[gpt-oss] Add IncompleteDetails to ResponsesRepsonse

[qandrew]

Purpose

• add IncompleteDetails to vllm's implementation of ResponsesResponses.
• first use case: if the max token length has been hit, we should add the reason for incompleteness. this is added in the non streaming version, the streaming version will be a follow up PR (bc it needs to be based off of [gpt-oss][1][bugfix] fix streaming final output #24466
• fix some formatting issues

I do realize there are other things we need to fix for IncompleteDetails; for now we'll guarantee that IF incomplete details is outputted, it is true. We'll still need to test flows like if the generator was interrupted abruptly, and output the reason why. And I don't think GPT-OSS does content filters for now but that is another IncompleteDetails implementation.

Test Plan

Server

(gpt_oss_edit) [[email protected] /data/users/axia/gitrepos/vllm (andrew/incomplete-details)]$ CUDA_VISIBLE_DEVICES=2,3 with-proxy vllm serve "/data/users/axia/checkpoints/gpt-oss-120b" -tp 2 --port 20001

Client

(gpt_oss_edit) [[email protected] /data/users/axia/gitrepos/vllm (andrew/incomplete-details)]$ curl http://localhost:20001/v1/responses   -H "Content-Type: application/json"   -N   -d '{
    "model": "/data/users/axia/checkpoints/gpt-oss-120b",
    "input": [
        {
            "role": "user",
            "content": "Write two paragraphs on the weather."
        }
    ],
    "temperature": 0.7,
    "max_output_tokens": 256
}' | jq
# outputs 
{
  "id": "resp_5b6f93e8efa4445497f6b6bc052b6dac",
  "created_at": 1757701130,
  "incomplete_details": {
    "reason": "max_output_tokens" # we get incomplete reasons here
  },
  "instructions": null,
  "metadata": null,
  "model": "/data/users/axia/checkpoints/gpt-oss-120b",
  "object": "response",
  "output": [
    {
      "id": "rs_541bc7e4806f4630b04703e72e14e025",
      "summary": [],
      "type": "reasoning",
      "content": [
        {
          "text": "User asks: \"Write two paragraphs on the weather.\" We need to produce two paragraphs about weather. Should be descriptive, could talk about different aspects. Probably just two paragraphs. Ensure it's coherent. No extra instructions.",
          "type": "reasoning_text"
        }
      ],
      "encrypted_content": null,
      "status": null
    },
    {
      "id": "msg_122778466dba478aa761b0542f2cf81f",
      "content": [
        {
          "annotations": [],
          "text": "The sky stretched a bruised violet this morning, the first hints of sunrise shyly peeking through a thin veil of low‑lying clouds. A gentle breeze whispered through the oak leaves, carrying the faint scent of damp earth and pine sap, while droplets of drizzle clung to the windowpanes like tiny crystal beads. The temperature hovered just above the dew point, making the air feel cool enough to pull a light sweater from the closet, yet not so cold as to bite the skin. As the sun climbed higher, its golden rays began to melt the lingering mist, turning the wet streets into shimmering ribbons that reflected the city’s hurried rhythm.\n\nBy afternoon, the weather shifted dramatically. Dark, billowing cumulonimbus towers gathered on the horizon, their edges tinged with electric blue, signalling an approaching thunderstorm. A sudden gust surged through the streets, rattling shutters and sending loose papers swirling in a chaotic dance. The first crack of thunder rolled like distant drums, followed quickly by a cascade of rain that",
          "type": "output_text",
          "logprobs": null
        }
      ],
      "role": "assistant",
      "status": "completed",
      "type": "message"
    }
  ],
  "parallel_tool_calls": true,
  "temperature": 0.7,
  "tool_choice": "auto",
  "tools": [],
  "top_p": 1.0,
  "background": false,
  "max_output_tokens": 256,
  "max_tool_calls": null,
  "previous_response_id": null,
  "prompt": null,
  "reasoning": null,
  "service_tier": "auto",
  "status": "incomplete",
  "text": null,
  "top_logprobs": null,
  "truncation": "disabled",
  "usage": {
    "input_tokens": 72,
    "input_tokens_details": {
      "cached_tokens": 64
    },
    "output_tokens": 256,
    "output_tokens_details": {
      "reasoning_tokens": 44,
      "tool_output_tokens": 0
    },
    "total_tokens": 328
  },
  "user": null
}

Client, where we don't hit max token limit

(gpt_oss_edit) [[email protected] /data/users/axia/gitrepos/vllm (andrew/incomplete-details)]$ curl http://localhost:20001/v1/responses   -H "Content-Type: application/json"   -N   -d '{
    "model": "/data/users/axia/checkpoints/gpt-oss-120b",
    "input": [
        {
            "role": "user",
            "content": "Write two words on the weather."
        }
    ],
    "temperature": 0.7,
    "max_output_tokens": 256
}' 
# output
{
  "id": "resp_60215e7212c044669bf5b85015fe19e4",
  "created_at": 1757701190,
  "incomplete_details": null,
  "instructions": null,
  "metadata": null,
  "model": "/data/users/axia/checkpoints/gpt-oss-120b",
  "object": "response",
  "output": [
    {
      "id": "rs_02262a1cf9cf4ff3891d07d282f2e176",
      "summary": [],
      "type": "reasoning",
      "content": [
        {
          "text": "The user asks: \"Write two words on the weather.\" Probably they want a short phrase of two words describing the weather. Could be \"sunny skies\", \"stormy night\", etc. Provide two words. Could also be a short description: \"Cloudy day\". Provide two words. Probably just output two words. I'll respond with two words.",
          "type": "reasoning_text"
        }
      ],
      "encrypted_content": null,
      "status": null
    },
    {
      "id": "msg_52743c005ca94adf9687a4ffebb0107e",
      "content": [
        {
          "annotations": [],
          "text": "Sunny skies.",
          "type": "output_text",
          "logprobs": null
        }
      ],
      "role": "assistant",
      "status": "completed",
      "type": "message"
    }
  ],
  "parallel_tool_calls": true,
  "temperature": 0.7,
  "tool_choice": "auto",
  "tools": [],
  "top_p": 1.0,
  "background": false,
  "max_output_tokens": 256,
  "max_tool_calls": null,
  "previous_response_id": null,
  "prompt": null,
  "reasoning": null,
  "service_tier": "auto",
  "status": "completed",
  "text": null,
  "top_logprobs": null,
  "truncation": "disabled",
  "usage": {
    "input_tokens": 72,
    "input_tokens_details": {
      "cached_tokens": 64
    },
    "output_tokens": 84,
    "output_tokens_details": {
      "reasoning_tokens": 72,
      "tool_output_tokens": 0
    },
    "total_tokens": 156
  },
  "user": null
}

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @qandrew.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[yeqcharlotte]

thanks for the test plan! cc: @heheda12345 @houseroad

[yeqcharlotte]

what's missing from current logic btw.

[qandrew]

I don't think VLLM baseline implementation supports content filter as an abort reason currently: https://github.com/vllm-project/vllm/blob/main/vllm/v1/request.py#L206

[qandrew]

if the parser still has messages (ie if the generator got cut abruptly, this should be incomplete and not completed.

[yeqcharlotte]

make sense please move your comment to the code :)

[qandrew]

I don't think VLLM baseline implementation supports content filter as an abort reason currently: https://github.com/vllm-project/vllm/blob/main/vllm/v1/request.py#L206

[yeqcharlotte]

@qandrew @houseroad this needs a rebase over the structured output disable commit