fix claude message tool streaming

Co-Authored-By: Codex <noreply@openai.com>

services/anthropic_protocol.py

@@ -0,0 +1,262 @@
+from __future__ import annotations
+
+import html
+import json
+import re
+import time
+import uuid
+from collections.abc import Callable, Iterable, Iterator
+from typing import Any
+
+XML_TOOL_RULE = """Tool output adapter: when calling tools, output ONLY this XML and no prose/markdown:
+<tool_calls><tool_call><tool_name>TOOL_NAME</tool_name><parameters><PARAM><![CDATA[value]]></PARAM></parameters></tool_call></tool_calls>"""
+
+
+def _tool_meta(tool: dict[str, object]) -> tuple[str, str, object]:
+    fn = tool.get("function") if isinstance(tool.get("function"), dict) else {}
+    name = str(tool.get("name") or fn.get("name") or "").strip()
+    desc = str(tool.get("description") or fn.get("description") or "").strip()
+    schema = tool.get("input_schema") or tool.get("parameters") or fn.get("input_schema") or fn.get("parameters") or {}
+    return name, desc, schema
+
+
+def build_tool_prompt(tools: object) -> str:
+    if not isinstance(tools, list):
+        return ""
+    blocks = []
+    for tool in tools:
+        if not isinstance(tool, dict):
+            continue
+        name, desc, schema = _tool_meta(tool)
+        if name:
+            blocks.append(f"Tool: {name}\nDescription: {desc}\nParameters: {json.dumps(schema, ensure_ascii=False)}")
+    if not blocks:
+        return ""
+    return "Available tools:\n" + "\n".join(blocks) + """
+
+Tool use rules:
+- If the user asks to list/read/search files, inspect project state, run a command, or answer from local code, you MUST call a suitable tool first. Do not say you cannot access files.
+- To call tools, output ONLY XML and no prose/markdown:
+<tool_calls><tool_call><tool_name>TOOL_NAME</tool_name><parameters><PARAM><![CDATA[value]]></PARAM></parameters></tool_call></tool_calls>
+- Put parameters under <parameters> using the exact schema names.
+""".strip()
+
+
+def merge_system(system: object, extra: str) -> object:
+    system = compact_system(system)
+    if _has_claude_code_system(system):
+        extra = XML_TOOL_RULE
+    if not extra:
+        return system
+    if isinstance(system, str) and system.strip():
+        return f"{system.strip()}\n\n{extra}"
+    if isinstance(system, list):
+        return [*system, {"type": "text", "text": extra}]
+    return extra
+
+
+def _has_claude_code_system(system: object) -> bool:
+    if isinstance(system, str):
+        return "You are Claude Code" in system
+    if isinstance(system, list):
+        return any(isinstance(item, dict) and "You are Claude Code" in str(item.get("text") or "") for item in system)
+    return False
+
+
+def compact_system(system: object) -> object:
+    if isinstance(system, str):
+        return _compact_system_text(system)
+    if isinstance(system, list):
+        result = []
+        for item in system:
+            if isinstance(item, dict) and str(item.get("type") or "") == "text":
+                copied = dict(item)
+                copied["text"] = _compact_system_text(str(item.get("text") or ""))
+                result.append(copied)
+            else:
+                result.append(item)
+        return result
+    return system
+
+
+def _compact_system_text(text: str) -> str:
+    return text or ""
+
+
+def _compact_message_text(text: str) -> str:
+    return text or ""
+
+
+def preprocess_payload(payload: dict[str, object], text_mapper: Callable[[str], str] | None = None) -> dict[str, object]:
+    payload["messages"] = preprocess_messages(payload.get("messages"), text_mapper)
+    payload["system"] = merge_system(payload.get("system"), build_tool_prompt(payload.get("tools")))
+    return payload
+
+
+def preprocess_messages(messages: object, text_mapper: Callable[[str], str] | None = None) -> object:
+    if not isinstance(messages, list):
+        return messages
+    mapper = text_mapper or (lambda text: text)
+    result = []
+    for message in messages:
+        if not isinstance(message, dict):
+            continue
+        item = dict(message)
+        content = item.get("content")
+        if isinstance(content, str):
+            item["content"] = _compact_message_text(mapper(content))
+        elif isinstance(content, list):
+            item["content"] = [_preprocess_block(block, mapper) for block in content]
+        result.append(item)
+    return result
+
+
+def _preprocess_block(block: object, text_mapper: Callable[[str], str]) -> object:
+    if not isinstance(block, dict):
+        return block
+    block_type = str(block.get("type") or "")
+    if block_type == "text":
+        item = dict(block)
+        item["text"] = _compact_message_text(text_mapper(str(block.get("text") or "")))
+        return item
+    if block_type == "tool_use":
+        return {"type": "text", "text": f"<tool_calls><tool_call><tool_name>{block.get('name') or ''}</tool_name><parameters>{json.dumps(block.get('input') or {}, ensure_ascii=False)}</parameters></tool_call></tool_calls>"}
+    if block_type == "tool_result":
+        return {"type": "text", "text": f"Tool result {block.get('tool_use_id') or ''}: {block.get('content') or ''}"}
+    return block
+
+
+def message_response(model: str, text: str, input_tokens: int, output_tokens: int, tools: object = None) -> dict[str, object]:
+    content, stop_reason = content_blocks(text, tools)
+    return {
+        "id": f"msg_{uuid.uuid4()}",
+        "type": "message",
+        "role": "assistant",
+        "model": model,
+        "content": content,
+        "stop_reason": stop_reason,
+        "stop_sequence": None,
+        "usage": {"input_tokens": input_tokens, "output_tokens": output_tokens},
+    }
+
+
+def content_blocks(text: str, tools: object = None) -> tuple[list[dict[str, object]], str]:
+    calls = parse_tool_calls(text) if isinstance(tools, list) and tools else []
+    text = strip_tool_markup(text)
+    if calls:
+        content = ([{"type": "text", "text": text}] if text else []) + [{"type": "tool_use", "id": f"toolu_{uuid.uuid4()}", "name": name, "input": args} for name, args in calls]
+        return content, "tool_use"
+    return [{"type": "text", "text": text}], "end_turn"
+
+
+def strip_tool_markup(text: str) -> str:
+    return re.sub(r"(?is)<tool_calls\b[^>]*>.*?</tool_calls>|<tool_call\b[^>]*>.*?</tool_call>|<function_call\b[^>]*>.*?</function_call>|<invoke\b[^>]*>.*?</invoke>", "", text or "").strip()
+
+
+def streamable_text(text: str) -> str:
+    text = text or ""
+    match = re.search(r"(?is)<tool_calls\b|<tool_call\b|<function_call\b|<invoke\b", text)
+    return text[:match.start()].rstrip() if match else text
+
+
+def parse_tool_calls(text: str) -> list[tuple[str, dict[str, object]]]:
+    text = re.sub(r"(?is)```.*?```", "", text or "").strip()
+    blocks = re.findall(r"(?is)<tool_call\b[^>]*>(.*?)</tool_call>|<function_call\b[^>]*>(.*?)</function_call>|<invoke\b[^>]*>(.*?)</invoke>", text)
+    result = []
+    for block in (next((part for part in match if part), "") for match in blocks):
+        name = xml_value(block, "tool_name") or xml_value(block, "name") or xml_value(block, "function")
+        params = xml_value(block, "parameters") or xml_value(block, "input") or xml_value(block, "arguments") or "{}"
+        if name:
+            result.append((name, parse_tool_params(params)))
+    return result
+
+
+def xml_value(text: str, tag: str) -> str:
+    match = re.search(rf"(?is)<{tag}\b[^>]*>(.*?)</{tag}>", text)
+    if not match:
+        return ""
+    value = match.group(1).strip()
+    cdata = re.fullmatch(r"(?is)<!\[CDATA\[(.*?)]]>", value)
+    return html.unescape(cdata.group(1) if cdata else value).strip()
+
+
+def parse_tool_params(raw: str) -> dict[str, object]:
+    raw = raw.strip()
+    try:
+        parsed = json.loads(raw)
+        return parsed if isinstance(parsed, dict) else {}
+    except Exception:
+        return {m.group(1): parse_tool_value(m.group(2)) for m in re.finditer(r"(?is)<([\w.-]+)\b[^>]*>(.*?)</\1>", raw)}
+
+
+def parse_tool_value(raw: str) -> object:
+    value = xml_value(f"<x>{raw}</x>", "x")
+    try:
+        return json.loads(value)
+    except Exception:
+        return value
+
+
+def stream_events(chunks: Iterable[dict[str, object]], model: str, input_tokens: int, output_tokens: Callable[[str], int], tools: object = None) -> Iterator[dict[str, object]]:
+    message_id = f"msg_{uuid.uuid4()}"
+    created = int(time.time())
+    current_text = ""
+    streamed_text = ""
+    tool_mode = isinstance(tools, list) and bool(tools)
+    tool_started = False
+    text_open = False
+    yield {"type": "message_start", "message": {"id": message_id, "type": "message", "role": "assistant", "model": model, "content": [], "stop_reason": None, "stop_sequence": None, "usage": {"input_tokens": input_tokens, "output_tokens": 0}}}
+    if not tool_mode:
+        text_open = True
+        yield {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}
+    for chunk in chunks:
+        choice = (chunk.get("choices") or [{}])[0]
+        delta = choice.get("delta") or {}
+        text_delta = delta.get("content", "") if isinstance(delta, dict) else ""
+        if text_delta:
+            current_text += text_delta
+            if not tool_started:
+                visible_text = current_text if not tool_mode else streamable_text(current_text)
+                if visible_text.startswith(streamed_text):
+                    text_delta = visible_text[len(streamed_text):]
+                    if text_delta:
+                        if not text_open:
+                            text_open = True
+                            yield {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}
+                        streamed_text = visible_text
+                        yield {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": text_delta}}
+                tool_started = tool_mode and visible_text != current_text
+        if choice.get("finish_reason"):
+            content, stop_reason = content_blocks(current_text, tools)
+            if text_open:
+                yield {"type": "content_block_stop", "index": 0}
+            if stop_reason == "tool_use":
+                start_index = 1 if text_open else 0
+                if content and content[0]["type"] == "text":
+                    remaining = str(content[0].get("text") or "")[len(streamed_text):]
+                    if remaining:
+                        if not text_open:
+                            yield {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}
+                        yield {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": remaining}}
+                        if not text_open:
+                            yield {"type": "content_block_stop", "index": 0}
+                    start_index = 1
+                    content = content[1:]
+                yield from _stream_buffered_blocks(content, start_index)
+            yield {"type": "message_delta", "delta": {"stop_reason": stop_reason, "stop_sequence": None}, "usage": {"output_tokens": output_tokens(current_text)}}
+            break
+    yield {"type": "message_stop", "created": created}
+
+
+def _stream_buffered_blocks(content: list[dict[str, object]], start_index: int = 0) -> Iterator[dict[str, object]]:
+    for offset, block in enumerate(content):
+        index = start_index + offset
+        if block["type"] == "tool_use":
+            start = {"type": "tool_use", "id": block["id"], "name": block["name"], "input": {}}
+            delta = {"type": "input_json_delta", "partial_json": json.dumps(block.get("input") or {}, ensure_ascii=False)}
+        else:
+            start = {"type": "text", "text": ""}
+            delta = {"type": "text_delta", "text": block.get("text") or ""}
+        yield {"type": "content_block_start", "index": index, "content_block": start}
+        yield {"type": "content_block_delta", "index": index, "delta": delta}
+        yield {"type": "content_block_stop", "index": index}

services/chatgpt_service.py

@@ -11,6 +11,7 @@ from typing import Any, Iterable, Iterator
 from fastapi import HTTPException
 
 from services.account_service import AccountService
+from services.anthropic_protocol import preprocess_payload
 from services.config import config
 from services.openai_backend_api import CODEX_IMAGE_MODEL, OpenAIBackendAPI
 from utils.helper import (
@@ -83,8 +84,14 @@ class ChatGPTService:
         return OpenAIBackendAPI(access_token=access_token)
 
     def _get_text_access_token(self) -> str:
-        tokens = self.account_service.list_tokens()
-        return tokens[0] if tokens else ""
+        try:
+            for access_token in self.account_service.list_tokens():
+                status = str((self.account_service.get_account(access_token) or {}).get("status") or "").strip()
+                if status not in {"禁用", "异常"}:
+                    return access_token
+        except Exception:
+            pass
+        return ""
 
     @staticmethod
     def _encode_images(images: Iterable[tuple[bytes, str, str]]) -> list[str]:
@@ -1071,6 +1078,7 @@ class ChatGPTService:
             raise HTTPException(status_code=502, detail={"error": str(exc)}) from exc
 
     def create_message(self, body: dict[str, object]) -> dict[str, object]:
+        body = preprocess_payload(dict(body))
         model = str(body.get("model") or "auto").strip() or "auto"
         messages = self._chat_messages_from_body(body)
         try:
@@ -1079,11 +1087,13 @@ class ChatGPTService:
                 model=model,
                 stream=False,
                 system=body.get("system"),
+                tools=body.get("tools"),
             )
         except Exception as exc:
             raise HTTPException(status_code=502, detail={"error": str(exc)}) from exc
 
     def stream_message(self, body: dict[str, object]) -> Iterator[dict[str, object]]:
+        body = preprocess_payload(dict(body))
         model = str(body.get("model") or "auto").strip() or "auto"
         messages = self._chat_messages_from_body(body)
         try:
@@ -1092,6 +1102,7 @@ class ChatGPTService:
                 model=model,
                 stream=True,
                 system=body.get("system"),
+                tools=body.get("tools"),
             )
         except Exception as exc:
             raise HTTPException(status_code=502, detail={"error": str(exc)}) from exc

services/openai_backend_api.py

@@ -13,6 +13,7 @@ from curl_cffi import requests
 from PIL import Image
 
 from services.account_service import account_service
+from services.anthropic_protocol import message_response, stream_events
 from services.config import config
 from services.proxy_service import proxy_settings
 from utils.helper import build_chat_image_markdown_content, ensure_ok, new_uuid, parse_sse_lines
@@ -1231,9 +1232,22 @@ class OpenAIBackendAPI:
                 return event
         return {}
 
+    def _conversation_payload_debug(self, payload: Dict[str, Any]) -> Dict[str, Any]:
+        raw = json.dumps(payload, ensure_ascii=False, separators=(",", ":"))
+        messages = payload.get("messages") if isinstance(payload.get("messages"), list) else []
+        details = []
+        for item in messages:
+            content = item.get("content") if isinstance(item, dict) else {}
+            parts = content.get("parts") if isinstance(content, dict) else []
+            text = "\n".join(str(part) for part in parts) if isinstance(parts, list) else ""
+            details.append({"role": (item.get("author") or {}).get("role") if isinstance(item, dict) and isinstance(item.get("author"), dict) else "", "chars": len(text), "preview": text[:120]})
+        return {"bytes": len(raw.encode("utf-8")), "chars": len(raw), "message_count": len(messages), "messages": details}
+
     def _stream_events(self, path: str, requirements: ChatRequirements, payload: Dict[str, Any]) -> Iterator[
         Dict[str, Any]]:
         """向 conversation 接口发起请求，并逐条产出 SSE 事件。"""
+        payload_debug = self._conversation_payload_debug(payload)
+        logger.info({"event": "conversation_payload_debug", "path": path, **payload_debug})
         response = self.session.post(
             self.base_url + path,
             headers=self._conversation_headers(path, requirements),
@@ -1241,6 +1255,8 @@ class OpenAIBackendAPI:
             timeout=300,
             stream=True,
         )
+        if response.status_code >= 400:
+            logger.warning({"event": "conversation_request_failed", "path": path, "status_code": response.status_code, "payload": payload_debug, "response_text": response.text[:500]})
         ensure_ok(response, path)
         yield from parse_sse_lines(response)
 
@@ -1445,95 +1461,14 @@ class OpenAIBackendAPI:
             }],
         }
 
-    def _anthropic_message_response(self, model: str, messages: list[Dict[str, Any]], text: str) -> Dict[str, Any]:
+    def _anthropic_message_response(self, model: str, messages: list[Dict[str, Any]], text: str, tools: Any = None) -> Dict[str, Any]:
         """把对话结果整理成 Anthropic `/v1/messages` 风格结构。"""
-        prompt_tokens = self._count_message_tokens(messages, model)
-        completion_tokens = self._count_text_tokens(text, model)
-        return {
-            "id": f"msg_{new_uuid()}",
-            "type": "message",
-            "role": "assistant",
-            "model": model,
-            "content": [{
-                "type": "text",
-                "text": text,
-            }],
-            "stop_reason": "end_turn",
-            "stop_sequence": None,
-            "usage": {
-                "input_tokens": prompt_tokens,
-                "output_tokens": completion_tokens,
-            },
-        }
+        return message_response(model, text, self._count_message_tokens(messages, model), self._count_text_tokens(text, model), tools)
 
-    def _stream_anthropic_messages(self, messages: list[Dict[str, Any]], model: str = "auto") -> Iterator[
+    def _stream_anthropic_messages(self, messages: list[Dict[str, Any]], model: str = "auto", tools: Any = None) -> Iterator[
         Dict[str, Any]]:
         """返回 Anthropic `/v1/messages` 风格的流式事件。"""
-        message_id = f"msg_{new_uuid()}"
-        created = int(time.time())
-        current_text = ""
-
-        yield {
-            "type": "message_start",
-            "message": {
-                "id": message_id,
-                "type": "message",
-                "role": "assistant",
-                "model": model,
-                "content": [],
-                "stop_reason": None,
-                "stop_sequence": None,
-                "usage": {
-                    "input_tokens": self._count_message_tokens(messages, model),
-                    "output_tokens": 0,
-                },
-            },
-        }
-        yield {
-            "type": "content_block_start",
-            "index": 0,
-            "content_block": {
-                "type": "text",
-                "text": "",
-            },
-        }
-
-        for chunk in self._stream_chat_completions(messages, model):
-            choice = (chunk.get("choices") or [{}])[0]
-            delta = choice.get("delta") or {}
-            text_delta = delta.get("content", "")
-            if text_delta:
-                current_text += text_delta
-                yield {
-                    "type": "content_block_delta",
-                    "index": 0,
-                    "delta": {
-                        "type": "text_delta",
-                        "text": text_delta,
-                    },
-                }
-
-            if choice.get("finish_reason"):
-                yield {
-                    "type": "content_block_stop",
-                    "index": 0,
-                }
-                yield {
-                    "type": "message_delta",
-                    "delta": {
-                        "stop_reason": "end_turn",
-                        "stop_sequence": None,
-                    },
-                    "usage": {
-                        "output_tokens": self._count_text_tokens(current_text, model),
-                    },
-                }
-                break
-
-        yield {
-            "type": "message_stop",
-            "created": created,
-        }
+        yield from stream_events(self._stream_chat_completions(messages, model), model, self._count_message_tokens(messages, model), lambda text: self._count_text_tokens(text, model), tools)
 
     def _iter_response_events(self, response: requests.Response) -> Iterator[Dict[str, Any]]:
         """按 Responses 接口事件格式直通输出，不额外包装标准事件。"""
@@ -1633,10 +1568,10 @@ class OpenAIBackendAPI:
         return self._chat_completion_response(model, normalized_messages, result["text"])
 
     def messages(self, messages: list[Dict[str, Any]], model: str = "auto", stream: bool = False,
-                 system: Any = None) -> Dict[str, Any] | Iterator[Dict[str, Any]]:
+                 system: Any = None, tools: Any = None) -> Dict[str, Any] | Iterator[Dict[str, Any]]:
         """返回 Anthropic `/v1/messages` 风格结果，支持 stream 模式。"""
         normalized_messages = self._normalize_messages(messages, system)
         if stream:
-            return self._stream_anthropic_messages(normalized_messages, model)
+            return self._stream_anthropic_messages(normalized_messages, model, tools)
         result = self._complete_chat(normalized_messages, model)
-        return self._anthropic_message_response(model, normalized_messages, result["text"])
+        return self._anthropic_message_response(model, normalized_messages, result["text"], tools)

test/test_account_image_capabilities.py

@@ -86,7 +86,7 @@ class AuthServiceTests(unittest.TestCase):
             self.assertEqual(item["role"], "user")
             self.assertEqual(item["name"], "Alice")
             self.assertTrue(item["enabled"])
-            self.assertTrue(raw_key.startswith("cg2a_user_"))
+            self.assertTrue(raw_key.startswith("sk-"))
 
             authed = service.authenticate(raw_key)
             self.assertIsNotNone(authed)