github/AutoGPT
1
1
Fork 0
mirror of https://github.com/Significant-Gravitas/AutoGPT.git synced 2026-02-10 06:45:28 -05:00
Code Issues Packages Projects Releases Wiki Activity

fix(backend/chat): Remove manual context truncation from SDK path, let SDK handle compaction

Browse Source
This commit is contained in:
Zamil Majdy
2026-02-10 13:52:49 +04:00
parent 587e11c60a
commit 43153a12e0
2 changed files with 11 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
autogpt_platform/backend/backend/api/features/chat/config.py
View File

@@ -27,13 +27,13 @@ class ChatConfig(BaseSettings):
session_ttl: int = Field(default=43200, description="Session TTL in seconds")
# Streaming Configuration
# Note: When using Claude Agent SDK, context management is handled automatically
# via the SDK's built-in compaction. This is mainly used for the fallback path.
# Only used by the non-SDK fallback path (Anthropic direct).
# The SDK path passes full history and relies on SDK's built-in compaction.
max_context_messages: int = Field(
default=100,
ge=1,
le=500,
description="Max context messages (SDK handles compaction automatically)",
description="Max context messages for non-SDK fallback path only",
)
stream_timeout: int = Field(default=300, description="Stream timeout in seconds")

22
autogpt_platform/backend/backend/api/features/chat/sdk/service.py
View File

@@ -142,8 +142,8 @@ async def _build_system_prompt(
def _format_conversation_history(session: ChatSession) -> str:
"""Format conversation history as a prompt context.
The SDK handles context compaction automatically, but we apply
max_context_messages as a safety guard to limit initial prompt size.
Passes full history to the SDK — the SDK handles context compaction
automatically when the context window approaches its limit.
"""
if not session.messages:
return ""
@@ -153,19 +153,12 @@ def _format_conversation_history(session: ChatSession) -> str:
if not messages:
return ""
# Apply max_context_messages limit as a safety guard
# (SDK handles compaction, but this prevents excessively large initial prompts)
max_messages = config.max_context_messages
if len(messages) > max_messages:
messages = messages[-max_messages:]
history_parts = ["<conversation_history>"]
for msg in messages:
if msg.role == "user":
history_parts.append(f"User: {msg.content or ''}")
elif msg.role == "assistant":
# Pass full content - SDK handles compaction automatically
history_parts.append(f"Assistant: {msg.content or ''}")
if msg.tool_calls:
for tc in msg.tool_calls:
@@ -174,11 +167,7 @@ def _format_conversation_history(session: ChatSession) -> str:
f" [Called tool: {func.get('name', 'unknown')}]"
)
elif msg.role == "tool":
# Truncate large tool results to avoid blowing context window
tool_content = msg.content or ""
if len(tool_content) > 500:
tool_content = tool_content[:500] + "... (truncated)"
history_parts.append(f" [Tool result: {tool_content}]")
history_parts.append(f" [Tool result: {msg.content or ''}]")
history_parts.append("</conversation_history>")
history_parts.append("")
@@ -333,6 +322,11 @@ async def stream_chat_completion_sdk(
else:
prompt = current_message
logger.info(
f"[SDK] Prompt built: {len(prompt)} chars, "
f"{len(session.messages)} messages in session"
)
# Guard against empty prompts
if not prompt.strip():
yield StreamError(