From 1c9dfc1e1a4178d3ed4ef694bb0d273903b4b6e1 Mon Sep 17 00:00:00 2001
From: WH-2099 <wh2099@pm.me>
Date: Wed, 21 Oct 2026 09:22:09 +0000
Subject: [PATCH] feat(model-runtime): expose incremental LLM result
 accumulation

---
 src/graphon/model_runtime/README.md           |  3 ++
 .../base/large_language_model.py              |  6 ++-
 .../test_llm_result_accumulator.py            | 49 +++++++++++++++++++
 3 files changed, 56 insertions(+), 2 deletions(-)
 create mode 100644 tests/model_runtime/test_llm_result_accumulator.py

diff --git a/src/graphon/model_runtime/README.md b/src/graphon/model_runtime/README.md
index 89f752e6..38b96836 100644
--- a/src/graphon/model_runtime/README.md
+++ b/src/graphon/model_runtime/README.md
@@ -39,6 +39,9 @@ blocks retain their order and block state; see the
 [SlimLLM](../dsl/slim/llm.py) shares this aggregation for blocking ordinary and
 structured output calls, covered by the [Slim tests](../../../tests/dsl/test_slim_llm.py).
 
+Hosts can display chunks while calling `consume(chunk)` on one [LLMResultAccumulator](model_providers/base/large_language_model.py) per invocation.
+After consuming the stream, call its `to_result(model=..., prompt_messages=...)` and use the returned `message` for replay or [JSON persistence](../../../tests/model_runtime/test_llm_result_accumulator.py).
+
 [Prompt filtering](../nodes/llm/llm_utils.py) preserves state on supported blocks
 and assistant messages with state but no content; see the
 [prompt tests](../../../tests/nodes/llm/test_llm_utils.py).
diff --git a/src/graphon/model_runtime/model_providers/base/large_language_model.py b/src/graphon/model_runtime/model_providers/base/large_language_model.py
index 9a89ee25..ac2eb602 100644
--- a/src/graphon/model_runtime/model_providers/base/large_language_model.py
+++ b/src/graphon/model_runtime/model_providers/base/large_language_model.py
@@ -118,7 +118,9 @@ def _merge_tool_call_delta(
 
 
 @dataclass(slots=True)
-class _LLMChunkAccumulator:
+class LLMResultAccumulator:
+    """Retain a complete result while a caller consumes and displays each chunk."""
+
     content: str = ""
     content_list: list[PromptMessageContentUnionTypes] = field(default_factory=list)
     usage: LLMUsage = field(default_factory=LLMUsage.empty_usage)
@@ -242,7 +244,7 @@ def _build_llm_result_from_chunks(
         A normalized `LLMResult` assembled from the consumed chunks.
 
     """
-    accumulator = _LLMChunkAccumulator()
+    accumulator = LLMResultAccumulator()
     try:
         accumulator.consume_all(chunks)
     except Exception:
diff --git a/tests/model_runtime/test_llm_result_accumulator.py b/tests/model_runtime/test_llm_result_accumulator.py
new file mode 100644
index 00000000..1e7aff98
--- /dev/null
+++ b/tests/model_runtime/test_llm_result_accumulator.py
@@ -0,0 +1,49 @@
+from graphon.model_runtime.entities.llm_entities import (
+    LLMResultChunk,
+    LLMResultChunkDelta,
+)
+from graphon.model_runtime.entities.message_entities import (
+    AssistantPromptMessage,
+    TextPromptMessageContent,
+)
+from graphon.model_runtime.model_providers.base import large_language_model
+
+
+def test_accumulated_result_preserves_message_through_json() -> None:
+    accumulator = large_language_model.LLMResultAccumulator()
+    tool_call = AssistantPromptMessage.ToolCall(
+        id="call-1",
+        type="function",
+        function=AssistantPromptMessage.ToolCall.ToolCallFunction(
+            name="lookup",
+            arguments='{"id":9007199254740993}',
+        ),
+    )
+    block = TextPromptMessageContent(data="", opaque_body={"signature": "sig"})
+    messages = [
+        AssistantPromptMessage(content="before"),
+        AssistantPromptMessage(
+            content=[block], tool_calls=[tool_call], opaque_body=False
+        ),
+        AssistantPromptMessage(content="after"),
+    ]
+    for index, message in enumerate(messages):
+        accumulator.consume(
+            LLMResultChunk(
+                model="test-model",
+                delta=LLMResultChunkDelta(index=index, message=message),
+            )
+        )
+
+    result = accumulator.to_result(model="test-model", prompt_messages=[])
+    restored_message = AssistantPromptMessage.model_validate_json(
+        result.message.model_dump_json()
+    )
+
+    assert restored_message.content == [
+        TextPromptMessageContent(data="before"),
+        TextPromptMessageContent(data="", opaque_body={"signature": "sig"}),
+        TextPromptMessageContent(data="after"),
+    ]
+    assert restored_message.tool_calls == [tool_call]
+    assert restored_message.opaque_body is False