[Bugfix][Responses API] Reuse streamed item ids in final harmony response

[he-yufeng]

Purpose

Fixes the harmony (gpt-oss) part of #59834; #59307 fixed the non-harmony path. When /v1/responses streams, response.completed rebuilds its output from the harmony messages with harmony_to_response_output. That rebuild mints a new item id and function call_id, which no longer match the ones sent in response.output_item.added/done, so strict clients abort.

• _process_harmony_streaming_events records, for each completed harmony message, the items sent in its output_item.done events (HarmonyContext.streamed_items_by_message).
• responses_full_generator matches each message to its own record by object identity. reuse_streamed_item_ids then copies id/call_id onto that message's rebuilt items by type. Content and incomplete status still come from the rebuild. Items that were never streamed (zero-delta) keep a new id.
• Streamed harmony reasoning items now use the rs_ prefix instead of msg_, matching the non-streaming path and the non-harmony path.

Matching per message matters. A built-in python call streams as code_interpreter_call but is rebuilt as reasoning. If ids were matched across messages, that call would take the streamed id of the next reasoning item.

Testing

vllm serve openai/gpt-oss-20b --no-enable-prefix-caching (1x H200), then the request from the issue:

curl -sN localhost:8000/v1/responses -H 'Content-Type: application/json' -d '{
  "model": "openai/gpt-oss-20b", "stream": true,
  "input": "Call the get_weather tool for Paris. You must call the tool now.",
  "tools": [{"type": "function", "name": "get_weather", "parameters": {"type": "object",
    "properties": {"city": {"type": "string"}}, "required": ["city"]}}]
}' | sed -n 's/^data: //p' | jq -c '
  if .type == "response.output_item.done" then ["done", .item.type, .item.id, .item.call_id]
  elif .type == "response.completed" then .response.output[] | ["completed", .type, .id, .call_id]
  else empty end'

main (5f30fc7): completed changes every id.

["done","reasoning","msg_832f7cd8a3888df4",null]
["done","function_call","fc_8019379c1a695c28","call_b07a9991be290d4c"]
["completed","reasoning","rs_b5c9be54d08ca702",null]
["completed","function_call","fc_bc041d3276faf22e","call_bc041d3276faf22e"]

This PR: completed keeps the streamed ids.

["done","reasoning","rs_a972e97aaf1f0eb1",null]
["done","function_call","fc_be95d8cf99657077","call_84160cffacc40ff2"]
["completed","reasoning","rs_a972e97aaf1f0eb1",null]
["completed","function_call","fc_be95d8cf99657077","call_84160cffacc40ff2"]

I ran the same check, 3 times each, on a forced tool call, a reasoning + text answer, and a max_output_tokens=40 cut-off (incomplete). Results: main had id mismatches in 9/9 runs, and this PR in 0/9.

Other tests:

• tests/entrypoints/openai/responses/test_serving_responses.py: 64 passed. The 2 new tests fail on main.
• tests/entrypoints/openai/responses/test_harmony.py (gpt-oss-20b e2e): 37 passed, 2 skipped, 1 xpassed on both main and this PR.

No change to generated content: only item ids change, so no eval is needed.

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[sfeng33]

Thanks @he-yufeng for tracking down the harmony path. I pushed a commit that switches the id overlay from a (type, name) scan across all messages to per-message pairing. The scan can give an item another item's id: a built-in python call streams as code_interpreter_call but is rebuilt as reasoning, so it took the streamed id of the next reasoning item. Each completed harmony message now records its own output_item.done items, and only that message's rebuilt items reuse those ids.

[sfeng33]

/ci run --allow-stale

[github-actions]

✅ Triggered Buildkite CI #92773 for commit 1d595d05d080.

⚠️ This PR is 6 commits behind upstream main. Running CI at your own risk because --allow-stale was requested; outdated CI configuration may cause failures. Before merging, merge or rebase onto the latest main, then rerun /ci run on the latest PR commit.

[vllm-agent]

CI selector (shadow): 3 test steps (3 jobs) instead of 38 (51 jobs)

Shadow mode: this changes nothing about what CI runs. It shows what the evidence-based selector would pick for this PR, next to today's rules. How it works.

Feedback welcome: reply here if it would skip a step this change needs, or runs something unrelated.

steps (jobs)	Today's rules	Selector	Would skip	Would add
NVIDIA, CPU and others	38 (51)	3 (3)	35 (48)	0 (0)
AMD mirrors	37 (45)	3 (3)	34 (42)	0 (0)

Selector would run (3)

• entrypoints-integration-multimodal
• entrypoints-integration-responses-api
• entrypoints-unit-tests

Would skip (today's rules run them) (35)

• basic-correctness ×2
• basic-correctness-cpu-offload
• basic-correctness-cumem
• basic-correctness-prefetch-offload
• basic-correctness-sleep-mode
• basic-models-test-other-cpu
• basic-models-tests-initialization
• basic-models-tests-other
• benchmarks-cli-test
• cpu-language-generation-and-pooling-model-tests ×3
• entrypoints-integration-api-server ×4
• entrypoints-integration-api-server-generate
• entrypoints-integration-api-server-openai-chat_completion
• entrypoints-integration-api-server-openai-completion
• entrypoints-integration-llm
• entrypoints-integration-pooling
• entrypoints-integration-speech_to_text
• examples
• kernels-fla-ops-test-b200
• kernels-mhc-test-b200
• kernels-root-misc-test-b200
• language-models-tests-granite-l4-compatibility
• language-models-tests-hybrid ×2
• language-models-tests-standard
• multi-modal-models-standard-1-qwen2
• multi-modal-models-standard-2-qwen3-gemma
• multi-modal-models-standard-3-llava-qwen2-vl
• multi-modal-models-standard-4-other-whisper
• multi-modal-processor ×4
• multi-modal-processor-cpu ×4
• openai-api-correctness
• rust-frontend-core-correctness
• rust-frontend-openai-coverage
• rust-frontend-serve-admin-coverage
• rust-frontend-tool-use

Would add (today's rules do not run them) (0)

none

AMD mirrors: would skip (34)

• basic-correctness ×2
• basic-correctness-cpu-offload
• basic-correctness-cumem
• basic-correctness-prefetch-offload
• basic-correctness-sleep-mode
• basic-models-tests-initialization
• basic-models-tests-other
• benchmarks-cli-test
• entrypoints-integration-api-server ×4
• entrypoints-integration-api-server-generate
• entrypoints-integration-api-server-openai-chat_completion
• entrypoints-integration-api-server-openai-completion
• entrypoints-integration-llm
• entrypoints-integration-pooling
• entrypoints-integration-speech_to_text
• examples
• kernels-fla-ops-test-b200
• kernels-mhc-test-b200
• kernels-root-misc-test-b200
• language-models-tests-granite-l4-compatibility
• language-models-tests-hybrid ×2
• language-models-tests-standard
• multi-modal-models-standard-1-qwen2
• multi-modal-models-standard-2-qwen3-gemma
• multi-modal-models-standard-3-llava-qwen2-vl
• multi-modal-models-standard-4-other-whisper
• multi-modal-processor ×4
• openai-api-correctness
• platform-tests
• pytorch-compilation-passes-unit-tests
• rust-frontend-core-correctness
• rust-frontend-openai-coverage
• rust-frontend-serve-admin-coverage
• rust-frontend-tool-use

AMD mirrors: would add (0)

none

_{5 changed files · base 5f30fc7031 · head 1d595d05d0 · Python record: build 92769 at 84bcbc6264 · kernel record: table 84bcbc6 (build 92769), map 84bcbc6 · not counted: 10 build steps, 5 A100 steps the generator no longer emits, 13 optional steps the selector would also run}