UPSTREAM PR #19773: server : merge contiguous Responses input items into a single assistant message

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#19773

The Responses API endpoint constructs separate chat completion messages for each input item, except for reasoning. This causes problems with many chat templates that expect content, reasoning, and tool calls to appear in a single assistant message.

This PR merges contiguous assistant inputs into a single message before passing them to templates.

This also preserves reasoning content that isn't coupled with a tool call. A few models, such as Ministral 3, support interleaved reasoning within regular messages. Models that don't typically handle pruning the reasoning in their own templates.

ref: ggml-org/llama.cpp#19765 (comment)
fixes #19513

cc @bfroemel

[loci-review]

No meaningful performance changes were detected across 111693 analyzed functions in the following binaries: build.bin.llama-tts, build.bin.libmtmd.so, build.bin.libllama.so, build.bin.llama-cvector-generator, build.bin.llama-tokenize, build.bin.libggml-base.so, build.bin.libggml-cpu.so, build.bin.libggml.so, build.bin.llama-bench, build.bin.llama-gemma3-cli, build.bin.llama-gguf-split, build.bin.llama-llava-cli, build.bin.llama-minicpmv-cli, build.bin.llama-quantize, build.bin.llama-qwen2vl-cli.

🔎 Full breakdown: Loci Inspector.
💬 Questions? Tag @loci-dev.