test(live_llm): cover OpenAI null content tool-call replay

Add a live multi-turn tool replay that asserts provider content=null stays
null through normalize + OpenAIHistoryAdapter and that the continuation
still answers. Mark gpt_4/gpt_5 families as supports_tool_replay.

tests/live_llm/README.md

@@ -63,8 +63,8 @@ export OPENROUTER_API_KEY="sk-or-v1-..."
 Coverage by provider:
 
 - Anthropic: structured output path, prompt caching metrics, thinking blocks, multi-turn tool replay
-- OpenAI GPT-4 class: structured outputs, prompt caching
-- OpenAI GPT-5 class (incl. gpt-5.x point-releases): structured outputs, prompt caching, `reasoning_effort`, `max_completion_tokens` routing
+- OpenAI GPT-4 class: structured outputs, prompt caching, multi-turn tool replay (null `content` preserved)
+- OpenAI GPT-5 class (incl. gpt-5.x point-releases): structured outputs, prompt caching, `reasoning_effort`, `max_completion_tokens` routing, multi-turn tool replay (null `content` preserved)
 - OpenAI transport → OpenRouter non-reasoning models (e.g. `inception/mercury-2`): non-chat / diffusion architectures must stay on `max_tokens`, no `reasoning_effort`, tool-calling parameter-schema compatibility is the canary for exotic OR-served providers
 - Gemini 2.5/3.0 classes: structured outputs, cached-content reuse, thought signatures, multi-turn tool replay
 - Gemini 3.1 class: thinking and tool replay coverage by default; structured-output/caching coverage should only be added once Google documents support for that path

tests/live_llm/model_matrix.py

@@ -58,6 +58,7 @@ MODEL_FAMILIES: tuple[LiveModelFamily, ...] = (
         default_models=("gpt-4.1",),
         supports_structured_output=True,
         supports_caching=True,
+        supports_tool_replay=True,
         docs_url="https://platform.openai.com/docs/models/gpt-4.1",
     ),
     LiveModelFamily(
@@ -68,6 +69,7 @@ MODEL_FAMILIES: tuple[LiveModelFamily, ...] = (
         supports_structured_output=True,
         supports_caching=True,
         supports_reasoning=True,
+        supports_tool_replay=True,
         docs_url="https://platform.openai.com/docs/models/gpt-5",
     ),
     # OpenAI-compatible transport → OpenRouter-served non-reasoning models.

tests/live_llm/test_live_openai.py

@@ -2,10 +2,13 @@ from __future__ import annotations
 
 import pytest
 
+from src.llm.history_adapters import OpenAIHistoryAdapter
 from src.llm.request_builder import execute_completion
 
 from .conftest import (
     StructuredLiveResponse,
+    execute_local_tool,
+    favorite_prime_tools,
     make_backend,
     make_large_system_prompt,
     require_provider_key,
@@ -30,6 +33,11 @@ _JSON_OBJECT_SPECS = tuple(
     for spec in get_live_model_specs(provider="openai")
     if spec.family == "openai_json_object"
 )
+_TOOL_REPLAY_SPECS = tuple(
+    spec
+    for spec in get_live_model_specs(provider="openai")
+    if spec.supports_tool_replay
+)
 
 
 @pytest.mark.asyncio
@@ -194,3 +202,83 @@ async def test_live_openai_json_object_structured_output(
     assert parse_calls == []
     assert create_calls, "expected a chat.completions.create call"
     assert create_calls[0]["kwargs"]["response_format"] == {"type": "json_object"}
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_spec", _TOOL_REPLAY_SPECS, ids=lambda spec: spec.id)
+async def test_live_openai_tool_replay_preserves_null_content(
+    model_spec: LiveModelSpec,
+) -> None:
+    """Tool-call turns with provider content=null must stay null through
+    normalize + history replay, and the continuation must still succeed."""
+    require_provider_key(model_spec)
+    # Leave reasoning_effort unset: gpt-5.4 rejects function tools with any
+    # explicit reasoning_effort other than 'none' on /v1/chat/completions.
+    backend, config = make_backend(model_spec)
+    tools = favorite_prime_tools()
+    adapter = OpenAIHistoryAdapter()
+
+    initial_messages = [
+        {
+            "role": "user",
+            "content": (
+                "Before answering, call the get_favorite_prime tool exactly once. "
+                "Do not answer with plain text on this turn. "
+                "After you receive the tool result, answer in one sentence that "
+                "includes the number and the word 'prime'."
+            ),
+        }
+    ]
+
+    first = await execute_completion(
+        backend,
+        config,
+        messages=initial_messages,
+        max_tokens=1024,
+        tools=tools,
+        tool_choice="required",
+    )
+
+    assert first.tool_calls, "OpenAI should issue a tool call in the first turn"
+    raw_message = first.raw_response.choices[0].message
+    raw_content = raw_message.content
+    if raw_content is None:
+        assert first.content is None
+    else:
+        assert first.content == raw_content
+
+    assistant_message = adapter.format_assistant_tool_message(first)
+    assert assistant_message["content"] is (
+        first.content if isinstance(first.content, str) else None
+    )
+    if raw_content is None:
+        assert assistant_message["content"] is None
+
+    tool_call = first.tool_calls[0]
+    tool_result = execute_local_tool(tool_call.name, tool_call.input)
+    replay_messages = initial_messages + [
+        assistant_message,
+        *adapter.format_tool_results(
+            [
+                {
+                    "tool_id": tool_call.id,
+                    "tool_name": tool_call.name,
+                    "result": tool_result,
+                }
+            ]
+        ),
+    ]
+
+    second = await execute_completion(
+        backend,
+        config,
+        messages=replay_messages,
+        max_tokens=1024,
+        tools=tools,
+        tool_choice="auto",
+    )
+
+    assert not second.tool_calls, "continuation should answer without another tool call"
+    assert isinstance(second.content, str)
+    assert "13" in second.content
+    assert "prime" in second.content.lower()