[Feedback]: Every prompt results in `Error during ACP method session/prompt`

[dekimir]

logs.zip

Feedback

Every prompt results in this error:

• Error during ACP method session/prompt
  ⎿ 400 Bad Request: 3 validation errors:
      {'type': 'missing', 'loc': ('body', 1, 'ChatCompletionMessageFunctionToolCallParam', 'function'), 'msg': 'Field required', 'input': {'id': '019fdcf3-e6aa-76e5-9b65-059d74ffa3c6',
    'type': 'function'}}
      {'type': 'missing', 'loc': ('body', 1, 'ChatCompletionMessageCustomToolCallParam', 'custom'), 'msg': 'Field required', 'input': {'id': '019fdcf3-e6aa-76e5-9b65-059d74ffa3c6',
    'type': 'function'}}
      {'type': 'literal_error', 'loc': ('body', 1, 'ChatCompletionMessageCustomToolCallParam', 'type'), 'msg': "Input should be 'custom'", 'input': 'function', 'ctx': {'expected':
    "'custom'"}}

      File "/atlas/.venv/lib/python3.12/site-packages/vllm/entrypoints/serve/utils/api_utils.py", line 40, in create_chat_completion
        POST /v1/chat/completions [{'type': 'missing', 'loc': ('body', 1, 'ChatCompletionMessageFunctionToolCallParam', 'function'), 'msg': 'Field required', 'input': {'id': '019fdcf3-
    e6aa-76e5-9b65-059d74ffa3c6', 'type': 'function'}}, {'type': 'missing', 'loc': ('body', 1, 'ChatCompletionMessageCustomToolCallParam', 'custom'), 'msg': 'Field required', 'input':
    {'id': '019fdcf3-e6aa-76e5-9b65-059d74ffa3c6', 'type': 'function'}}, {'type': 'literal_error', 'loc': ('body', 1, 'ChatCompletionMessageCustomToolCallParam', 'type'), 'msg': "Input
    should be 'custom'", 'input': 'function',

This started after updating to 1.0.15.

Session

• Pool version: 1.0.15
• OS/arch: darwin/arm64
• Session ID: 019fdcf3-b172-775a-bc42-793434e9d088
• Agent server: Poolside
• Agent command: /Users/dejanmircevski/.local/bin/pool
• Model: riemann1_turbo4
• Timestamp: 2026-08-07T18:05:07+02:00

Logs

Attach this zip before submitting:

/var/folders/17/pmk0k7k16b36zqcfvdwvtlhr0000gn/T/pool-logs/20260807-180507-019fdcf3-b172-775a-bc42-793434e9d088/logs.zip

Logs and trajectory contain full prompt and response text from the session. Review before submitting.

[IgorGanapolsky]

The 400 body is a client serialization bug, not vLLM rejecting a valid OpenAI request.

pool is sending a prior-turn tool call as { "id": "…", "type": "function" } with no function object. vLLM then tries both ChatCompletionMessageFunctionToolCallParam and ChatCompletionMessageCustomToolCallParam and both fail (missing function, and type is not custom).

That matches “works for a few turns, then every later prompt 400s”: the stub gets appended to history and every subsequent session/prompt replays it.

Suggested contract: never echo a tool_call into /v1/chat/completions unless function.name (and arguments) are present; drop or repair incomplete stubs before the next turn.

pool 1.0.15 on macOS here; source is not in this repo so this stays a contract note. Logs in the original report already show the exact payload. Not affiliated.