OpenAI multi-turn tool history reaches chat templates with string-valued arguments

Row: SERVE-TOOL-HISTORY-ARGS

Bug

The OpenAI Chat Completions protocol carries assistant.tool_calls[].function.arguments as a JSON-encoded string. vllm.cpp preserves that string through ChatMessage and BuildMessages() and hands it directly to Jinja. This diverges from pinned vLLM vllm/entrypoints/chat_utils.py::_postprocess_messages, which json.loads() each historical argument string before chat-template rendering (empty/null becomes {}).

Gemma 4 exposes the failure because its canonical DSL requires an arguments mapping. A valid first call such as {"command":"date"} is rendered into history as <|tool_call>call:terminal{"command":"date"}<tool_call|> rather than canonical <|tool_call>call:terminal{command:<|"|>date<|"|>}<tool_call|>. On the next turn Gemma imitates the quoted form; the parser faithfully returns a literal quote-wrapped key such as {"\"command\"":"\"ls\""}, and clients cannot find command. Retries recursively escape it.

Reproduction evidence

• ~/llms/logs/agentic_tool_loop_20260811-212603.json: turns 1-2 have valid arguments; turn 3 is the first completed-history turn and emits the literal quote-wrapped key/value.
• The same pattern repeats in agentic_tool_loop_20260811-202653.json.
• Fresh-context controls emit valid {command: ...}-derived JSON.
• Current vllm.cpp src/vllm/entrypoints/chat_template.cpp stores fn["arguments"] = tc.function.arguments.
• Pinned vLLM converts strings to objects in vllm/entrypoints/chat_utils.py:1915-1951.
• Google/vLLM Gemma 4 canonical templates now fail closed unless historical arguments are mappings.

Expected behavior

Before rendering any chat template, vllm.cpp should mirror vLLM message post-processing:

1. Decode valid JSON argument strings to JSON values.
2. Normalize empty/null arguments to {}.
3. Reject invalid JSON instead of teaching the model malformed history.
4. Preserve already-structured arguments internally.
5. Gate a complete user → assistant tool call → tool result → next-generation Gemma render, plus non-Gemma template compatibility.

This must be fixed in vllm.cpp, not in Hermes or another client.

Start in src/vllm/entrypoints/chat_template.cpp, where fn["arguments"] is currently stored, and compare the pinned vLLM handling in vllm/entrypoints/chat_utils.py:1915-1951. Trace BuildMessages() through chat-template rendering, then add coverage for a complete tool-call history with Gemma 4 and a non-Gemma template, including valid, empty/null, structured, and invalid arguments.