[Bug]: standalone Responses API input, streaming, and tool continuation failures with Agentic API

hello, i'm Francisco, maintainer of vLLM Agentic API. i'm testing streaming, conversation continuation, and tool calls with Qwen3-8B and GPT-OSS-20B against TensorRT-LLM's standalone /v1/responses. our Dynamo → TensorRT-LLM integration worked, but the standalone path exposed these issues:

• input conversion: typed messages with string content become empty on the Qwen path; empty strings raise an error.
• output replay: returned logprobs:null can fail input validation. summary-only, empty, or multipart reasoning items can also fail or lose content.
• streaming: item IDs, call IDs, and output indices can disagree with the terminal response. Harmony tool-call events are missing, and incomplete responses get a response.completed event.
• continuation: streamed responses aren't saved by the frontend path even with storage enabled. previous_response_id with store:false also drops prior history.
• tool results: stored chat history can lose tool calls; GPT-OSS/Harmony misses some call channels/recipients and fails to pair replayed calls with their results.
• instructions: Harmony continuation reuses old instructions and tool definitions instead of the current request's.

i have local fixes for these. the tests pass and through it works using vLLM Agentic API on an NVIDIA GB10, using TensorRT-LLM 1.3.0rc26 with patched serving code, OpenAI SDK 3.8.0, and Pydantic 2.13.4. this tested one frontend with frontend postprocessing, not a full rebuild of main.

the first draft, #19222, covers input conversion/replay and includes regression tests: 86 unit tests and pre-commit checks passed. the live results above cover the combined local fixes, not that PR alone. i'll keep streaming/storage and tool-history/Harmony fixes in separate PRs.

does this split make sense, and are any of these behaviors intentional restrictions of the standalone API?

Start by reviewing draft #19222 and its regression tests for input conversion and replay, then compare standalone /v1/responses behavior with the listed streaming, storage, tool-history, and Harmony cases. The remaining work is planned as separate pull requests; done means each behavior has agreed scope and corresponding regression coverage.