简述
部分按 OpenAI 流式规范分片发送工具调用参数、且续片把 id/name 写成空字符串 "" 的模型(如 glm-5.2),其工具调用会被 tool_call_accumulator 判定为 Arguments are invalid JSON. 而整轮失败重试。根因不在模型:模型实际发出了合法 JSON,但累加器只保留了每个工具调用的第一片 {,后续参数片段被丢弃。
环境
- OpenBitFun Desktop: 1.0.1(Windows)
- 模型:
glm-5.2(provider: sensenova,OpenAI 兼容流式)
- 出现会话:
20260923T213930、20260924T163057(前者参数解析失败 9 次,后者 8 次)
复现
让 glm-5.2 在一轮里并行发起两个工具调用(如 ExecCommand + Glob)。app.log 出现:
[ERROR][openbitfun_agent_stream::tool_call_accumulator] Tool call arguments parsing failed at boundary=finish_reason: tool_id=call_9072c7daa3af466782cee2ea, tool_name=Glob, raw_len=1, error=EOF while parsing an object at line 1 column 1, parse_error_category=eof, parse_error_is_eof=true, completion=Unknown
[ERROR][openbitfun_agent_stream] SSE Error: Has invalid tool calls
[ERROR][openbitfun_core::agentic::tools::pipeline::tool_pipeline] Tool execution failed: error=Validation error: Arguments are invalid JSON.
[WARN][openbitfun_core::agentic::execution::round_executor] Retrying stream because provider returned only invalid tool arguments: ... local_retry=1/10
关键:raw_len=1 —— 累加器里每个工具调用只剩 1 个字符 {。
根因(SSE 证据)
模型把同一工具调用的 arguments 拆成多个 SSE chunk 发送,首片带 id/name,续片把它们写成空串:
17: tool_calls[0] id="call_b45..." name="ExecCommand" arguments="{"
18: tool_calls[0] id="" name="" arguments="\"cmd\": "
19: tool_calls[0] id="" name="" arguments="\"Get-"
...(续片累积命令字符串)
28: tool_calls[0] id="" name="" arguments="\"}"
29: tool_calls[1] id="call_907..." name="Glob" arguments="{"
30: tool_calls[1] id="" name="" arguments="\"pattern\": "
33: tool_calls[1] id="" name="" arguments="}"
34: finish_reason="tool_calls"
把续片 arguments 顺序拼接,得到的是合法 JSON:
- ExecCommand →
{"cmd": "Get-Date; Write-Output \"PowerShell 工具运行正常"}
- Glob →
{"pattern": "*.md"}
但累加器只保留了首片的 {,续片全部丢失 → 解析到 EOF while parsing an object at line 1 column 1。
为什么"有些模型"出问题(对比)
同一台机器上 qwen3.8-27b-uncensored 基本正常。其续片是把 id/name 字段省略(而非空串):
"tool_calls":[{"index":0,"type":"function","function":{"arguments":"{"}}]
"tool_calls":[{"index":0,"type":"function","function":{"arguments":"\"file_path\":"}}]
glm-5.2 的续片则是空串:
"tool_calls":[{"index":0,"id":"","type":"function","function":{"name":"","arguments":"\"cmd\": "}}]
两种写法都符合 OpenAI 流式语义,但实现细节不同:省略字段 → serde 反序列化为 None,累加器走续片追加分支(正常);空串 "" → 累加器按 id 查归属工具调用,查不到(真 id 是 call_xxx),于是丢弃该参数片段,缓冲区永远停在 {。
(注:9/21 那次 qwen 的唯一失败是另一类:boundary=graceful_shutdown、raw_len=67、completion=Interrupted,是流被中途打断导致参数真的截断,且已累积到 67 字符,反证 qwen 的累加本身是工作的。)
期望行为
续片无论 id/name 是省略还是空串,都应视为"无 id/无 name"的续片,按 index 追加到对应工具调用的参数缓冲区。两种写法都应正确合并出完整 JSON。
修复建议(openbitfun_agent_stream::tool_call_accumulator)
- 续片归属应按
index 匹配,而不是按 id;或
- 当
id 为空串 / name 为空串时,等价于 None,走续片追加分支,而非当成新 id 去查表。
影响范围
凡按 OpenAI 规范分片发送 arguments、且续片带空串 id/name 的模型都会 100% 失败并触发整轮重试(local_retry 最多 10 次)。续片字段被省略的模型不受影响。已用本机真实 SSE 流量验证(上述日志片段为原始截取,仅脱敏用户路径)。
简述
部分按 OpenAI 流式规范分片发送工具调用参数、且续片把
id/name写成空字符串""的模型(如glm-5.2),其工具调用会被tool_call_accumulator判定为Arguments are invalid JSON.而整轮失败重试。根因不在模型:模型实际发出了合法 JSON,但累加器只保留了每个工具调用的第一片{,后续参数片段被丢弃。环境
glm-5.2(provider: sensenova,OpenAI 兼容流式)20260923T213930、20260924T163057(前者参数解析失败 9 次,后者 8 次)复现
让
glm-5.2在一轮里并行发起两个工具调用(如ExecCommand+Glob)。app.log出现:关键:
raw_len=1—— 累加器里每个工具调用只剩 1 个字符{。根因(SSE 证据)
模型把同一工具调用的
arguments拆成多个 SSE chunk 发送,首片带id/name,续片把它们写成空串:把续片
arguments顺序拼接,得到的是合法 JSON:{"cmd": "Get-Date; Write-Output \"PowerShell 工具运行正常"}{"pattern": "*.md"}但累加器只保留了首片的
{,续片全部丢失 → 解析到EOF while parsing an object at line 1 column 1。为什么"有些模型"出问题(对比)
同一台机器上
qwen3.8-27b-uncensored基本正常。其续片是把id/name字段省略(而非空串):glm-5.2 的续片则是空串:
两种写法都符合 OpenAI 流式语义,但实现细节不同:省略字段 → serde 反序列化为
None,累加器走续片追加分支(正常);空串""→ 累加器按id查归属工具调用,查不到(真 id 是call_xxx),于是丢弃该参数片段,缓冲区永远停在{。(注:9/21 那次 qwen 的唯一失败是另一类:
boundary=graceful_shutdown、raw_len=67、completion=Interrupted,是流被中途打断导致参数真的截断,且已累积到 67 字符,反证 qwen 的累加本身是工作的。)期望行为
续片无论
id/name是省略还是空串,都应视为"无 id/无 name"的续片,按index追加到对应工具调用的参数缓冲区。两种写法都应正确合并出完整 JSON。修复建议(
openbitfun_agent_stream::tool_call_accumulator)index匹配,而不是按id;或id为空串 /name为空串时,等价于None,走续片追加分支,而非当成新 id 去查表。影响范围
凡按 OpenAI 规范分片发送
arguments、且续片带空串id/name的模型都会 100% 失败并触发整轮重试(local_retry最多 10 次)。续片字段被省略的模型不受影响。已用本机真实 SSE 流量验证(上述日志片段为原始截取,仅脱敏用户路径)。