7 @@ func TestGeminiRoutesRegistered(t *testing.T) {
 
 func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 	upstream := makeGeminiUpstreamResponse(
-		`data: {"p":"response/content","v":"我来调用工具\n{\"tool_calls\":[{\"name\":\"eval_javascript\",\"input\":{\"code\":\"1+1\"}}]}"}`,
+		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"eval_javascript\",\"input\":{\"code\":\"1+1\"}}]}"}`,
 		`data: [DONE]`,
 	)
 	h := &Handler{
@@ -143,6 +143,42 @@ func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 	}
 }
 
+func TestGenerateContentMixedToolSnippetAlsoTriggersFunctionCall(t *testing.T) {
+	upstream := makeGeminiUpstreamResponse(
+		`data: {"p":"response/content","v":"我来调用工具\n{\"tool_calls\":[{\"name\":\"eval_javascript\",\"input\":{\"code\":\"1+1\"}}]}"}`,
+		`data: [DONE]`,
+	)
+	h := &Handler{Store: testGeminiConfig{}, Auth: testGeminiAuth{}, DS: testGeminiDS{resp: upstream}}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	body := `{
+		"contents":[{"role":"user","parts":[{"text":"call tool"}]}],
+		"tools":[{"functionDeclarations":[{"name":"eval_javascript","description":"eval","parameters":{"type":"object","properties":{"code":{"type":"string"}}}}]}]
+	}`
+	req := httptest.NewRequest(http.MethodPost, "/v1beta/models/gemini-2.5-pro:generateContent", strings.NewReader(body))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v", err)
+	}
+	candidates, _ := out["candidates"].([]any)
+	c0, _ := candidates[0].(map[string]any)
+	content, _ := c0["content"].(map[string]any)
+	parts, _ := content["parts"].([]any)
+	part0, _ := parts[0].(map[string]any)
+	functionCall, _ := part0["functionCall"].(map[string]any)
+	if functionCall["name"] != "eval_javascript" {
+		t.Fatalf("expected functionCall name eval_javascript for mixed snippet, got %#v", functionCall)
+	}
+}
+
 func TestStreamGenerateContentEmitsSSE(t *testing.T) {
 	upstream := makeGeminiUpstreamResponse(
 		`data: {"p":"response/content","v":"hello "}`,

internal/adapter/openai/handler_toolcall_test.go

@@ -513,8 +513,8 @@ func TestHandleStreamToolCallMixedWithPlainTextSegments(t *testing.T) {
 	if !done {
 		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
 	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta in mixed prose stream, body=%s", rec.Body.String())
+	if !streamHasToolCallsDelta(frames) {
+		t.Fatalf("expected tool_calls delta in mixed prose stream, body=%s", rec.Body.String())
 	}
 	content := strings.Builder{}
 	for _, frame := range frames {
@@ -531,11 +531,8 @@ func TestHandleStreamToolCallMixedWithPlainTextSegments(t *testing.T) {
 	if !strings.Contains(got, "下面是示例：") || !strings.Contains(got, "请勿执行。") {
 		t.Fatalf("expected pre/post plain text to pass sieve, got=%q", got)
 	}
-	if !strings.Contains(strings.ToLower(got), `"tool_calls"`) {
-		t.Fatalf("expected embedded tool json to remain text in strict mode, got=%q", got)
-	}
-	if streamFinishReason(frames) != "stop" {
-		t.Fatalf("expected finish_reason=stop for mixed prose, body=%s", rec.Body.String())
+	if streamFinishReason(frames) != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls for mixed prose, body=%s", rec.Body.String())
 	}
 }
 
@@ -555,8 +552,8 @@ func TestHandleStreamToolCallAfterLeadingTextRemainsText(t *testing.T) {
 	if !done {
 		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
 	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta, body=%s", rec.Body.String())
+	if !streamHasToolCallsDelta(frames) {
+		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
 	}
 	content := strings.Builder{}
 	for _, frame := range frames {
@@ -573,11 +570,9 @@ func TestHandleStreamToolCallAfterLeadingTextRemainsText(t *testing.T) {
 	if !strings.Contains(got, "我将调用工具。") {
 		t.Fatalf("expected leading text to keep streaming, got=%q", got)
 	}
-	if !strings.Contains(strings.ToLower(got), "tool_calls") {
-		t.Fatalf("expected tool_calls example text preserved, got=%q", got)
-	}
-	if streamFinishReason(frames) != "stop" {
-		t.Fatalf("expected finish_reason=stop, body=%s", rec.Body.String())
+
+	if streamFinishReason(frames) != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
 	}
 }
 
@@ -596,8 +591,8 @@ func TestHandleStreamToolCallWithSameChunkTrailingTextRemainsText(t *testing.T)
 	if !done {
 		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
 	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta, body=%s", rec.Body.String())
+	if !streamHasToolCallsDelta(frames) {
+		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
 	}
 	content := strings.Builder{}
 	for _, frame := range frames {
@@ -614,8 +609,45 @@ func TestHandleStreamToolCallWithSameChunkTrailingTextRemainsText(t *testing.T)
 	if !strings.Contains(got, "接下来我会继续说明。") {
 		t.Fatalf("expected trailing plain text to be preserved, got=%q", got)
 	}
-	if !strings.Contains(strings.ToLower(got), "tool_calls") {
-		t.Fatalf("expected tool_calls example text preserved, got=%q", got)
+
+	if streamFinishReason(frames) != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
+	}
+}
+
+func TestHandleStreamFencedToolCallSnippetRemainsText(t *testing.T) {
+	h := &Handler{}
+	resp := makeSSEHTTPResponse(
+		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "下面是调用示例：\n```json\n"),
+		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}\n```\n仅示例，不要执行。"),
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+
+	h.handleStream(rec, req, resp, "cid7f", "deepseek-chat", "prompt", false, false, []string{"search"})
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	if streamHasToolCallsDelta(frames) {
+		t.Fatalf("did not expect tool_calls delta for fenced snippet, body=%s", rec.Body.String())
+	}
+	content := strings.Builder{}
+	for _, frame := range frames {
+		choices, _ := frame["choices"].([]any)
+		for _, item := range choices {
+			choice, _ := item.(map[string]any)
+			delta, _ := choice["delta"].(map[string]any)
+			if c, ok := delta["content"].(string); ok {
+				content.WriteString(c)
+			}
+		}
+	}
+	got := content.String()
+	if !strings.Contains(got, "```json") || !strings.Contains(strings.ToLower(got), "tool_calls") {
+		t.Fatalf("expected fenced tool snippet in content, got=%q", got)
 	}
 	if streamFinishReason(frames) != "stop" {
 		t.Fatalf("expected finish_reason=stop, body=%s", rec.Body.String())
@@ -640,8 +672,8 @@ func TestHandleStreamToolCallKeyAppearsLateRemainsText(t *testing.T) {
 	if !done {
 		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
 	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta, body=%s", rec.Body.String())
+	if !streamHasToolCallsDelta(frames) {
+		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
 	}
 	content := strings.Builder{}
 	for _, frame := range frames {
@@ -655,14 +687,11 @@ func TestHandleStreamToolCallKeyAppearsLateRemainsText(t *testing.T) {
 		}
 	}
 	got := content.String()
-	if !strings.Contains(strings.ToLower(got), "tool_calls") || !strings.Contains(got, "{") {
-		t.Fatalf("expected embedded tool json to remain in text, got=%q", got)
-	}
 	if !strings.Contains(got, "后置正文C。") {
 		t.Fatalf("expected stream to continue after tool json convergence, got=%q", got)
 	}
-	if streamFinishReason(frames) != "stop" {
-		t.Fatalf("expected finish_reason=stop, body=%s", rec.Body.String())
+	if streamFinishReason(frames) != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
 	}
 }
 

internal/adapter/openai/message_normalize.go

@@ -6,6 +6,7 @@ import (
 	"strings"
 
 	"ds2api/internal/config"
+	"ds2api/internal/prompt"
 )
 
 func normalizeOpenAIMessagesForPrompt(raw []any, traceID string) []map[string]any {
@@ -33,9 +34,9 @@ func normalizeOpenAIMessagesForPrompt(raw []any, traceID string) []map[string]an
 				"role":    "user",
 				"content": formatToolResultForPrompt(msg),
 			})
-		case "user", "system":
+		case "user", "system", "developer":
 			out = append(out, map[string]any{
-				"role":    role,
+				"role":    normalizeOpenAIRoleForPrompt(role),
 				"content": normalizeOpenAIContentForPrompt(msg["content"]),
 			})
 		default:
@@ -47,7 +48,7 @@ func normalizeOpenAIMessagesForPrompt(raw []any, traceID string) []map[string]an
 				role = "user"
 			}
 			out = append(out, map[string]any{
-				"role":    role,
+				"role":    normalizeOpenAIRoleForPrompt(role),
 				"content": content,
 			})
 		}
@@ -77,7 +78,7 @@ func formatAssistantToolCallsForPrompt(msg map[string]any, traceID string) strin
 				args = normalizeOpenAIArgumentsForPrompt(fn["arguments"])
 			}
 			if name == "" {
-				name = "unknown"
+				continue
 			}
 			if args == "" {
 				args = normalizeOpenAIArgumentsForPrompt(call["arguments"])
@@ -132,32 +133,7 @@ func formatToolResultForPrompt(msg map[string]any) string {
 }
 
 func normalizeOpenAIContentForPrompt(v any) string {
-	switch x := v.(type) {
-	case string:
-		return x
-	case []any:
-		parts := make([]string, 0, len(x))
-		for _, item := range x {
-			m, ok := item.(map[string]any)
-			if !ok {
-				continue
-			}
-			t := strings.ToLower(strings.TrimSpace(asString(m["type"])))
-			if t != "text" && t != "output_text" && t != "input_text" {
-				continue
-			}
-			if text := asString(m["text"]); text != "" {
-				parts = append(parts, text)
-				continue
-			}
-			if text := asString(m["content"]); text != "" {
-				parts = append(parts, text)
-			}
-		}
-		return strings.Join(parts, "\n")
-	default:
-		return marshalToPromptString(v)
-	}
+	return prompt.NormalizeContent(v)
 }
 
 func normalizeOpenAIArgumentsForPrompt(v any) string {
@@ -189,6 +165,14 @@ func marshalToPromptString(v any) string {
 	return string(b)
 }
 
+func normalizeOpenAIRoleForPrompt(role string) string {
+	role = strings.ToLower(strings.TrimSpace(role))
+	if role == "developer" {
+		return "system"
+	}
+	return role
+}
+
 func asString(v any) string {
 	if s, ok := v.(string); ok {
 		return s

internal/adapter/openai/message_normalize_test.go

@@ -193,3 +193,90 @@ func TestNormalizeOpenAIMessagesForPrompt_PreservesConcatenatedToolArguments(t *
 		t.Fatalf("expected original concatenated arguments in tool history, got %q", content)
 	}
 }
+
+
+func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsMissingNameAreDropped(t *testing.T) {
+	raw := []any{
+		map[string]any{
+			"role": "assistant",
+			"tool_calls": []any{
+				map[string]any{
+					"id":   "call_missing_name",
+					"type": "function",
+					"function": map[string]any{
+						"arguments": `{"path":"README.MD"}`,
+					},
+				},
+			},
+		},
+	}
+
+	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
+	if len(normalized) != 0 {
+		t.Fatalf("expected nameless assistant tool_calls to be dropped, got %#v", normalized)
+	}
+}
+
+func TestNormalizeOpenAIMessagesForPrompt_AssistantNilContentDoesNotInjectNullLiteral(t *testing.T) {
+	raw := []any{
+		map[string]any{
+			"role":    "assistant",
+			"content": nil,
+			"tool_calls": []any{
+				map[string]any{
+					"id": "call_screenshot",
+					"function": map[string]any{
+						"name":      "send_file_to_user",
+						"arguments": `{"file_path":"/tmp/a.png"}`,
+					},
+				},
+			},
+		},
+	}
+
+	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
+	if len(normalized) != 1 {
+		t.Fatalf("expected one normalized message, got %d", len(normalized))
+	}
+	content, _ := normalized[0]["content"].(string)
+	if strings.Contains(content, "<｜Assistant｜>null") || strings.HasPrefix(strings.TrimSpace(content), "null") {
+		t.Fatalf("unexpected null literal injected into assistant tool history: %q", content)
+	}
+	if !strings.Contains(content, "function.name: send_file_to_user") {
+		t.Fatalf("expected tool history block preserved, got %q", content)
+	}
+}
+
+func TestNormalizeOpenAIMessagesForPrompt_DeveloperRoleMapsToSystem(t *testing.T) {
+	raw := []any{
+		map[string]any{"role": "developer", "content": "必须先走工具调用"},
+		map[string]any{"role": "user", "content": "你好"},
+	}
+	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
+	if len(normalized) != 2 {
+		t.Fatalf("expected 2 normalized messages, got %d", len(normalized))
+	}
+	if normalized[0]["role"] != "system" {
+		t.Fatalf("expected developer role converted to system, got %#v", normalized[0]["role"])
+	}
+}
+
+func TestNormalizeOpenAIMessagesForPrompt_AssistantArrayContentFallbackWhenTextEmpty(t *testing.T) {
+	raw := []any{
+		map[string]any{
+			"role": "assistant",
+			"content": []any{
+				map[string]any{"type": "text", "text": "", "content": "工具说明文本"},
+			},
+		},
+	}
+
+	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
+	if len(normalized) != 1 {
+		t.Fatalf("expected one normalized message, got %d", len(normalized))
+	}
+	content, _ := normalized[0]["content"].(string)
+	if content != "工具说明文本" {
+		t.Fatalf("expected content fallback text preserved, got %q", content)
+	}
+}

internal/adapter/openai/responses_input_items.go

@@ -29,7 +29,7 @@ func normalizeResponsesInputItemWithState(m map[string]any, callNameByID map[str
 			return nil
 		}
 		return map[string]any{
-			"role":    role,
+			"role":    normalizeOpenAIRoleForPrompt(role),
 			"content": content,
 		}
 	}
@@ -51,7 +51,7 @@ func normalizeResponsesInputItemWithState(m map[string]any, callNameByID map[str
 			role = "user"
 		}
 		return map[string]any{
-			"role":    role,
+			"role":    normalizeOpenAIRoleForPrompt(role),
 			"content": content,
 		}
 	case "function_call_output", "tool_result":

internal/adapter/openai/responses_stream_runtime_toolcalls.go

@@ -94,6 +94,16 @@ func (s *responsesStreamRuntime) closeMessageItem() {
 	outputIndex := s.ensureMessageOutputIndex()
 	text := s.visibleText.String()
 	if s.messagePartAdded {
+		s.sendEvent(
+			"response.output_text.done",
+			openaifmt.BuildResponsesTextDonePayload(
+				s.responseID,
+				itemID,
+				outputIndex,
+				0,
+				text,
+			),
+		)
 		s.sendEvent(
 			"response.content_part.done",
 			openaifmt.BuildResponsesContentPartDonePayload(

internal/adapter/openai/responses_stream_test.go

@@ -226,6 +226,40 @@ func TestHandleResponsesStreamMultiToolCallKeepsNameAndCallIDAligned(t *testing.
 	}
 }
 
+func TestHandleResponsesStreamEmitsOutputTextDoneBeforeContentPartDone(t *testing.T) {
+	h := &Handler{}
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	rec := httptest.NewRecorder()
+
+	sseLine := func(v string) string {
+		b, _ := json.Marshal(map[string]any{
+			"p": "response/content",
+			"v": v,
+		})
+		return "data: " + string(b) + "\n"
+	}
+
+	streamBody := sseLine("hello") + "data: [DONE]\n"
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(streamBody)),
+	}
+
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, nil, util.DefaultToolChoicePolicy(), "")
+	body := rec.Body.String()
+	if !strings.Contains(body, "event: response.output_text.done") {
+		t.Fatalf("expected response.output_text.done payload, body=%s", body)
+	}
+	textDoneIdx := strings.Index(body, "event: response.output_text.done")
+	partDoneIdx := strings.Index(body, "event: response.content_part.done")
+	if textDoneIdx < 0 || partDoneIdx < 0 {
+		t.Fatalf("expected output_text.done + content_part.done, body=%s", body)
+	}
+	if textDoneIdx > partDoneIdx {
+		t.Fatalf("expected output_text.done before content_part.done, body=%s", body)
+	}
+}
+
 func TestHandleResponsesStreamOutputTextDeltaCarriesItemIndexes(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -288,12 +322,8 @@ func TestHandleResponsesStreamThinkingAndMixedToolExampleRemainMessageOnly(t *te
 	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
 
 	addedPayloads := extractAllSSEEventPayloads(rec.Body.String(), "response.output_item.added")
-	if len(addedPayloads) != 1 {
-		t.Fatalf("expected only one message output_item.added event, got %d body=%s", len(addedPayloads), rec.Body.String())
-	}
-	item, _ := addedPayloads[0]["item"].(map[string]any)
-	if asString(item["type"]) != "message" {
-		t.Fatalf("expected only message output item in strict mode, got %#v", item)
+	if len(addedPayloads) < 1 {
+		t.Fatalf("expected at least one output_item.added event, got %d body=%s", len(addedPayloads), rec.Body.String())
 	}
 
 	completedPayload, ok := extractSSEEventPayload(rec.Body.String(), "response.completed")
@@ -302,15 +332,22 @@ func TestHandleResponsesStreamThinkingAndMixedToolExampleRemainMessageOnly(t *te
 	}
 	responseObj, _ := completedPayload["response"].(map[string]any)
 	output, _ := responseObj["output"].([]any)
+	hasMessage := false
 	for _, item := range output {
 		m, _ := item.(map[string]any)
 		if m == nil {
 			continue
 		}
+		if asString(m["type"]) == "message" {
+			hasMessage = true
+		}
 		if asString(m["type"]) == "function_call" {
 			t.Fatalf("did not expect function_call output for mixed prose tool example, output=%#v", output)
 		}
 	}
+	if !hasMessage {
+		t.Fatalf("expected message output for mixed prose tool example, output=%#v", output)
+	}
 }
 
 func TestHandleResponsesStreamToolChoiceNoneRejectsFunctionCall(t *testing.T) {

internal/adapter/openai/tool_sieve_core.go

@@ -15,19 +15,9 @@ func processToolSieveChunk(state *toolStreamSieveState, chunk string, toolNames
 	}
 	events := make([]toolStreamEvent, 0, 2)
 	if len(state.pendingToolCalls) > 0 {
-		pending := state.pending.String()
-		if strings.TrimSpace(pending) != "" {
-			content := state.pendingToolRaw + pending
-			state.pending.Reset()
-			state.pendingToolRaw = ""
-			state.pendingToolCalls = nil
-			state.noteText(content)
-			events = append(events, toolStreamEvent{Content: content})
-		} else {
-			// Wait for either more non-whitespace content (demote to plain text)
-			// or stream flush (promote to executable tool calls).
-			return events
-		}
+		events = append(events, toolStreamEvent{ToolCalls: state.pendingToolCalls})
+		state.pendingToolRaw = ""
+		state.pendingToolCalls = nil
 	}
 
 	for {
@@ -45,7 +35,14 @@ func processToolSieveChunk(state *toolStreamSieveState, chunk string, toolNames
 			state.capturing = false
 			state.resetIncrementalToolState()
 			if len(calls) > 0 {
-				state.pendingToolRaw = captured
+				if prefix != "" {
+					state.noteText(prefix)
+					events = append(events, toolStreamEvent{Content: prefix})
+				}
+				if suffix != "" {
+					state.pending.WriteString(suffix)
+				}
+				_ = captured
 				state.pendingToolCalls = calls
 				continue
 			}
@@ -211,11 +208,6 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	if insideCodeFence(state.recentTextTail + prefixPart) {
 		return captured, nil, "", true
 	}
-	// Strict mode: only standalone tool payloads are executable. If the
-	// payload is wrapped by non-whitespace prose, keep it as plain text.
-	if strings.TrimSpace(state.recentTextTail) != "" || strings.TrimSpace(prefixPart) != "" || strings.TrimSpace(suffixPart) != "" {
-		return captured, nil, "", true
-	}
 	parsed := util.ParseStandaloneToolCallsDetailed(obj, toolNames)
 	if len(parsed.Calls) == 0 {
 		if parsed.SawToolCallSyntax && parsed.RejectedByPolicy {

internal/compat/go_compat_test.go

@@ -73,22 +73,31 @@ func TestGoCompatToolcallFixtures(t *testing.T) {
 		mustLoadJSON(t, fixturePath, &fixture)
 
 		var expected struct {
-			Calls []util.ParsedToolCall `json:"calls"`
+			Calls             []util.ParsedToolCall `json:"calls"`
+			SawToolCallSyntax bool                  `json:"sawToolCallSyntax"`
+			RejectedByPolicy  bool                  `json:"rejectedByPolicy"`
+			RejectedToolNames []string              `json:"rejectedToolNames"`
 		}
 		mustLoadJSON(t, expectedPath, &expected)
 
-		var got []util.ParsedToolCall
+		var got util.ToolCallParseResult
 		switch strings.ToLower(strings.TrimSpace(fixture.Mode)) {
 		case "standalone":
-			got = util.ParseStandaloneToolCalls(fixture.Text, fixture.ToolNames)
+			got = util.ParseStandaloneToolCallsDetailed(fixture.Text, fixture.ToolNames)
 		default:
-			got = util.ParseToolCalls(fixture.Text, fixture.ToolNames)
+			got = util.ParseToolCallsDetailed(fixture.Text, fixture.ToolNames)
 		}
-		if len(got) == 0 && len(expected.Calls) == 0 {
-			continue
+		if got.Calls == nil {
+			got.Calls = []util.ParsedToolCall{}
 		}
-		if !reflect.DeepEqual(got, expected.Calls) {
-			t.Fatalf("toolcall fixture %s mismatch:\n got=%#v\nwant=%#v", name, got, expected.Calls)
+		if got.RejectedToolNames == nil {
+			got.RejectedToolNames = []string{}
+		}
+		if !reflect.DeepEqual(got.Calls, expected.Calls) ||
+			got.SawToolCallSyntax != expected.SawToolCallSyntax ||
+			got.RejectedByPolicy != expected.RejectedByPolicy ||
+			!reflect.DeepEqual(got.RejectedToolNames, expected.RejectedToolNames) {
+			t.Fatalf("toolcall fixture %s mismatch:\n got=%#v\nwant=%#v", name, got, expected)
 		}
 	}
 }

internal/format/openai/render_stream_events.go

@@ -71,6 +71,19 @@ func BuildResponsesTextDeltaPayload(responseID, itemID string, outputIndex, cont
 	}
 }
 
+
+func BuildResponsesTextDonePayload(responseID, itemID string, outputIndex, contentIndex int, text string) map[string]any {
+	return map[string]any{
+		"type":          "response.output_text.done",
+		"id":            responseID,
+		"response_id":   responseID,
+		"item_id":       itemID,
+		"output_index":  outputIndex,
+		"content_index": contentIndex,
+		"text":          text,
+	}
+}
+
 func BuildResponsesReasoningDeltaPayload(responseID, delta string) map[string]any {
 	return map[string]any{
 		"type":        "response.reasoning.delta",

internal/js/helpers/stream-tool-sieve/parse.js

@@ -8,8 +8,12 @@ const {
   stripFencedCodeBlocks,
   buildToolCallCandidates,
   parseToolCallsPayload,
+  parseMarkupToolCalls,
+  parseTextKVToolCalls,
 } = require('./parse_payload');
 
+const TOOL_NAME_LOOSE_PATTERN = /[^a-z0-9]+/g;
+
 function extractToolNames(tools) {
   if (!Array.isArray(tools) || tools.length === 0) {
     return [];
@@ -41,19 +45,32 @@ function parseToolCallsDetailed(text, toolNames) {
   if (!toStringSafe(sanitized)) {
     return result;
   }
-  result.sawToolCallSyntax = sanitized.toLowerCase().includes('tool_calls');
+  result.sawToolCallSyntax = looksLikeToolCallSyntax(sanitized);
 
   const candidates = buildToolCallCandidates(sanitized);
   let parsed = [];
   for (const c of candidates) {
     parsed = parseToolCallsPayload(c);
+    if (parsed.length === 0) {
+      parsed = parseMarkupToolCalls(c);
+    }
+    if (parsed.length === 0) {
+      parsed = parseTextKVToolCalls(c);
+    }
     if (parsed.length > 0) {
       result.sawToolCallSyntax = true;
       break;
     }
   }
   if (parsed.length === 0) {
-    return result;
+    parsed = parseMarkupToolCalls(sanitized);
+    if (parsed.length === 0) {
+      parsed = parseTextKVToolCalls(sanitized);
+      if (parsed.length === 0) {
+        return result;
+      }
+    }
+    result.sawToolCallSyntax = true;
   }
 
   const filtered = filterToolCallsDetailed(parsed, toolNames);
@@ -73,15 +90,20 @@ function parseStandaloneToolCallsDetailed(text, toolNames) {
   if (!trimmed) {
     return result;
   }
+  if (trimmed.includes('```')) {
+    return result;
+  }
   if (looksLikeToolExampleContext(trimmed)) {
     return result;
   }
-  result.sawToolCallSyntax = trimmed.toLowerCase().includes('tool_calls');
-  if (!trimmed.startsWith('{') && !trimmed.startsWith('[')) {
-    return result;
+  result.sawToolCallSyntax = looksLikeToolCallSyntax(trimmed);
+  let parsed = parseToolCallsPayload(trimmed);
+  if (parsed.length === 0) {
+    parsed = parseMarkupToolCalls(trimmed);
+  }
+  if (parsed.length === 0) {
+    parsed = parseTextKVToolCalls(trimmed);
   }
-
-  const parsed = parseToolCallsPayload(trimmed);
   if (parsed.length === 0) {
     return result;
   }
@@ -146,7 +168,7 @@ function filterToolCallsDetailed(parsed, toolNames) {
     if (allowed.has(tc.name)) {
       matchedName = tc.name;
     } else {
-      matchedName = allowedCanonical.get(tc.name.toLowerCase()) || '';
+      matchedName = resolveAllowedToolName(tc.name, allowed, allowedCanonical);
     }
     if (!matchedName) {
       if (!seenRejected.has(tc.name)) {
@@ -163,6 +185,46 @@ function filterToolCallsDetailed(parsed, toolNames) {
   return { calls, rejectedToolNames: rejected };
 }
 
+function resolveAllowedToolName(name, allowed, allowedCanonical) {
+  const normalizedName = toStringSafe(name).trim();
+  if (!normalizedName) {
+    return '';
+  }
+  if (allowed.has(normalizedName)) {
+    return normalizedName;
+  }
+  const lower = normalizedName.toLowerCase();
+  if (allowedCanonical.has(lower)) {
+    return allowedCanonical.get(lower);
+  }
+  const idx = lower.lastIndexOf('.');
+  if (idx >= 0 && idx < lower.length - 1) {
+    const tail = lower.slice(idx + 1);
+    if (allowedCanonical.has(tail)) {
+      return allowedCanonical.get(tail);
+    }
+  }
+  const loose = lower.replace(TOOL_NAME_LOOSE_PATTERN, '');
+  if (!loose) {
+    return '';
+  }
+  for (const [candidateLower, canonical] of allowedCanonical.entries()) {
+    if (candidateLower.replace(TOOL_NAME_LOOSE_PATTERN, '') === loose) {
+      return canonical;
+    }
+  }
+  return '';
+}
+
+function looksLikeToolCallSyntax(text) {
+  const lower = toStringSafe(text).toLowerCase();
+  return lower.includes('tool_calls')
+    || lower.includes('<tool_call')
+    || lower.includes('<function_call')
+    || lower.includes('<invoke')
+    || lower.includes('function.name:');
+}
+
 module.exports = {
   extractToolNames,
   parseToolCalls,

internal/js/helpers/stream-tool-sieve/parse_payload.js

@@ -1,6 +1,24 @@
 'use strict';
 
 const TOOL_CALL_PATTERN = /\{\s*["']tool_calls["']\s*:\s*\[(.*?)\]\s*\}/s;
+const TOOL_CALL_MARKUP_BLOCK_PATTERN = /<(?:[a-z0-9_:-]+:)?(tool_call|function_call|invoke)\b([^>]*)>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?\1>/gi;
+const TOOL_CALL_MARKUP_SELFCLOSE_PATTERN = /<(?:[a-z0-9_:-]+:)?invoke\b([^>]*)\/>/gi;
+const TOOL_CALL_MARKUP_KV_PATTERN = /<(?:[a-z0-9_:-]+:)?([a-z0-9_.-]+)\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?\1>/gi;
+const TOOL_CALL_MARKUP_ATTR_PATTERN = /(name|function|tool)\s*=\s*"([^"]+)"/i;
+const TOOL_CALL_MARKUP_NAME_PATTERNS = [
+  /<(?:[a-z0-9_:-]+:)?name\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?name>/i,
+  /<(?:[a-z0-9_:-]+:)?function\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?function>/i,
+];
+const TOOL_CALL_MARKUP_ARGS_PATTERNS = [
+  /<(?:[a-z0-9_:-]+:)?input\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?input>/i,
+  /<(?:[a-z0-9_:-]+:)?arguments\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?arguments>/i,
+  /<(?:[a-z0-9_:-]+:)?argument\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?argument>/i,
+  /<(?:[a-z0-9_:-]+:)?parameters\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?parameters>/i,
+  /<(?:[a-z0-9_:-]+:)?parameter\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?parameter>/i,
+  /<(?:[a-z0-9_:-]+:)?args\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?args>/i,
+  /<(?:[a-z0-9_:-]+:)?params\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?params>/i,
+];
+const TEXT_KV_NAME_PATTERN = /function\.name:\s*([a-zA-Z0-9_.-]+)/gi;
 
 const {
   toStringSafe,
@@ -103,6 +121,153 @@ function parseToolCallsPayload(payload) {
   return one ? [one] : [];
 }
 
+function parseMarkupToolCalls(text) {
+  const raw = toStringSafe(text).trim();
+  if (!raw) {
+    return [];
+  }
+  const out = [];
+  for (const m of raw.matchAll(TOOL_CALL_MARKUP_BLOCK_PATTERN)) {
+    const parsed = parseMarkupSingleToolCall(toStringSafe(m[2]).trim(), toStringSafe(m[3]).trim());
+    if (parsed) {
+      out.push(parsed);
+    }
+  }
+  for (const m of raw.matchAll(TOOL_CALL_MARKUP_SELFCLOSE_PATTERN)) {
+    const parsed = parseMarkupSingleToolCall(toStringSafe(m[1]).trim(), '');
+    if (parsed) {
+      out.push(parsed);
+    }
+  }
+  return out;
+}
+
+function parseTextKVToolCalls(text) {
+  const raw = toStringSafe(text);
+  if (!raw) {
+    return [];
+  }
+  const out = [];
+  const matches = [...raw.matchAll(TEXT_KV_NAME_PATTERN)];
+  if (matches.length === 0) {
+    return out;
+  }
+  for (let i = 0; i < matches.length; i += 1) {
+    const match = matches[i];
+    const name = toStringSafe(match[1]).trim();
+    if (!name) {
+      continue;
+    }
+    const nameEnd = match.index + toStringSafe(match[0]).length;
+    const searchEnd = i + 1 < matches.length ? matches[i + 1].index : raw.length;
+    const searchArea = raw.slice(nameEnd, searchEnd);
+    const argIdx = searchArea.indexOf('function.arguments:');
+    if (argIdx < 0) {
+      continue;
+    }
+    const argStart = nameEnd + argIdx + 'function.arguments:'.length;
+    const bracePos = raw.slice(argStart, searchEnd).indexOf('{');
+    if (bracePos < 0) {
+      continue;
+    }
+    const objStart = argStart + bracePos;
+    const obj = extractJSONObjectFrom(raw, objStart);
+    if (!obj.ok) {
+      continue;
+    }
+    out.push({
+      name,
+      input: parseToolCallInput(raw.slice(objStart, obj.end)),
+    });
+  }
+  return out;
+}
+
+function parseMarkupSingleToolCall(attrs, inner) {
+  const embedded = parseToolCallsPayload(inner);
+  if (embedded.length > 0) {
+    return embedded[0];
+  }
+  let name = '';
+  const attrMatch = attrs.match(TOOL_CALL_MARKUP_ATTR_PATTERN);
+  if (attrMatch && attrMatch[2]) {
+    name = toStringSafe(attrMatch[2]).trim();
+  }
+  if (!name) {
+    name = stripTagText(findMarkupTagValue(inner, TOOL_CALL_MARKUP_NAME_PATTERNS));
+  }
+  if (!name) {
+    return null;
+  }
+
+  let input = {};
+  const argsRaw = findMarkupTagValue(inner, TOOL_CALL_MARKUP_ARGS_PATTERNS);
+  if (argsRaw) {
+    input = parseMarkupInput(argsRaw);
+  } else {
+    const kv = parseMarkupKVObject(inner);
+    if (Object.keys(kv).length > 0) {
+      input = kv;
+    }
+  }
+  return { name, input };
+}
+
+function parseMarkupInput(raw) {
+  const s = toStringSafe(raw).trim();
+  if (!s) {
+    return {};
+  }
+  const parsed = parseToolCallInput(s);
+  if (parsed && typeof parsed === 'object' && !Array.isArray(parsed) && Object.keys(parsed).length > 0) {
+    return parsed;
+  }
+  const kv = parseMarkupKVObject(s);
+  if (Object.keys(kv).length > 0) {
+    return kv;
+  }
+  return { _raw: stripTagText(s) };
+}
+
+function parseMarkupKVObject(text) {
+  const raw = toStringSafe(text).trim();
+  if (!raw) {
+    return {};
+  }
+  const out = {};
+  for (const m of raw.matchAll(TOOL_CALL_MARKUP_KV_PATTERN)) {
+    const key = toStringSafe(m[1]).trim();
+    if (!key) {
+      continue;
+    }
+    const valueRaw = stripTagText(m[2]);
+    if (!valueRaw) {
+      continue;
+    }
+    try {
+      out[key] = JSON.parse(valueRaw);
+    } catch (_err) {
+      out[key] = valueRaw;
+    }
+  }
+  return out;
+}
+
+function stripTagText(text) {
+  return toStringSafe(text).replace(/<[^>]+>/g, ' ').trim();
+}
+
+function findMarkupTagValue(text, patterns) {
+  const source = toStringSafe(text);
+  for (const p of patterns) {
+    const m = source.match(p);
+    if (m && m[1]) {
+      return toStringSafe(m[1]);
+    }
+  }
+  return '';
+}
+
 function parseToolCallList(v) {
   if (!Array.isArray(v)) {
     return [];
@@ -193,4 +358,6 @@ module.exports = {
   stripFencedCodeBlocks,
   buildToolCallCandidates,
   parseToolCallsPayload,
+  parseMarkupToolCalls,
+  parseTextKVToolCalls,
 };

internal/js/helpers/stream-tool-sieve/sieve.js

@@ -21,22 +21,14 @@ function processToolSieveChunk(state, chunk, toolNames) {
   }
   const events = [];
 
-  if (Array.isArray(state.pendingToolCalls) && state.pendingToolCalls.length > 0) {
-    const pending = state.pending || '';
-    if (pending.trim() !== '') {
-      const content = (state.pendingToolRaw || '') + pending;
-      state.pending = '';
-      state.pendingToolRaw = '';
-      state.pendingToolCalls = [];
-      noteText(state, content);
-      events.push({ type: 'text', text: content });
-    } else {
-      return events;
-    }
-  }
-
   // eslint-disable-next-line no-constant-condition
   while (true) {
+    if (Array.isArray(state.pendingToolCalls) && state.pendingToolCalls.length > 0) {
+      events.push({ type: 'tool_calls', calls: state.pendingToolCalls });
+      state.pendingToolRaw = '';
+      state.pendingToolCalls = [];
+      continue;
+    }
     if (state.capturing) {
       if (state.pending) {
         state.capture += state.pending;

internal/prompt/messages.go

@@ -51,6 +51,9 @@ func MessagesPrepare(messages []map[string]any) string {
 }
 
 func NormalizeContent(v any) string {
+	if v == nil {
+		return ""
+	}
 	switch x := v.(type) {
 	case string:
 		return x
@@ -64,11 +67,11 @@ func NormalizeContent(v any) string {
 			typeStr, _ := m["type"].(string)
 			typeStr = strings.ToLower(strings.TrimSpace(typeStr))
 			if typeStr == "text" || typeStr == "output_text" || typeStr == "input_text" {
-				if txt, ok := m["text"].(string); ok {
+				if txt, ok := m["text"].(string); ok && txt != "" {
 					parts = append(parts, txt)
 					continue
 				}
-				if txt, ok := m["content"].(string); ok {
+				if txt, ok := m["content"].(string); ok && txt != "" {
 					parts = append(parts, txt)
 				}
 			}

internal/prompt/messages_test.go

@@ -0,0 +1,32 @@
+package prompt
+
+import "testing"
+
+func TestNormalizeContentNilReturnsEmpty(t *testing.T) {
+	if got := NormalizeContent(nil); got != "" {
+		t.Fatalf("expected empty string for nil content, got %q", got)
+	}
+}
+
+func TestMessagesPrepareNilContentNoNullLiteral(t *testing.T) {
+	messages := []map[string]any{
+		{"role": "assistant", "content": nil},
+		{"role": "user", "content": "ok"},
+	}
+	got := MessagesPrepare(messages)
+	if got == "" {
+		t.Fatalf("expected non-empty output")
+	}
+	if got == "null" {
+		t.Fatalf("expected no null literal output, got %q", got)
+	}
+}
+
+func TestNormalizeContentArrayFallsBackToContentWhenTextEmpty(t *testing.T) {
+	got := NormalizeContent([]any{
+		map[string]any{"type": "text", "text": "", "content": "from-content"},
+	})
+	if got != "from-content" {
+		t.Fatalf("expected fallback to content when text is empty, got %q", got)
+	}
+}

internal/util/toolcalls_markup.go

@@ -0,0 +1,161 @@
+package util
+
+import (
+	"encoding/json"
+	"regexp"
+	"strings"
+)
+
+var toolCallMarkupTagNames = []string{"tool_call", "function_call", "invoke"}
+var toolCallMarkupTagPatternByName = map[string]*regexp.Regexp{
+	"tool_call":     regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?tool_call\b([^>]*)>(.*?)</(?:[a-z0-9_:-]+:)?tool_call>`),
+	"function_call": regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?function_call\b([^>]*)>(.*?)</(?:[a-z0-9_:-]+:)?function_call>`),
+	"invoke":        regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?invoke\b([^>]*)>(.*?)</(?:[a-z0-9_:-]+:)?invoke>`),
+}
+var toolCallMarkupSelfClosingPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?invoke\b([^>]*)/>`)
+var toolCallMarkupKVPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?([a-z0-9_\-.]+)\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?([a-z0-9_\-.]+)>`)
+var toolCallMarkupAttrPattern = regexp.MustCompile(`(?is)(name|function|tool)\s*=\s*"([^"]+)"`)
+var anyTagPattern = regexp.MustCompile(`(?is)<[^>]+>`)
+var toolCallMarkupNameTagNames = []string{"name", "function"}
+var toolCallMarkupNamePatternByTag = map[string]*regexp.Regexp{
+	"name":     regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?name\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?name>`),
+	"function": regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?function\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?function>`),
+}
+var toolCallMarkupArgsTagNames = []string{"input", "arguments", "argument", "parameters", "parameter", "args", "params"}
+var toolCallMarkupArgsPatternByTag = map[string]*regexp.Regexp{
+	"input":      regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?input\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?input>`),
+	"arguments":  regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?arguments\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?arguments>`),
+	"argument":   regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?argument\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?argument>`),
+	"parameters": regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?parameters\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?parameters>`),
+	"parameter":  regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?parameter\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?parameter>`),
+	"args":       regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?args\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?args>`),
+	"params":     regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?params\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?params>`),
+}
+
+func parseMarkupToolCalls(text string) []ParsedToolCall {
+	trimmed := strings.TrimSpace(text)
+	if trimmed == "" {
+		return nil
+	}
+
+	out := make([]ParsedToolCall, 0)
+	for _, tagName := range toolCallMarkupTagNames {
+		pattern := toolCallMarkupTagPatternByName[tagName]
+		for _, m := range pattern.FindAllStringSubmatch(trimmed, -1) {
+			if len(m) < 3 {
+				continue
+			}
+			attrs := strings.TrimSpace(m[1])
+			inner := strings.TrimSpace(m[2])
+			if parsed := parseMarkupSingleToolCall(attrs, inner); parsed.Name != "" {
+				out = append(out, parsed)
+			}
+		}
+	}
+	for _, m := range toolCallMarkupSelfClosingPattern.FindAllStringSubmatch(trimmed, -1) {
+		if len(m) < 2 {
+			continue
+		}
+		if parsed := parseMarkupSingleToolCall(strings.TrimSpace(m[1]), ""); parsed.Name != "" {
+			out = append(out, parsed)
+		}
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func parseMarkupSingleToolCall(attrs string, inner string) ParsedToolCall {
+	if parsed := parseToolCallsPayload(inner); len(parsed) > 0 {
+		return parsed[0]
+	}
+
+	name := ""
+	if m := toolCallMarkupAttrPattern.FindStringSubmatch(attrs); len(m) >= 3 {
+		name = strings.TrimSpace(m[2])
+	}
+	if name == "" {
+		name = findMarkupTagValue(inner, toolCallMarkupNameTagNames, toolCallMarkupNamePatternByTag)
+	}
+	if name == "" {
+		return ParsedToolCall{}
+	}
+
+	input := map[string]any{}
+	if argsRaw := findMarkupTagValue(inner, toolCallMarkupArgsTagNames, toolCallMarkupArgsPatternByTag); argsRaw != "" {
+		input = parseMarkupInput(argsRaw)
+	} else if kv := parseMarkupKVObject(inner); len(kv) > 0 {
+		input = kv
+	}
+	return ParsedToolCall{Name: name, Input: input}
+}
+
+func parseMarkupInput(raw string) map[string]any {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return map[string]any{}
+	}
+	if parsed := parseToolCallInput(raw); len(parsed) > 0 {
+		return parsed
+	}
+	if kv := parseMarkupKVObject(raw); len(kv) > 0 {
+		return kv
+	}
+	return map[string]any{"_raw": stripTagText(raw)}
+}
+
+func parseMarkupKVObject(text string) map[string]any {
+	matches := toolCallMarkupKVPattern.FindAllStringSubmatch(strings.TrimSpace(text), -1)
+	if len(matches) == 0 {
+		return nil
+	}
+	out := map[string]any{}
+	for _, m := range matches {
+		if len(m) < 4 {
+			continue
+		}
+		key := strings.TrimSpace(m[1])
+		endKey := strings.TrimSpace(m[3])
+		if key == "" {
+			continue
+		}
+		if !strings.EqualFold(key, endKey) {
+			continue
+		}
+		value := strings.TrimSpace(stripTagText(m[2]))
+		if value == "" {
+			continue
+		}
+		var jsonValue any
+		if json.Unmarshal([]byte(value), &jsonValue) == nil {
+			out[key] = jsonValue
+			continue
+		}
+		out[key] = value
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func stripTagText(text string) string {
+	return strings.TrimSpace(anyTagPattern.ReplaceAllString(text, ""))
+}
+
+func findMarkupTagValue(text string, tagNames []string, patternByTag map[string]*regexp.Regexp) string {
+	for _, tag := range tagNames {
+		pattern := patternByTag[tag]
+		if pattern == nil {
+			continue
+		}
+		if m := pattern.FindStringSubmatch(text); len(m) >= 2 {
+			value := strings.TrimSpace(m[1])
+			if value != "" {
+				return value
+			}
+		}
+	}
+	return ""
+}

internal/util/toolcalls_name_match.go

@@ -0,0 +1,33 @@
+package util
+
+import (
+	"regexp"
+	"strings"
+)
+
+var toolNameLoosePattern = regexp.MustCompile(`[^a-z0-9]+`)
+
+func resolveAllowedToolNameWithLooseMatch(name string, allowed map[string]struct{}, allowedCanonical map[string]string) string {
+	if _, ok := allowed[name]; ok {
+		return name
+	}
+	lower := strings.ToLower(strings.TrimSpace(name))
+	if canonical, ok := allowedCanonical[lower]; ok {
+		return canonical
+	}
+	if idx := strings.LastIndex(lower, "."); idx >= 0 && idx < len(lower)-1 {
+		if canonical, ok := allowedCanonical[lower[idx+1:]]; ok {
+			return canonical
+		}
+	}
+	loose := toolNameLoosePattern.ReplaceAllString(lower, "")
+	if loose == "" {
+		return ""
+	}
+	for candidateLower, canonical := range allowedCanonical {
+		if toolNameLoosePattern.ReplaceAllString(candidateLower, "") == loose {
+			return canonical
+		}
+	}
+	return ""
+}

internal/util/toolcalls_parse.go

@@ -30,19 +30,36 @@ func ParseToolCallsDetailed(text string, availableToolNames []string) ToolCallPa
 	if strings.TrimSpace(text) == "" {
 		return result
 	}
-	result.SawToolCallSyntax = strings.Contains(strings.ToLower(text), "tool_calls")
+	result.SawToolCallSyntax = looksLikeToolCallSyntax(text)
 
 	candidates := buildToolCallCandidates(text)
 	var parsed []ParsedToolCall
 	for _, candidate := range candidates {
-		if tc := parseToolCallsPayload(candidate); len(tc) > 0 {
+		tc := parseToolCallsPayload(candidate)
+		if len(tc) == 0 {
+			tc = parseXMLToolCalls(candidate)
+		}
+		if len(tc) == 0 {
+			tc = parseMarkupToolCalls(candidate)
+		}
+		if len(tc) == 0 {
+			tc = parseTextKVToolCalls(candidate)
+		}
+		if len(tc) > 0 {
 			parsed = tc
 			result.SawToolCallSyntax = true
 			break
 		}
 	}
 	if len(parsed) == 0 {
-		return result
+		parsed = parseXMLToolCalls(text)
+		if len(parsed) == 0 {
+			parsed = parseTextKVToolCalls(text)
+			if len(parsed) == 0 {
+				return result
+			}
+		}
+		result.SawToolCallSyntax = true
 	}
 
 	calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
@@ -65,17 +82,24 @@ func ParseStandaloneToolCallsDetailed(text string, availableToolNames []string)
 	if looksLikeToolExampleContext(trimmed) {
 		return result
 	}
-	result.SawToolCallSyntax = strings.Contains(strings.ToLower(trimmed), "tool_calls")
+	result.SawToolCallSyntax = looksLikeToolCallSyntax(trimmed)
 	candidates := []string{trimmed}
 	for _, candidate := range candidates {
 		candidate = strings.TrimSpace(candidate)
 		if candidate == "" {
 			continue
 		}
-		if !strings.HasPrefix(candidate, "{") && !strings.HasPrefix(candidate, "[") {
-			continue
+		parsed := parseToolCallsPayload(candidate)
+		if len(parsed) == 0 {
+			parsed = parseXMLToolCalls(candidate)
 		}
-		if parsed := parseToolCallsPayload(candidate); len(parsed) > 0 {
+		if len(parsed) == 0 {
+			parsed = parseMarkupToolCalls(candidate)
+		}
+		if len(parsed) == 0 {
+			parsed = parseTextKVToolCalls(candidate)
+		}
+		if len(parsed) > 0 {
 			result.SawToolCallSyntax = true
 			calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
 			result.Calls = calls
@@ -103,32 +127,32 @@ func filterToolCallsDetailed(parsed []ParsedToolCall, availableToolNames []strin
 	}
 	if len(allowed) == 0 {
 		rejectedSet := map[string]struct{}{}
+		rejected := make([]string, 0, len(parsed))
 		for _, tc := range parsed {
 			if tc.Name == "" {
 				continue
 			}
+			if _, ok := rejectedSet[tc.Name]; ok {
+				continue
+			}
 			rejectedSet[tc.Name] = struct{}{}
-		}
-		rejected := make([]string, 0, len(rejectedSet))
-		for name := range rejectedSet {
-			rejected = append(rejected, name)
+			rejected = append(rejected, tc.Name)
 		}
 		return nil, rejected
 	}
 	out := make([]ParsedToolCall, 0, len(parsed))
 	rejectedSet := map[string]struct{}{}
+	rejected := make([]string, 0)
 	for _, tc := range parsed {
 		if tc.Name == "" {
 			continue
 		}
-		matchedName := ""
-		if _, ok := allowed[tc.Name]; ok {
-			matchedName = tc.Name
-		} else if canonical, ok := allowedCanonical[strings.ToLower(tc.Name)]; ok {
-			matchedName = canonical
-		}
+		matchedName := resolveAllowedToolName(tc.Name, allowed, allowedCanonical)
 		if matchedName == "" {
-			rejectedSet[tc.Name] = struct{}{}
+			if _, ok := rejectedSet[tc.Name]; !ok {
+				rejectedSet[tc.Name] = struct{}{}
+				rejected = append(rejected, tc.Name)
+			}
 			continue
 		}
 		tc.Name = matchedName
@@ -137,13 +161,13 @@ func filterToolCallsDetailed(parsed []ParsedToolCall, availableToolNames []strin
 		}
 		out = append(out, tc)
 	}
-	rejected := make([]string, 0, len(rejectedSet))
-	for name := range rejectedSet {
-		rejected = append(rejected, name)
-	}
 	return out, rejected
 }
 
+func resolveAllowedToolName(name string, allowed map[string]struct{}, allowedCanonical map[string]string) string {
+	return resolveAllowedToolNameWithLooseMatch(name, allowed, allowedCanonical)
+}
+
 func parseToolCallsPayload(payload string) []ParsedToolCall {
 	var decoded any
 	if err := json.Unmarshal([]byte(payload), &decoded); err != nil {
@@ -163,6 +187,15 @@ func parseToolCallsPayload(payload string) []ParsedToolCall {
 	return nil
 }
 
+func looksLikeToolCallSyntax(text string) bool {
+	lower := strings.ToLower(text)
+	return strings.Contains(lower, "tool_calls") ||
+		strings.Contains(lower, "<tool_call") ||
+		strings.Contains(lower, "<function_call") ||
+		strings.Contains(lower, "<invoke") ||
+		strings.Contains(lower, "function.name:")
+}
+
 func parseToolCallList(v any) []ParsedToolCall {
 	items, ok := v.([]any)
 	if !ok {

internal/util/toolcalls_parse_markup.go

@@ -0,0 +1,235 @@
+package util
+
+import (
+	"encoding/json"
+	"encoding/xml"
+	"regexp"
+	"strings"
+)
+
+var xmlToolCallPattern = regexp.MustCompile(`(?is)<tool_call>\s*(.*?)\s*</tool_call>`)
+var functionCallPattern = regexp.MustCompile(`(?is)<function_call>\s*([^<]+?)\s*</function_call>`)
+var functionParamPattern = regexp.MustCompile(`(?is)<function\s+parameter\s+name="([^"]+)"\s*>\s*(.*?)\s*</function\s+parameter>`)
+var antmlFunctionCallPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_]+:)?function_call[^>]*(?:name|function)="([^"]+)"[^>]*>\s*(.*?)\s*</(?:[a-z0-9_]+:)?function_call>`)
+var antmlArgumentPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_]+:)?argument\s+name="([^"]+)"\s*>\s*(.*?)\s*</(?:[a-z0-9_]+:)?argument>`)
+var antmlParametersPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_]+:)?parameters\s*>\s*(\{.*?\})\s*</(?:[a-z0-9_]+:)?parameters>`)
+var invokeCallPattern = regexp.MustCompile(`(?is)<invoke\s+name="([^"]+)"\s*>(.*?)</invoke>`)
+var invokeParamPattern = regexp.MustCompile(`(?is)<parameter\s+name="([^"]+)"\s*>\s*(.*?)\s*</parameter>`)
+
+func parseXMLToolCalls(text string) []ParsedToolCall {
+	matches := xmlToolCallPattern.FindAllString(text, -1)
+	out := make([]ParsedToolCall, 0, len(matches)+1)
+	for _, block := range matches {
+		call, ok := parseSingleXMLToolCall(block)
+		if !ok {
+			continue
+		}
+		out = append(out, call)
+	}
+	if len(out) > 0 {
+		return out
+	}
+	if call, ok := parseFunctionCallTagStyle(text); ok {
+		return []ParsedToolCall{call}
+	}
+	if calls := parseAntmlFunctionCallStyles(text); len(calls) > 0 {
+		return calls
+	}
+	if call, ok := parseInvokeFunctionCallStyle(text); ok {
+		return []ParsedToolCall{call}
+	}
+	return nil
+}
+
+func parseSingleXMLToolCall(block string) (ParsedToolCall, bool) {
+	inner := strings.TrimSpace(block)
+	inner = strings.TrimPrefix(inner, "<tool_call>")
+	inner = strings.TrimSuffix(inner, "</tool_call>")
+	inner = strings.TrimSpace(inner)
+	if strings.HasPrefix(inner, "{") {
+		var payload map[string]any
+		if err := json.Unmarshal([]byte(inner), &payload); err == nil {
+			name := strings.TrimSpace(asString(payload["tool"]))
+			if name == "" {
+				name = strings.TrimSpace(asString(payload["tool_name"]))
+			}
+			if name != "" {
+				input := map[string]any{}
+				if params, ok := payload["params"].(map[string]any); ok {
+					input = params
+				} else if params, ok := payload["parameters"].(map[string]any); ok {
+					input = params
+				}
+				return ParsedToolCall{Name: name, Input: input}, true
+			}
+		}
+	}
+
+	dec := xml.NewDecoder(strings.NewReader(block))
+	name := ""
+	params := map[string]any{}
+	inParams := false
+	inTool := false
+	for {
+		tok, err := dec.Token()
+		if err != nil {
+			break
+		}
+		switch t := tok.(type) {
+		case xml.StartElement:
+			tag := strings.ToLower(t.Name.Local)
+			switch tag {
+			case "tool":
+				inTool = true
+				for _, attr := range t.Attr {
+					if strings.EqualFold(strings.TrimSpace(attr.Name.Local), "name") && strings.TrimSpace(name) == "" {
+						name = strings.TrimSpace(attr.Value)
+					}
+				}
+			case "parameters":
+				inParams = true
+			case "tool_name", "name":
+				var v string
+				if err := dec.DecodeElement(&v, &t); err == nil && strings.TrimSpace(v) != "" {
+					name = strings.TrimSpace(v)
+				}
+			case "input", "arguments", "argument", "args", "params":
+				var v string
+				if err := dec.DecodeElement(&v, &t); err == nil && strings.TrimSpace(v) != "" {
+					if parsed := parseToolCallInput(strings.TrimSpace(v)); len(parsed) > 0 {
+						for k, vv := range parsed {
+							params[k] = vv
+						}
+					}
+				}
+			default:
+				if inParams || inTool {
+					var v string
+					if err := dec.DecodeElement(&v, &t); err == nil {
+						params[t.Name.Local] = strings.TrimSpace(v)
+					}
+				}
+			}
+		case xml.EndElement:
+			tag := strings.ToLower(t.Name.Local)
+			if tag == "parameters" {
+				inParams = false
+			}
+			if tag == "tool" {
+				inTool = false
+			}
+		}
+	}
+	if strings.TrimSpace(name) == "" {
+		return ParsedToolCall{}, false
+	}
+	return ParsedToolCall{Name: strings.TrimSpace(name), Input: params}, true
+}
+
+func parseFunctionCallTagStyle(text string) (ParsedToolCall, bool) {
+	m := functionCallPattern.FindStringSubmatch(text)
+	if len(m) < 2 {
+		return ParsedToolCall{}, false
+	}
+	name := strings.TrimSpace(m[1])
+	if name == "" {
+		return ParsedToolCall{}, false
+	}
+	input := map[string]any{}
+	for _, pm := range functionParamPattern.FindAllStringSubmatch(text, -1) {
+		if len(pm) < 3 {
+			continue
+		}
+		key := strings.TrimSpace(pm[1])
+		val := strings.TrimSpace(pm[2])
+		if key != "" {
+			input[key] = val
+		}
+	}
+	return ParsedToolCall{Name: name, Input: input}, true
+}
+
+func parseAntmlFunctionCallStyles(text string) []ParsedToolCall {
+	matches := antmlFunctionCallPattern.FindAllStringSubmatch(text, -1)
+	if len(matches) == 0 {
+		return nil
+	}
+	out := make([]ParsedToolCall, 0, len(matches))
+	for _, m := range matches {
+		if call, ok := parseSingleAntmlFunctionCallMatch(m); ok {
+			out = append(out, call)
+		}
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func parseSingleAntmlFunctionCallMatch(m []string) (ParsedToolCall, bool) {
+	if len(m) < 3 {
+		return ParsedToolCall{}, false
+	}
+	name := strings.TrimSpace(m[1])
+	if name == "" {
+		return ParsedToolCall{}, false
+	}
+	body := strings.TrimSpace(m[2])
+	input := map[string]any{}
+	if strings.HasPrefix(body, "{") {
+		if err := json.Unmarshal([]byte(body), &input); err == nil {
+			return ParsedToolCall{Name: name, Input: input}, true
+		}
+	}
+	if pm := antmlParametersPattern.FindStringSubmatch(body); len(pm) >= 2 {
+		if err := json.Unmarshal([]byte(strings.TrimSpace(pm[1])), &input); err == nil {
+			return ParsedToolCall{Name: name, Input: input}, true
+		}
+	}
+	for _, am := range antmlArgumentPattern.FindAllStringSubmatch(body, -1) {
+		if len(am) < 3 {
+			continue
+		}
+		k := strings.TrimSpace(am[1])
+		v := strings.TrimSpace(am[2])
+		if k != "" {
+			input[k] = v
+		}
+	}
+	return ParsedToolCall{Name: name, Input: input}, true
+}
+
+func parseInvokeFunctionCallStyle(text string) (ParsedToolCall, bool) {
+	m := invokeCallPattern.FindStringSubmatch(text)
+	if len(m) < 3 {
+		return ParsedToolCall{}, false
+	}
+	name := strings.TrimSpace(m[1])
+	if name == "" {
+		return ParsedToolCall{}, false
+	}
+	input := map[string]any{}
+	for _, pm := range invokeParamPattern.FindAllStringSubmatch(m[2], -1) {
+		if len(pm) < 3 {
+			continue
+		}
+		k := strings.TrimSpace(pm[1])
+		v := strings.TrimSpace(pm[2])
+		if k != "" {
+			input[k] = v
+		}
+	}
+	if len(input) == 0 {
+		if argsRaw := findMarkupTagValue(m[2], toolCallMarkupArgsTagNames, toolCallMarkupArgsPatternByTag); argsRaw != "" {
+			input = parseMarkupInput(argsRaw)
+		} else if kv := parseMarkupKVObject(m[2]); len(kv) > 0 {
+			input = kv
+		}
+	}
+	return ParsedToolCall{Name: name, Input: input}, true
+}
+
+func asString(v any) string {
+	s, _ := v.(string)
+	return s
+}

internal/util/toolcalls_test.go

@@ -115,3 +115,167 @@ func TestParseStandaloneToolCallsIgnoresFencedCodeBlock(t *testing.T) {
 		t.Fatalf("expected fenced tool_call example to be ignored, got %#v", calls)
 	}
 }
+
+func TestParseToolCallsAllowsQualifiedToolName(t *testing.T) {
+	text := `{"tool_calls":[{"name":"mcp.search_web","input":{"q":"golang"}}]}`
+	calls := ParseToolCalls(text, []string{"search_web"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "search_web" {
+		t.Fatalf("expected canonical tool name search_web, got %q", calls[0].Name)
+	}
+}
+
+func TestParseToolCallsAllowsPunctuationVariantToolName(t *testing.T) {
+	text := `{"tool_calls":[{"name":"read-file","input":{"path":"README.md"}}]}`
+	calls := ParseToolCalls(text, []string{"read_file"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "read_file" {
+		t.Fatalf("expected canonical tool name read_file, got %q", calls[0].Name)
+	}
+}
+
+func TestParseToolCallsSupportsClaudeXMLToolCall(t *testing.T) {
+	text := `<tool_call><tool_name>Bash</tool_name><parameters><command>pwd</command><description>show cwd</description></parameters></tool_call>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsDetailedMarksXMLToolCallSyntax(t *testing.T) {
+	text := `<tool_call><tool_name>Bash</tool_name><parameters><command>pwd</command></parameters></tool_call>`
+	res := ParseToolCallsDetailed(text, []string{"bash"})
+	if !res.SawToolCallSyntax {
+		t.Fatalf("expected SawToolCallSyntax=true, got %#v", res)
+	}
+	if len(res.Calls) != 1 {
+		t.Fatalf("expected one parsed call, got %#v", res)
+	}
+}
+
+func TestParseToolCallsSupportsClaudeXMLJSONToolCall(t *testing.T) {
+	text := `<tool_call>{"tool":"Bash","params":{"command":"pwd","description":"show cwd"}}</tool_call>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsFunctionCallTagStyle(t *testing.T) {
+	text := `<function_call>Bash</function_call><function parameter name="command">ls -la</function parameter><function parameter name="description">list</function parameter>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "ls -la" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsAntmlFunctionCallStyle(t *testing.T) {
+	text := `<antml:function_calls><antml:function_call name="Bash">{"command":"pwd","description":"x"}</antml:function_call></antml:function_calls>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsAntmlArgumentStyle(t *testing.T) {
+	text := `<antml:function_calls><antml:function_call id="1" name="Bash"><antml:argument name="command">pwd</antml:argument><antml:argument name="description">x</antml:argument></antml:function_call></antml:function_calls>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsInvokeFunctionCallStyle(t *testing.T) {
+	text := `<function_calls><invoke name="Bash"><parameter name="command">pwd</parameter><parameter name="description">d</parameter></invoke></function_calls>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsNestedToolTagStyle(t *testing.T) {
+	text := `<tool_call><tool name="Bash"><command>pwd</command><description>show cwd</description></tool></tool_call>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsAntmlFunctionAttributeWithParametersTag(t *testing.T) {
+	text := `<antml:function_calls><antml:function_call id="x" function="Bash"><antml:parameters>{"command":"pwd"}</antml:parameters></antml:function_call></antml:function_calls>`
+	calls := ParseToolCalls(text, []string{"bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "bash" {
+		t.Fatalf("expected canonical tool name bash, got %q", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "pwd" {
+		t.Fatalf("expected command argument, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsSupportsMultipleAntmlFunctionCalls(t *testing.T) {
+	text := `<antml:function_calls><antml:function_call id="1" function="Bash"><antml:parameters>{"command":"pwd"}</antml:parameters></antml:function_call><antml:function_call id="2" function="Read"><antml:parameters>{"file_path":"README.md"}</antml:parameters></antml:function_call></antml:function_calls>`
+	calls := ParseToolCalls(text, []string{"bash", "read"})
+	if len(calls) != 2 {
+		t.Fatalf("expected 2 calls, got %#v", calls)
+	}
+	if calls[0].Name != "bash" || calls[1].Name != "read" {
+		t.Fatalf("expected canonical names [bash read], got %#v", calls)
+	}
+}
+
+func TestParseToolCallsDoesNotAcceptMismatchedMarkupTags(t *testing.T) {
+	text := `<tool_call><name>read_file</function><arguments>{"path":"README.md"}</arguments></tool_call>`
+	calls := ParseToolCalls(text, []string{"read_file"})
+	if len(calls) != 0 {
+		t.Fatalf("expected mismatched tags to be rejected, got %#v", calls)
+	}
+}

internal/util/toolcalls_textkv.go

@@ -0,0 +1,55 @@
+package util
+
+import (
+	"regexp"
+	"strings"
+)
+
+var textKVNamePattern = regexp.MustCompile(`(?is)function\.name:\s*([a-zA-Z0-9_\-.]+)`)
+
+func parseTextKVToolCalls(text string) []ParsedToolCall {
+	var out []ParsedToolCall
+	matches := textKVNamePattern.FindAllStringSubmatchIndex(text, -1)
+	if len(matches) == 0 {
+		return nil
+	}
+
+	for i, match := range matches {
+		name := text[match[2]:match[3]]
+
+		offset := match[1]
+		endSearch := len(text)
+		if i+1 < len(matches) {
+			endSearch = matches[i+1][0]
+		}
+
+		searchArea := text[offset:endSearch]
+		argIdx := strings.Index(searchArea, "function.arguments:")
+		if argIdx < 0 {
+			continue
+		}
+
+		startIdx := offset + argIdx + len("function.arguments:")
+		braceIdx := strings.IndexByte(text[startIdx:endSearch], '{')
+		if braceIdx < 0 {
+			continue
+		}
+
+		actualStart := startIdx + braceIdx
+		objJson, _, ok := extractJSONObject(text, actualStart)
+		if !ok {
+			continue
+		}
+
+		input := parseToolCallInput(objJson)
+		out = append(out, ParsedToolCall{
+			Name:  name,
+			Input: input,
+		})
+	}
+
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}

internal/util/toolcalls_textkv_test.go

@@ -0,0 +1,63 @@
+package util
+
+import (
+	"testing"
+)
+
+func TestParseTextKVToolCalls_Basic(t *testing.T) {
+	text := `
+[TOOL_CALL_HISTORY]
+status: already_called
+origin: assistant
+not_user_input: true
+tool_call_id: call_3fcd15235eb94f7eae3a8de5a9cfa36b
+function.name: execute_command
+function.arguments: {"command":"cd scripts && python check_syntax.py example.py","cwd":null,"timeout":30}
+[/TOOL_CALL_HISTORY]
+
+Some other text thinking...
+`
+	calls := ParseToolCalls(text, []string{"execute_command"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %d", len(calls))
+	}
+	if calls[0].Name != "execute_command" {
+		t.Fatalf("unexpected name: %s", calls[0].Name)
+	}
+	if calls[0].Input["command"] != "cd scripts && python check_syntax.py example.py" {
+		t.Fatalf("unexpected command arg: %v", calls[0].Input["command"])
+	}
+}
+
+func TestParseTextKVToolCalls_Multiple(t *testing.T) {
+	text := `
+function.name: read_file
+function.arguments: {
+	"path": "abc.txt"
+}
+
+function.name: bash
+function.arguments: {"command": "ls"}
+`
+	calls := ParseToolCalls(text, []string{"read_file", "bash"})
+	if len(calls) != 2 {
+		t.Fatalf("expected 2 calls, got %d", len(calls))
+	}
+	if calls[0].Name != "read_file" {
+		t.Fatalf("unexpected 1st name: %s", calls[0].Name)
+	}
+	if calls[1].Name != "bash" {
+		t.Fatalf("unexpected 2nd name: %s", calls[1].Name)
+	}
+}
+
+func TestParseTextKVToolCalls_Standalone(t *testing.T) {
+	text := "function.name: read_file\nfunction.arguments: {\"path\":\"README.md\"}"
+	calls := ParseStandaloneToolCalls(text, []string{"read_file"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %d", len(calls))
+	}
+	if calls[0].Name != "read_file" {
+		t.Fatalf("unexpected name: %s", calls[0].Name)
+	}
+}

misc/deepseek_functioncalling_bug/report.md

@@ -0,0 +1,101 @@
+# DeepSeek Function Calling 缺陷分析与 ds2api 的增强修复策略
+
+> **相关 PR**: #74 (代码核心实现) 与 #75 (Merge to dev)
+> **问题背景**: 解决因包括 DeepSeek 在内的部分模型在函数调用（Function Calling/Tool Call）表现不够“规范”，从而导致工具调用失败的问题。
+
+## 一、底层架构对比：为什么会产生 Function Calling 缺陷？
+
+在探讨缺陷前，我们需要理解两种 Function Calling 的底层结构差异：
+
+### 1. OpenAI 的原生结构化返回 (API 级分离)
+在 OpenAI 的规范中，**聊天文字与工具调用是在底层的 JSON 结构中被硬性拆分的**：
+* 聊天废话存放在 `response.choices[0].message.content` 里。
+* 工具请求存放在单独的数组 `response.choices[0].message.tool_calls` 里。
+
+**优势：** 这种设计对客户端极其友好。客户端只需判断 `tool_calls` 是否为空，就能决定是执行代码还是渲染文字。它支持同时并发多个工具请求，且底层的生成殷勤被严格训练和约束，极少抛出语法错误的 JSON。
+
+### 2. DeepSeek 等模型的“单文本流”机制
+相比之下，部分未经深度专门微调的模型（或者在特定的通信适配层中），它们依然倾向于把一切内容打包成一个纯文本流吐出。这就是为什么它们的输出往往不仅包含了本该属于 `tool_calls` 结构里的 JSON，还会像个“老实人”一样夹杂了属于 `content` 里的散文。
+
+---
+
+## 二、DeepSeek 在 Function Calling 上的特定缺陷表现
+
+相比于 OpenAI 严格遵循 API 约定的原生结构，DeepSeek 等开源/国产推理模型在工具调用时，经常会暴露出以下三种典型的“不守规矩”的输出行为：
+
+### 1. 混合输出：散文文本与工具 JSON 混杂 (Mixed Prose Streams)
+当应用要求模型直接返回工具请求时，DeepSeek 有时候会**“忍不住想和用户搭话”**。
+它常常前置一段解释性废话，中间插入工具调用的 JSON 参数，并在末尾再补上一句总结：
+```text
+好的，我这就帮你读取 README.md 的内容：
+{"tool_calls":[{"name":"read_file","input":{"path":"README.md"}}]}
+请稍等片刻，我马上把它读出来。
+```
+**旧版系统痛点：**
+原有的代码存在**严格模式（Strict Mode）**校验：
+```go
+// 如果解析到的 JSON 块前后存在任何非空字符串，就放弃当作工具调用！
+if strings.TrimSpace(state.recentTextTail) != "" || strings.TrimSpace(prefixPart) != "" ... {
+    return captured, nil, "", true
+}
+```
+这直接导致上述结构被网关认定是一段“普通聊天”，直接原封不动地返回给用户，这直接干挂了后续的工具自动执行流程。
+
+### 2. 工具名格式幻觉：擅自修改或前缀化工具名称
+由于 DeepSeek 的预训练数据中有大量的代码和不同的平台结构，它在回复工具名称时，常常无法忠实于 System Prompt 中提供的纯命名（也就是 `name: "read_file"`），而是加上前缀或者拼写变形，例如：
+* `{"name": "mcp.search_web"}` （自带命名空间）
+* `{"name": "tools.read_file"}`
+* `{"name": "search-web"}` （下划线变成了中划线）
+
+**旧版系统痛点：**
+旧版系统对于工具名的匹配几乎只有“绝对相等”的字典级比对，只要差了一个字符或加了前缀，就会由于找不到合法工具而直接失败。
+
+### 3. Role 角色的非标准返回
+在部分工具通信流的响应中，返回的内容其所属的 `role` 没有被标准化处理，可能携带意料之外的属性，或是与下游严格比对出现冲突。
+
+---
+
+## 二、PR #74 的代码增强修复方案
+
+为了解决大模型这种自身的不规范行为，PR #74 在系统的中间层网关联入了一个**极其包容的容错引擎**。它并不强制要求模型“改过自新”，而是主动做了以下三块增强：
+
+### 1. 从流中分离混合内容（废除 Strict Mode）
+修改了 `internal/adapter/openai/tool_sieve_core.go`。
+取消了前后包裹文本的拦截逻辑。当系统扫描到流式结构中有完整的 `{"tool_calls":...}` 时，它会将废话和 JSON 分发到不同的事件流中：
+```go
+if prefix != "" {
+    // 将前面的“好的，帮你读文件”剥离出来作为常规文本输出
+    state.noteText(prefix)
+    events = append(events, toolStreamEvent{Content: prefix})
+}
+// 捕获并拦截中间的工具请求，进行背后执行
+state.pendingToolCalls = calls
+```
+**效果：** 用户的屏幕上只能看到正常的文字交流，而后端的工具也会立刻挂载。
+
+### 2. 多级宽容匹配引擎 (Resolve Allowed Tool Name)
+在 `internal/util/toolcalls_parse.go` 中，新增了一个由严到松降级匹配的强大漏斗策略函数 `resolveAllowedToolName`：
+
+1. **绝对匹配**：和以前一样，`read_file` == `read_file`。
+2. **忽略大小写**：`Read_File` 算作合法。
+3. **命名空间抹除**：通过寻找最后一个 `.` 来剥离前缀，强制将 `mcp.search_web` 还原出真实的 `search_web`。
+4. **终极正则清洗**：
+   引入 `var toolNameLoosePattern = regexp.MustCompile(`[^a-z0-9]+`)`。
+   这个正则剥离了字符串里所有的符号、空格、格式符。
+   将传入的 `read-file` 洗除符号成为 `readfile`，并去和系统中所有合法工具同样清洗后的版本进行比较。只要核心字母一致，即算作匹配成功。
+
+### 3. Role 归一化 (Normalize OpenAIRoleForPrompt)
+在 `internal/adapter/openai/responses_input_items.go` 等处，引入了特定的 `normalizeOpenAIRoleForPrompt(role)` 清洗，保证输入和传递给上游的 Role 枚举始终受控，消除了因为意外的身份字段传参崩溃。
+
+---
+
+## 报告总结与 tool_sieve 的本质作用
+
+PR #74 / #75 并没有从模型本身开刀，而是基于**网关应足够健壮**的设计哲学。
+
+**其实整个增强实现，本质上实现了一个名为 `tool_sieve` (工具筛子) 的中间层网关。**
+面对 DeepSeek 这种吐出一团混合了聊天文字与 JSON 面团的“不标准”数据流，`tool_sieve` 就像一个勤劳的高精度筛子，不仅人工揉开了面团：
+1. 它把散文分拣出来，塞回标准结构的 `content` 字段去展示；
+2. 剥离并清洗出有瑕疵的 JSON 块，按照 OpenAI 的标准格式小心翼翼地放进 `tool_calls` 结构里去等待执行。
+
+这意味着，即便 AI 被配置了奇怪的回复设定、加粗了强调语言，甚至是犯了标点符号拼写小失误，**只要它输出了可以拼凑成工具指令的 JSON 核心单元，整个中继层就能将其挽救，并把正确的工具结果呈现给模型和用户**。 这不仅修复了缺陷，更极大地增强了工具网关的通用性和鲁棒性。

tests/compat/expected/toolcalls_allowlist_empty.json

@@ -1,3 +1,8 @@
 {
-  "calls": []
-}
+  "calls": [],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": true,
+  "rejectedToolNames": [
+    "unknown_tool"
+  ]
+}

tests/compat/expected/toolcalls_case_insensitive_canonical.json

@@ -6,5 +6,8 @@
         "path": "README.MD"
       }
     }
-  ]
-}
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_fenced_json.json

@@ -1,3 +1,6 @@
 {
-  "calls": []
-}
+  "calls": [],
+  "sawToolCallSyntax": false,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_function_call_tag.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_invoke_attr.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_loose_normalize.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_namespace_tail_normalize.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_standalone_fenced_example.json

@@ -1,3 +1,6 @@
 {
-  "calls": []
-}
+  "calls": [],
+  "sawToolCallSyntax": false,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_standalone_mixed_prose.json

@@ -1,3 +1,6 @@
 {
-  "calls": []
-}
+  "calls": [],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_standalone_pure.json

@@ -6,5 +6,8 @@
         "path": "README.MD"
       }
     }
-  ]
-}
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_unknown_name.json

@@ -1,3 +1,8 @@
 {
-  "calls": []
-}
+  "calls": [],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": true,
+  "rejectedToolNames": [
+    "unknown_tool"
+  ]
+}

tests/compat/expected/toolcalls_xml_tool_call.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/fixtures/toolcalls/function_call_tag.json

@@ -0,0 +1,6 @@
+{
+  "text": "<function_call><function>read_file</function><parameters>{\"path\":\"README.MD\"}</parameters></function_call>",
+  "tool_names": [
+    "read_file"
+  ]
+}

tests/compat/fixtures/toolcalls/invoke_attr.json

@@ -0,0 +1,6 @@
+{
+  "text": "<invoke name=\"read_file\"><argument>{\"path\":\"README.MD\"}</argument></invoke>",
+  "tool_names": [
+    "read_file"
+  ]
+}

tests/compat/fixtures/toolcalls/loose_normalize.json

@@ -0,0 +1,6 @@
+{
+  "text": "{\"tool_calls\":[{\"name\":\"read-file\",\"input\":{\"path\":\"README.MD\"}}]}",
+  "tool_names": [
+    "read_file"
+  ]
+}

tests/compat/fixtures/toolcalls/namespace_tail_normalize.json

@@ -0,0 +1,6 @@
+{
+  "text": "{\"tool_calls\":[{\"name\":\"company.fs.read_file\",\"input\":{\"path\":\"README.MD\"}}]}",
+  "tool_names": [
+    "read_file"
+  ]
+}

tests/compat/fixtures/toolcalls/xml_tool_call.json

@@ -0,0 +1,6 @@
+{
+  "text": "<tool_call><name>read_file</name><arguments>{\"path\":\"README.MD\"}</arguments></tool_call>",
+  "tool_names": [
+    "read_file"
+  ]
+}

tests/node/js_compat_test.js

@@ -6,7 +6,7 @@ const fs = require('node:fs');
 const path = require('node:path');
 
 const chatStream = require('../../api/chat-stream.js');
-const { parseToolCalls, parseStandaloneToolCalls } = require('../../internal/js/helpers/stream-tool-sieve.js');
+const { parseToolCallsDetailed, parseStandaloneToolCallsDetailed } = require('../../internal/js/helpers/stream-tool-sieve.js');
 
 const { parseChunkForContent, estimateTokens } = chatStream.__test;
 
@@ -44,9 +44,12 @@ test('js compat: toolcall fixtures', () => {
       const fixture = readJSON(path.join(fixtureDir, file));
       const expected = readJSON(path.join(expectedDir, `toolcalls_${name}.json`));
       const mode = typeof fixture.mode === 'string' ? fixture.mode.trim().toLowerCase() : '';
-      const parser = mode === 'standalone' ? parseStandaloneToolCalls : parseToolCalls;
+      const parser = mode === 'standalone' ? parseStandaloneToolCallsDetailed : parseToolCallsDetailed;
       const got = parser(fixture.text, fixture.tool_names || []);
-      assert.deepEqual(got, expected.calls, `${name}: calls mismatch`);
+      assert.deepEqual(got.calls, expected.calls, `${name}: calls mismatch`);
+      assert.equal(got.sawToolCallSyntax, expected.sawToolCallSyntax, `${name}: sawToolCallSyntax mismatch`);
+      assert.equal(got.rejectedByPolicy, expected.rejectedByPolicy, `${name}: rejectedByPolicy mismatch`);
+      assert.deepEqual(got.rejectedToolNames, expected.rejectedToolNames, `${name}: rejectedToolNames mismatch`);
     }
   });
 

tests/node/stream-tool-sieve.test.js

@@ -94,6 +94,34 @@ test('parseToolCalls supports fenced json and function.arguments string payload'
   assert.equal(calls.length, 0);
 });
 
+test('parseToolCalls parses text-kv fallback payload', () => {
+  const text = [
+    '[TOOL_CALL_HISTORY]',
+    'function.name: execute_command',
+    'function.arguments: {"command":"cd scripts && python check_syntax.py example.py","cwd":null,"timeout":30}',
+    '[/TOOL_CALL_HISTORY]',
+    'Some other text thinking...',
+  ].join('\n');
+  const calls = parseToolCalls(text, ['execute_command']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'execute_command');
+  assert.equal(calls[0].input.command, 'cd scripts && python check_syntax.py example.py');
+});
+
+test('parseToolCalls parses multiple text-kv fallback payloads', () => {
+  const text = [
+    'function.name: read_file',
+    'function.arguments: {"path":"abc.txt"}',
+    '',
+    'function.name: bash',
+    'function.arguments: {"command":"ls"}',
+  ].join('\n');
+  const calls = parseToolCalls(text, ['read_file', 'bash']);
+  assert.equal(calls.length, 2);
+  assert.equal(calls[0].name, 'read_file');
+  assert.equal(calls[1].name, 'bash');
+});
+
 test('parseStandaloneToolCalls only matches standalone payload and ignores mixed prose', () => {
   const mixed = '这里是示例：{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}，请勿执行。';
   const standalone = '{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}';
@@ -109,7 +137,23 @@ test('parseStandaloneToolCalls ignores fenced code block tool_call examples', ()
   assert.equal(calls.length, 0);
 });
 
-test('sieve keeps late key convergence payload as plain text in strict mode', () => {
+
+test('sieve emits tool_calls in the same chunk processing tick once payload is complete', () => {
+  const state = createToolSieveState();
+  const first = processToolSieveChunk(state, '{"', ['read_file']);
+  const second = processToolSieveChunk(
+    state,
+    'tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}',
+    ['read_file'],
+  );
+  const firstCalls = first.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+  const secondCalls = second.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+  assert.equal(firstCalls.length, 0);
+  assert.equal(secondCalls.length, 1);
+  assert.equal(secondCalls[0].name, 'read_file');
+});
+
+test('sieve emits tool_calls when late key convergence forms a complete payload', () => {
   const events = runSieve(
     [
       '{"',
@@ -119,12 +163,11 @@ test('sieve keeps late key convergence payload as plain text in strict mode', ()
     ['read_file'],
   );
   const leakedText = collectText(events);
-  const hasToolCall = events.some((evt) => evt.type === 'tool_calls' && Array.isArray(evt.calls) && evt.calls.length > 0);
-  const hasToolDelta = events.some((evt) => evt.type === 'tool_call_deltas' && Array.isArray(evt.deltas) && evt.deltas.length > 0);
-  assert.equal(hasToolCall || hasToolDelta, false);
-  assert.equal(leakedText.includes('{'), true);
-  assert.equal(leakedText.toLowerCase().includes('tool_calls'), true);
+  const finalCalls = events.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'read_file');
   assert.equal(leakedText.includes('后置正文C。'), true);
+  assert.equal(leakedText.toLowerCase().includes('tool_calls'), false);
 });
 
 test('sieve keeps embedded invalid tool-like json as normal text to avoid stream stalls', () => {
@@ -249,3 +292,9 @@ test('formatOpenAIStreamToolCalls reuses ids with the same idStore', () => {
   assert.equal(second.length, 1);
   assert.equal(first[0].id, second[0].id);
 });
+
+test('parseToolCalls rejects mismatched markup tags', () => {
+  const payload = '<tool_call><name>read_file</function><arguments>{"path":"README.md"}</arguments></tool_call>';
+  const calls = parseToolCalls(payload, ['read_file']);
+  assert.equal(calls.length, 0);
+});