[kodurd]

Motivation

Responses replies are returned with store=False, so continuing a conversation means sending the previous output items back in input. One assistant turn is reported as several output items — its text first, then its calls — and replaying that turn does not reconstruct it as one turn.

messages_from_input starts a new assistant message for every function_call item unless it directly follows another call, so a turn that contained text and a tool call comes back as two consecutive assistant messages:

# replaying the endpoint's own output: message item + function_call item
[{'role': 'user', 'content': 'weather?'},
 {'role': 'assistant', 'content': 'Let me check. '},            # <- item 1
 {'role': 'assistant', 'content': None, 'tool_calls': [...]},   # <- item 2
 {'role': 'tool', 'tool_call_id': 'call_abc', 'content': 'sunny'}]

/v1/messages and /v1/chat/completions represent that turn as a single assistant message carrying tool_calls.

Chat templates render one turn per message, so the prompt the model receives is wrong in a different way for each template. Rendered through MultimodalProcessor._get_text_prompt_input, same conversation, before this change:

deepseek-v4  …<|Assistant|></think>Let me check. <|end of sentence|>\n\n<|DSML|tool_calls>…
mistral      [INST] weather? [/INST]Let me check. </s></s>sunny
llama2       …weather? [/INST] Let me check. </s><s>[INST]  [/INST] </s><s>[INST] sunny [/INST]

The tool call arrives after an end-of-turn marker, in a turn that was never opened; mistral gets a doubled </s>; llama2 gains an empty user exchange. With this change all three templates produce exactly the prompt the equivalent /v1/messages history produces.

Merged #5044 fixed the neighbouring case, where two adjacent function_call items became two turns instead of one parallel invocation. A call that follows the assistant's own text is still split off.

Modification

• Join a function_call to the preceding assistant message whenever the previous message is one, instead of only when the previous item was itself a call. The assistant message is the turn being continued, so this covers the adjacent-call case too and #5044's previous_item_type bookkeeping is no longer needed; its comment about keeping every id visible to templates is kept.
• A function_call that does not follow an assistant message still starts its own content: None assistant message. That is the shape test_responses_maps_function_call_history_to_chat_messages asserts, and it is unchanged.
• Two regression tests. The text-then-call case fails on main. The parallel-call case passes on main since fix(serve): preserve tool-result image ownership #5044 and is kept as a guard, so a future change to this branch cannot regroup those calls.

Not touched: /v1/messages and /v1/chat/completions already produce the single-message shape.

BC-breaking (Optional)

None at the protocol level. A replayed history that previously rendered as two assistant turns now renders as one, which is what the other two endpoints already did.

Use cases (Optional)

client = OpenAI(base_url='http://127.0.0.1:23333/v1', api_key='x')

first = client.responses.create(model=MODEL, input='weather in Beijing?', tools=TOOLS)
# the model answers with text and a tool call -> two output items

second = client.responses.create(
    model=MODEL,
    input=[*first.output, {'type': 'function_call_output', 'call_id': call_id, 'output': 'sunny'}],
    tools=TOOLS,
)

Before this change the prompt behind second carries an end-of-turn marker between the assistant text and its tool call.

Checklist

• pytest tests/test_lmdeploy/serve — 666 passed against 664 on main, plus the 2 new tests
• The same 3 failures appear with and without this change, on main as well: test_tool_images.py::test_internvl_tool_image_placeholders_across_protocols[internvl3-*]. tests/test_lmdeploy/conftest.py reads internvl3.jinja with Path.read_text() and no encoding, and that template contains non-ASCII text, so it raises UnicodeDecodeError on a Windows machine whose ANSI codepage is not UTF-8. Unrelated to this change and invisible on Linux CI.
• pre-commit run --all-files — all hooks pass
• pylint lmdeploy/serve/openai/responses/request.py — same message set and exit code as main; only line numbers move
• interrogate --fail-under 70 lmdeploy — passes at 71.7%
• No GPU available here, so the engine-level suites and the end-to-end tests were not run locally

[kodurd]

Updated the branch to current main. The diff is unchanged, and tests/test_lmdeploy/serve/openai/responses/ is 152 passed locally on the new base.

@RunningLeon @silencelamb could one of you approve the workflow run here and have a look? CI has never run on this PR, while #5058 got its runs authorized and came back green on all three, so a run is the only thing missing on this one.

The behaviour is still on main: a function_call item that follows assistant text starts a second assistant turn instead of joining the turn it was emitted in, so a model that returns text and then calls a tool gets its own history back split in two.

[kodurd]

unit_test failed on the dependency fetch, not on the change: the runner could not clone fmtlib/fmt (fatal: could not read Username for 'https://github.com', three attempts), so the editable build never finished and no test ran. The job was over in 49 seconds. lint and pr_functions_test were green on the same commit, and unit_test passed on another branch six minutes later, so it looks like a transient fetch problem on the runner.

Updated the branch to current main, which also picks up #5074. Could you re-approve the run?