Skip to content

Building AI agents, atomically

fix: include new input in the context token budget

liangzhechen655a0bb9b862b86
diff --git a/atomic-agents/atomic_agents/agents/atomic_agent.py b/atomic-agents/atomic_agents/agents/atomic_agent.py
index ab656e7..ebcd289 100644
--- a/atomic-agents/atomic_agents/agents/atomic_agent.py
+++ b/atomic-agents/atomic_agents/agents/atomic_agent.py
@@ -297,7 +297,8 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
Called before building the messages list. Uses the full context token count
(system prompt + history + tools) via get_context_token_count().
- Removes oldest turns one at a time until the context fits within the limit.
+ Removes oldest turns one at a time until the context fits within the limit,
+ preserving the current turn so the request never loses its latest input.
Turn-preserving: always removes complete turns, never individual messages.
@@ -327,6 +328,8 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
for turn_id in turn_ids_ordered:
if total_tokens <= self.max_context_tokens:
break
+ if turn_id == self.history.current_turn_id:
+ continue
self.history.delete_turn_id(turn_id)
new_result = self.get_context_token_count()
@@ -577,14 +580,12 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
self.client, instructor.core.client.AsyncInstructor
), "The run method is not supported for async clients. Use run_async instead."
- # Trim history BEFORE adding new user message to protect the new input
- self._trim_context()
-
if user_input:
self.history.initialize_turn()
self.current_user_input = user_input
self.history.add_message("user", user_input)
+ self._trim_context()
self._prepare_messages()
response = self.client.chat.completions.create(
messages=self.messages,
@@ -614,13 +615,12 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
self.client, instructor.core.client.AsyncInstructor
), "The run_stream method is not supported for async clients. Use run_async instead."
- self._trim_context()
-
if user_input:
self.history.initialize_turn()
self.current_user_input = user_input
self.history.add_message("user", user_input)
+ self._trim_context()
self._prepare_messages()
response_stream = self.client.chat.completions.create_partial(
@@ -658,13 +658,12 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
"""
assert isinstance(self.client, instructor.core.client.AsyncInstructor), "The run_async method is for async clients."
- self._trim_context()
-
if user_input:
self.history.initialize_turn()
self.current_user_input = user_input
self.history.add_message("user", user_input)
+ self._trim_context()
self._prepare_messages()
response = await self.client.chat.completions.create(
@@ -686,12 +685,12 @@ class AtomicAgent[InputSchema: BaseIOSchema, OutputSchema: BaseIOSchema]:
OutputSchema: Partial responses from the chat agent.
"""
assert isinstance(self.client, instructor.core.client.AsyncInstructor), "The run_async method is for async clients."
- self._trim_context()
if user_input:
self.history.initialize_turn()
self.current_user_input = user_input
self.history.add_message("user", user_input)
+ self._trim_context()
self._prepare_messages()
response_stream = self.client.chat.completions.create_partial(
diff --git a/atomic-agents/tests/agents/test_atomic_agent.py b/atomic-agents/tests/agents/test_atomic_agent.py
index 9bfdd47..5dfcc3e 100644
--- a/atomic-agents/tests/agents/test_atomic_agent.py
+++ b/atomic-agents/tests/agents/test_atomic_agent.py
@@ -1,5 +1,7 @@
+import json
+
import pytest
-from unittest.mock import Mock, call, patch
+from unittest.mock import AsyncMock, Mock, call, patch
from enum import Enum
from pydantic import BaseModel, Field
from pydantic import ValidationError
@@ -1271,6 +1273,67 @@ def test_prepare_messages_keeps_system_for_openai(mock_instructor, mock_system_p
# --- max_context_tokens and _trim_context tests ---
[email protected]("method", ["run", "run_stream", "run_async", "run_async_stream"])
[email protected]("input_size", [15, 40])
+@patch("atomic_agents.agents.atomic_agent.get_token_counter")
+async def test_context_budget_includes_new_input(
+ mock_get_token_counter, method, input_size, mock_instructor, mock_instructor_async, mock_system_prompt_generator
+):
+ """All completion APIs budget the actual request and preserve the latest input."""
+ history = ChatHistory()
+ history.add_message("user", BasicChatInputSchema(chat_message="o" * 20))
+ client = mock_instructor_async if "async" in method else mock_instructor
+ response = BasicChatOutputSchema(chat_message="Response")
+ if "async" in method:
+ client.chat.completions.create = AsyncMock(return_value=response)
+
+ async def partials(**kwargs):
+ yield response
+
+ client.chat.completions.create_partial = Mock(side_effect=partials)
+
+ def count_context(model, system_messages, history_messages, tools):
+ history_tokens = sum(len(json.loads(message["content"])["chat_message"]) for message in history_messages)
+ return TokenCountResult(total=10 + history_tokens, system_prompt=10, history=history_tokens, tools=0, model=model)
+
+ mock_get_token_counter.return_value.count_context.side_effect = count_context
+ agent = AtomicAgent[BasicChatInputSchema, BasicChatOutputSchema](
+ AgentConfig(
+ client=client,
+ model="gpt-5-mini",
+ history=history,
+ system_prompt_generator=mock_system_prompt_generator,
+ max_context_tokens=30,
+ )
+ )
+ user_input = BasicChatInputSchema(chat_message="n" * input_size)
+
+ async def invoke():
+ if method == "run":
+ agent.run(user_input)
+ elif method == "run_stream":
+ list(agent.run_stream(user_input))
+ elif method == "run_async":
+ await agent.run_async(user_input)
+ else:
+ async for _ in agent.run_async_stream(user_input):
+ pass
+
+ if input_size == 40:
+ with pytest.raises(ValueError, match="max_context_tokens"):
+ await invoke()
+ client.chat.completions.create.assert_not_called()
+ client.chat.completions.create_partial.assert_not_called()
+ assert history.history[-1].content == user_input
+ else:
+ await invoke()
+ completion = client.chat.completions.create_partial if "stream" in method else client.chat.completions.create
+ request_history = completion.call_args.kwargs["messages"][1:]
+ assert request_history == [{"role": "user", "content": user_input.model_dump_json()}]
+ assert count_context(agent.model, [], request_history, None).total <= agent.max_context_tokens
+
+
def test_trim_context_no_op_when_max_context_tokens_unset(mock_instructor, mock_system_prompt_generator):
"""When max_context_tokens is None, _trim_context returns without any action."""
history = ChatHistory()
@@ -1391,8 +1454,8 @@ def test_trim_context_raises_when_single_turn_exceeds_limit(
@patch("atomic_agents.agents.atomic_agent.get_token_counter")
-def test_trim_context_called_before_user_message_in_run(mock_get_token_counter, mock_instructor, mock_system_prompt_generator):
- """_trim_context runs before the new user message is added to history."""
+def test_trim_context_preserves_new_user_message_in_run(mock_get_token_counter, mock_instructor, mock_system_prompt_generator):
+ """Old turns can be trimmed while the new user input is preserved."""
history = ChatHistory()
# Existing turn in history
@@ -1424,7 +1487,7 @@ def test_trim_context_called_before_user_message_in_run(mock_get_token_counter,
agent.run(BasicChatInputSchema(chat_message="New message"))
- # Old turn was trimmed BEFORE new message was added
+ # The old turn was trimmed and the new message was preserved.
assert history.get_message_count() == 2 # assistant response + new user message
assert history.history[0].content.chat_message == "New message" # new message is first