fix(codebuddy): 指纹中和支持文本块形态 content（B1.3 收尾）

2026-09-21 直证（28+8 请求交错对照，6s 间隔排除频率窗口假因）：
- 块形态 content（[{type:text}]）的指纹同样触发 11128——原实现只处理
  字符串形态，块形态请求会绕过中和直接被拒
- 替换占位符后的块形态/system/assistant/billing 残留形态全部通过
- 小写变体也触发（上游大小写不敏感）；glm-5.3-flash 复验与模型无关

sanitize_channel_markers 拆出 _neutralize_text，str 与文本块列表共用；
非文本块/非 str 块内文/user 角色照旧不动。

Author: RobbsLuo

TECHNICAL.md

@@ -184,6 +184,8 @@ class ErrKind(StrEnum):
 > 会话一旦把指纹写进历史，后续每轮（含压缩请求本身）都持续 11128。
 > 处理：出站前 `sanitize_channel_markers` 把指纹替换为占位符
 > （`CODEBUDDY_SANITIZE_CHANNEL_MARKERS=false` 关闭），客户端会话历史不受影响。
+> `content` 为文本块列表（`[{"type": "text", ...}]`）时逐块处理：2026-09-21
+> 直证块形态指纹同样触发 11128，替换占位符后通过。
 > 曾误落 INVALID → 跳过上游全部凭证、零重试直接 400（`invalid_request`），是
 > deepseek-v4.1-flash / glm-5.3-flash 报「not available on any configured
 > upstream」的根因。

src/provider/codebuddy/client.py

@@ -92,14 +92,25 @@ def _clean_history_tool_calls(body: dict[str, Any]) -> None:
     ]
 
 
+def _neutralize_text(text: str) -> tuple[str, int]:
+    """单段正文的中和：返回（替换后文本, 命中处数）。"""
+    hits = 0
+    for marker in CHANNEL_MARKERS:
+        if marker in text:
+            hits += text.count(marker)
+            text = text.replace(marker, _CHANNEL_MARKER_STANDIN)
+    return text, hits
+
+
 def sanitize_channel_markers(body: dict[str, Any]) -> int:
     """中和出站 system/assistant 正文里的「伪装其他厂商官方客户端」指纹。
 
     上游内容风控（11128 Illegal API invocation）对该类指纹整单拒绝：
     与凭证无关（换号无效）、确定性复现、user/tool 角色 / tool_calls 参数 /
     reasoning 均不触发，仅 system/assistant 的 content 命中（2026-09 实测，
-    见 TECHNICAL.md §3.2）。只改写出站副本，客户端会话历史保持原样。
-    返回替换处数（用于日志与测试）。
+    见 TECHNICAL.md §3.2）。content 为文本块列表（{"type": "text", ...}）
+    时逐块处理——块形态指纹同样触发（2026-09-21 直证）。只改写出站副本，
+    客户端会话历史保持原样。返回替换处数（用于日志与测试）。
     """
     messages = body.get("messages")
     if not isinstance(messages, list):
@@ -111,15 +122,20 @@ def sanitize_channel_markers(body: dict[str, Any]) -> int:
         if message.get("role") not in ("system", "assistant"):
             continue
         content = message.get("content")
-        if not isinstance(content, str) or not content:
-            continue
-        replaced = content
-        for marker in CHANNEL_MARKERS:
-            if marker in replaced:
-                replaced_total += replaced.count(marker)
-                replaced = replaced.replace(marker, _CHANNEL_MARKER_STANDIN)
-        if replaced != content:
-            message["content"] = replaced
+        if isinstance(content, str):
+            replaced, hits = _neutralize_text(content)
+            if hits:
+                replaced_total += hits
+                message["content"] = replaced
+        elif isinstance(content, list):
+            for block in content:
+                if not (isinstance(block, dict) and block.get("type") == "text"
+                        and isinstance(block.get("text"), str)):
+                    continue
+                replaced, hits = _neutralize_text(block["text"])
+                if hits:
+                    replaced_total += hits
+                    block["text"] = replaced
     return replaced_total
 
 

tests/test_m1b_codebuddy.py

@@ -1806,6 +1806,38 @@ def test_sanitize_channel_markers_neutralizes_prompt_roles_only():
     assert marker in args                          # tool_calls 参数不动
 
 
+def test_sanitize_channel_markers_neutralizes_text_blocks():
+    """content 为文本块列表时逐块中和（2026-09-21 实测块形态指纹同样 11128）。"""
+    from src.provider.codebuddy.client import CHANNEL_MARKERS, sanitize_channel_markers
+
+    marker = CHANNEL_MARKERS[0]
+    body = {"messages": [
+        {"role": "system", "content": [
+            {"type": "text", "text": f"前缀 {marker} 后缀"},
+            {"type": "text", "text": "无指纹"},
+            {"type": "image_url", "image_url": {"url": marker}},  # 非文本块不动
+            {"type": "text", "text": 123},                        # 非 str 不动
+            "junk",                                               # 非 dict 不动
+        ]},
+        {"role": "assistant", "content": [
+            {"type": "text", "text": marker},
+            {"type": "text", "text": f"{marker} x{marker}"},
+        ]},
+        {"role": "user", "content": [{"type": "text", "text": marker}]},  # user 不动
+    ]}
+    assert sanitize_channel_markers(body) == 4
+    blocks = body["messages"][0]["content"]
+    assert blocks[0]["text"] == "前缀 [external-client-identity] 后缀"
+    assert blocks[1]["text"] == "无指纹"
+    assert blocks[2]["image_url"]["url"] == marker
+    assert blocks[3]["text"] == 123
+    assert blocks[4] == "junk"
+    assistant_blocks = body["messages"][1]["content"]
+    assert assistant_blocks[0]["text"] == "[external-client-identity]"
+    assert assistant_blocks[1]["text"] == "[external-client-identity] x[external-client-identity]"
+    assert body["messages"][2]["content"][0]["text"] == marker
+
+
 def test_sanitize_channel_markers_counts_repeats_and_skips_noise():
     """同条消息多指纹多出处计数；非 list/非 dict/非 str content 安全跳过。"""
     from src.provider.codebuddy.client import CHANNEL_MARKERS, sanitize_channel_markers