[Bug/错误]: Responses → Chat Completions 将并行 function calls 拆分为无效的连续 assistant 消息

[Snowy117]

[Bug/错误]: Responses → Chat Completions 将并行 function calls 拆分为无效的连续 assistant 消息

Pre-submission Checklist / 提交前检查

• I searched the existing issues and did not find this exact bug.
  • PR #590 修复了 Bailian outbound transformer 中类似的排序问题，但通用的 OpenAI Responses inbound 转换仍受影响。
• I have read the documentation.
• I am using the latest version of AxonHub.

FAQ Check / FAQ 检查

• I checked the FAQ issue and did not find this problem covered there.

Bug Description / Bug 描述

当发送到 /v1/responses 的请求包含两个连续的 function_call input item，后面跟着对应的两个 function_call_output item 时，AxonHub 会将每个 function call 分别转换为一个独立的 Chat Completions assistant 消息。

合法的 Responses input：

function_call        call_a
function_call        call_b
function_call_output call_a
function_call_output call_b

AxonHub 当前将其转发到 Chat Completions 渠道时变为：

assistant tool_calls=[call_a]
assistant tool_calls=[call_b]
tool      tool_call_id=call_a
tool      tool_call_id=call_b

这是非法的 Chat Completions 顺序。第一条 assistant 消息声明了 call_a 后，必须立即跟随对应的 tool 消息，而不是另一条 assistant 消息。严格的上游会拒绝该请求并报告 call_a 缺少响应。

这两个连续的 Responses function_call item 代表同一次 assistant turn 中的并行调用，应合并为一条 Chat Completions assistant 消息。

Steps to Reproduce / 复现步骤

export AXONHUB_BASE_URL='https://your-axonhub.example.com'
export AXONHUB_API_KEY='replace-with-your-api-key'

curl --silent --show-error \
  --request POST \
  "${AXONHUB_BASE_URL}/v1/responses" \
  --header "Authorization: Bearer ${AXONHUB_API_KEY}" \
  --header 'Content-Type: application/json' \
  --header 'User-Agent: axonhub-consecutive-function-calls-repro/1.0' \
  --data-binary @- <<'JSON'
{
  "model": "kimi-k3",
  "store": false,
  "stream": false,
  "input": [
    {
      "role": "user",
      "content": "Reply with exactly OK."
    },
    {
      "type": "function_call",
      "call_id": "call_a",
      "name": "first_tool",
      "arguments": "{}"
    },
    {
      "type": "function_call",
      "call_id": "call_b",
      "name": "second_tool",
      "arguments": "{}"
    },
    {
      "type": "function_call_output",
      "call_id": "call_a",
      "output": "first result"
    },
    {
      "type": "function_call_output",
      "call_id": "call_b",
      "output": "second result"
    },
    {
      "role": "user",
      "content": "Reply with exactly OK."
    }
  ],
  "max_output_tokens": 32
}
JSON

kimi官方渠道。

Expected Behavior / 期望行为

AxonHub 应将连续的 function_call item 合并为一条内部 assistant 消息，以保留单次 assistant turn。Chat Completions payload 应具有以下形状：

[
  {
    "role": "assistant",
    "tool_calls": [
      {
        "id": "call_a",
        "type": "function",
        "function": { "name": "first_tool", "arguments": "{}" }
      },
      {
        "id": "call_b",
        "type": "function",
        "function": { "name": "second_tool", "arguments": "{}" }
      }
    ]
  },
  {
    "role": "tool",
    "tool_call_id": "call_a",
    "content": "first result"
  },
  {
    "role": "tool",
    "tool_call_id": "call_b",
    "content": "second result"
  },
  {
    "role": "user",
    "content": "Reply with exactly OK."
  }
]

上游请求应正常完成，而非返回 HTTP 400。

Error Message / 错误消息

HTTP/2 400

{
  "error": {
    "message": "an assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: call_a",
    "type": "InvalidParameter",
    "code": "InvalidParameter"
  }
}

Operating System / 操作系统

Docker on Linux.

AxonHub Version / AxonHub 版本

Image: looplj/axonhub:v1.0.0-beta5
OCI version label: v1.0.0-beta5
OCI revision label: d061ac7df6aef0c5ec6cdfa9dc5002546a1c5a57

Usage Scenario / 使用场景

API.

API Format / API 格式

• 入站：OpenAI - Responses（/v1/responses）。
• 出站渠道：OpenAI - Chat Completions（/v1/chat/completions）。

Channel Type / 渠道类型

kimi OpenAI-compatible Chat Completions 渠道。

Logs / 日志

保存的入站请求结构正确：

# requests.id = 58800
1  <implicit message>     user
2  function_call          call_a  first_tool
3  function_call          call_b  second_tool
4  function_call_output   call_a
5  function_call_output   call_b
6  <implicit message>     user

保存的出站执行结构无效：

# request_executions.id = 71643
1  user
2  assistant  tool_calls=call_a
3  assistant  tool_calls=call_b
4  tool       tool_call_id=call_a
5  tool       tool_call_id=call_b
6  user

执行记录中包含：

status: failed
response_status_code: 400
format: openai/chat_completions
error: an assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: call_a

Additional Context / 其他补充信息

可能原因（以下内容由GPT生成）

问题出现在：

llm/transformer/openai/responses/inbound.go

convertInputToMessages() 对 reasoning item 有特殊处理：convertReasoningWithFollowing() 会向前扫描并将后续的 function_call item 合并到一条 assistant 消息中。

但裸 function_call 由 convertItemToMessage() 处理，该方法会为每个调用创建一条只含一个 tool call 的新 assistant 消息。因此，连续的裸 function_call item 会变成连续的 assistant 消息：

convertInputToMessages                       约第 329 行
convertReasoningWithFollowing                约第 388 行
convertReasoningWithFollowing/function_call  约第 417 行
convertItemToMessage/function_call           约第 519 行

当前源码参考：

axonhub/llm/transformer/openai/responses/inbound.go

Line 329 in 35a8e5b

func convertInputToMessages(input *Input) ([]llm.Message, error) {

建议修复

当 convertInputToMessages() 遇到裸 function_call 或 custom_tool_call 时，收集同一 Responses turn 中的所有连续 tool-call item，并追加到一条 llm.Message{Role: "assistant", ToolCalls: [...]} 中。遇到 message、tool output 或其他分隔 turn 的 item 时停止合并。

建议的回归用例：

1. function_call, function_call 变为一条包含两个 tool calls 的 assistant 消息。
2. function_call, function_call, function_call_output, function_call_output 保持合法的 Chat Completions 顺序。
3. 由 outputs 或 user message 分隔的 tool-call 组不会被跨 turn 合并。
4. 对连续的 custom_tool_call item 也应用同等的合并逻辑（在可表示的范围内）。

PR #590 可作为有价值的参考：它针对 Bailian outbound 合并了连续的纯 tool-call assistant 消息，但未修复通用的 Responses inbound 表示——该问题在渠道特定的 outbound 转换之前就产生了无效的消息序列。