Merge remote-tracking branch 'origin/dev' into lluisagusti/secrt-1926-agent-generation-completes-but-ui-does-not-update-refresh

This commit is contained in:
Lluis Agusti
2026-02-12 18:53:12 +08:00
6 changed files with 336 additions and 33 deletions

View File

@@ -2,7 +2,7 @@ import asyncio
import logging
import uuid
from datetime import UTC, datetime
from typing import Any
from typing import Any, cast
from weakref import WeakValueDictionary
from openai.types.chat import (
@@ -104,6 +104,26 @@ class ChatSession(BaseModel):
successful_agent_runs: dict[str, int] = {}
successful_agent_schedules: dict[str, int] = {}
def add_tool_call_to_current_turn(self, tool_call: dict) -> None:
"""Attach a tool_call to the current turn's assistant message.
Searches backwards for the most recent assistant message (stopping at
any user message boundary). If found, appends the tool_call to it.
Otherwise creates a new assistant message with the tool_call.
"""
for msg in reversed(self.messages):
if msg.role == "user":
break
if msg.role == "assistant":
if not msg.tool_calls:
msg.tool_calls = []
msg.tool_calls.append(tool_call)
return
self.messages.append(
ChatMessage(role="assistant", content="", tool_calls=[tool_call])
)
@staticmethod
def new(user_id: str) -> "ChatSession":
return ChatSession(
@@ -172,6 +192,47 @@ class ChatSession(BaseModel):
successful_agent_schedules=successful_agent_schedules,
)
@staticmethod
def _merge_consecutive_assistant_messages(
messages: list[ChatCompletionMessageParam],
) -> list[ChatCompletionMessageParam]:
"""Merge consecutive assistant messages into single messages.
Long-running tool flows can create split assistant messages: one with
text content and another with tool_calls. Anthropic's API requires
tool_result blocks to reference a tool_use in the immediately preceding
assistant message, so these splits cause 400 errors via OpenRouter.
"""
if len(messages) < 2:
return messages
result: list[ChatCompletionMessageParam] = [messages[0]]
for msg in messages[1:]:
prev = result[-1]
if prev.get("role") != "assistant" or msg.get("role") != "assistant":
result.append(msg)
continue
prev = cast(ChatCompletionAssistantMessageParam, prev)
curr = cast(ChatCompletionAssistantMessageParam, msg)
curr_content = curr.get("content") or ""
if curr_content:
prev_content = prev.get("content") or ""
prev["content"] = (
f"{prev_content}\n{curr_content}" if prev_content else curr_content
)
curr_tool_calls = curr.get("tool_calls")
if curr_tool_calls:
prev_tool_calls = prev.get("tool_calls")
prev["tool_calls"] = (
list(prev_tool_calls) + list(curr_tool_calls)
if prev_tool_calls
else list(curr_tool_calls)
)
return result
def to_openai_messages(self) -> list[ChatCompletionMessageParam]:
messages = []
for message in self.messages:
@@ -258,7 +319,7 @@ class ChatSession(BaseModel):
name=message.name or "",
)
)
return messages
return self._merge_consecutive_assistant_messages(messages)
async def _get_session_from_cache(session_id: str) -> ChatSession | None:

View File

@@ -1,4 +1,16 @@
from typing import cast
import pytest
from openai.types.chat import (
ChatCompletionAssistantMessageParam,
ChatCompletionMessageParam,
ChatCompletionToolMessageParam,
ChatCompletionUserMessageParam,
)
from openai.types.chat.chat_completion_message_tool_call_param import (
ChatCompletionMessageToolCallParam,
Function,
)
from .model import (
ChatMessage,
@@ -117,3 +129,205 @@ async def test_chatsession_db_storage(setup_test_user, test_user_id):
loaded.tool_calls is not None
), f"Tool calls missing for {orig.role} message"
assert len(orig.tool_calls) == len(loaded.tool_calls)
# --------------------------------------------------------------------------- #
# _merge_consecutive_assistant_messages #
# --------------------------------------------------------------------------- #
_tc = ChatCompletionMessageToolCallParam(
id="tc1", type="function", function=Function(name="do_stuff", arguments="{}")
)
_tc2 = ChatCompletionMessageToolCallParam(
id="tc2", type="function", function=Function(name="other", arguments="{}")
)
def test_merge_noop_when_no_consecutive_assistants():
"""Messages without consecutive assistants are returned unchanged."""
msgs = [
ChatCompletionUserMessageParam(role="user", content="hi"),
ChatCompletionAssistantMessageParam(role="assistant", content="hello"),
ChatCompletionUserMessageParam(role="user", content="bye"),
]
merged = ChatSession._merge_consecutive_assistant_messages(msgs)
assert len(merged) == 3
assert [m["role"] for m in merged] == ["user", "assistant", "user"]
def test_merge_splits_text_and_tool_calls():
"""The exact bug scenario: text-only assistant followed by tool_calls-only assistant."""
msgs = [
ChatCompletionUserMessageParam(role="user", content="build agent"),
ChatCompletionAssistantMessageParam(
role="assistant", content="Let me build that"
),
ChatCompletionAssistantMessageParam(
role="assistant", content="", tool_calls=[_tc]
),
ChatCompletionToolMessageParam(role="tool", content="ok", tool_call_id="tc1"),
]
merged = ChatSession._merge_consecutive_assistant_messages(msgs)
assert len(merged) == 3
assert merged[0]["role"] == "user"
assert merged[2]["role"] == "tool"
a = cast(ChatCompletionAssistantMessageParam, merged[1])
assert a["role"] == "assistant"
assert a.get("content") == "Let me build that"
assert a.get("tool_calls") == [_tc]
def test_merge_combines_tool_calls_from_both():
"""Both consecutive assistants have tool_calls — they get merged."""
msgs: list[ChatCompletionAssistantMessageParam] = [
ChatCompletionAssistantMessageParam(
role="assistant", content="text", tool_calls=[_tc]
),
ChatCompletionAssistantMessageParam(
role="assistant", content="", tool_calls=[_tc2]
),
]
merged = ChatSession._merge_consecutive_assistant_messages(msgs) # type: ignore[arg-type]
assert len(merged) == 1
a = cast(ChatCompletionAssistantMessageParam, merged[0])
assert a.get("tool_calls") == [_tc, _tc2]
assert a.get("content") == "text"
def test_merge_three_consecutive_assistants():
"""Three consecutive assistants collapse into one."""
msgs: list[ChatCompletionAssistantMessageParam] = [
ChatCompletionAssistantMessageParam(role="assistant", content="a"),
ChatCompletionAssistantMessageParam(role="assistant", content="b"),
ChatCompletionAssistantMessageParam(
role="assistant", content="", tool_calls=[_tc]
),
]
merged = ChatSession._merge_consecutive_assistant_messages(msgs) # type: ignore[arg-type]
assert len(merged) == 1
a = cast(ChatCompletionAssistantMessageParam, merged[0])
assert a.get("content") == "a\nb"
assert a.get("tool_calls") == [_tc]
def test_merge_empty_and_single_message():
"""Edge cases: empty list and single message."""
assert ChatSession._merge_consecutive_assistant_messages([]) == []
single: list[ChatCompletionMessageParam] = [
ChatCompletionUserMessageParam(role="user", content="hi")
]
assert ChatSession._merge_consecutive_assistant_messages(single) == single
# --------------------------------------------------------------------------- #
# add_tool_call_to_current_turn #
# --------------------------------------------------------------------------- #
_raw_tc = {
"id": "tc1",
"type": "function",
"function": {"name": "f", "arguments": "{}"},
}
_raw_tc2 = {
"id": "tc2",
"type": "function",
"function": {"name": "g", "arguments": "{}"},
}
def test_add_tool_call_appends_to_existing_assistant():
"""When the last assistant is from the current turn, tool_call is added to it."""
session = ChatSession.new(user_id="u")
session.messages = [
ChatMessage(role="user", content="hi"),
ChatMessage(role="assistant", content="working on it"),
]
session.add_tool_call_to_current_turn(_raw_tc)
assert len(session.messages) == 2 # no new message created
assert session.messages[1].tool_calls == [_raw_tc]
def test_add_tool_call_creates_assistant_when_none_exists():
"""When there's no current-turn assistant, a new one is created."""
session = ChatSession.new(user_id="u")
session.messages = [
ChatMessage(role="user", content="hi"),
]
session.add_tool_call_to_current_turn(_raw_tc)
assert len(session.messages) == 2
assert session.messages[1].role == "assistant"
assert session.messages[1].tool_calls == [_raw_tc]
def test_add_tool_call_does_not_cross_user_boundary():
"""A user message acts as a boundary — previous assistant is not modified."""
session = ChatSession.new(user_id="u")
session.messages = [
ChatMessage(role="assistant", content="old turn"),
ChatMessage(role="user", content="new message"),
]
session.add_tool_call_to_current_turn(_raw_tc)
assert len(session.messages) == 3 # new assistant was created
assert session.messages[0].tool_calls is None # old assistant untouched
assert session.messages[2].role == "assistant"
assert session.messages[2].tool_calls == [_raw_tc]
def test_add_tool_call_multiple_times():
"""Multiple long-running tool calls accumulate on the same assistant."""
session = ChatSession.new(user_id="u")
session.messages = [
ChatMessage(role="user", content="hi"),
ChatMessage(role="assistant", content="doing stuff"),
]
session.add_tool_call_to_current_turn(_raw_tc)
# Simulate a pending tool result in between (like _yield_tool_call does)
session.messages.append(
ChatMessage(role="tool", content="pending", tool_call_id="tc1")
)
session.add_tool_call_to_current_turn(_raw_tc2)
assert len(session.messages) == 3 # user, assistant, tool — no extra assistant
assert session.messages[1].tool_calls == [_raw_tc, _raw_tc2]
def test_to_openai_messages_merges_split_assistants():
"""End-to-end: session with split assistants produces valid OpenAI messages."""
session = ChatSession.new(user_id="u")
session.messages = [
ChatMessage(role="user", content="build agent"),
ChatMessage(role="assistant", content="Let me build that"),
ChatMessage(
role="assistant",
content="",
tool_calls=[
{
"id": "tc1",
"type": "function",
"function": {"name": "create_agent", "arguments": "{}"},
}
],
),
ChatMessage(role="tool", content="done", tool_call_id="tc1"),
ChatMessage(role="assistant", content="Saved!"),
ChatMessage(role="user", content="show me an example run"),
]
openai_msgs = session.to_openai_messages()
# The two consecutive assistants at index 1,2 should be merged
roles = [m["role"] for m in openai_msgs]
assert roles == ["user", "assistant", "tool", "assistant", "user"]
# The merged assistant should have both content and tool_calls
merged = cast(ChatCompletionAssistantMessageParam, openai_msgs[1])
assert merged.get("content") == "Let me build that"
tc_list = merged.get("tool_calls")
assert tc_list is not None and len(list(tc_list)) == 1
assert list(tc_list)[0]["id"] == "tc1"

View File

@@ -800,9 +800,13 @@ async def stream_chat_completion(
# Build the messages list in the correct order
messages_to_save: list[ChatMessage] = []
# Add assistant message with tool_calls if any
# Add assistant message with tool_calls if any.
# Use extend (not assign) to preserve tool_calls already added by
# _yield_tool_call for long-running tools.
if accumulated_tool_calls:
assistant_response.tool_calls = accumulated_tool_calls
if not assistant_response.tool_calls:
assistant_response.tool_calls = []
assistant_response.tool_calls.extend(accumulated_tool_calls)
logger.info(
f"Added {len(accumulated_tool_calls)} tool calls to assistant message"
)
@@ -1404,13 +1408,9 @@ async def _yield_tool_call(
operation_id=operation_id,
)
# Save assistant message with tool_call FIRST (required by LLM)
assistant_message = ChatMessage(
role="assistant",
content="",
tool_calls=[tool_calls[yield_idx]],
)
session.messages.append(assistant_message)
# Attach the tool_call to the current turn's assistant message
# (or create one if this is a tool-only response with no text).
session.add_tool_call_to_current_turn(tool_calls[yield_idx])
# Then save pending tool result
pending_message = ChatMessage(

View File

@@ -21,43 +21,71 @@ logger = logging.getLogger(__name__)
class HumanInTheLoopBlock(Block):
"""
This block pauses execution and waits for human approval or modification of the data.
Pauses execution and waits for human approval or rejection of the data.
When executed, it creates a pending review entry and sets the node execution status
to REVIEW. The execution will remain paused until a human user either:
- Approves the data (with or without modifications)
- Rejects the data
When executed, this block creates a pending review entry and sets the node execution
status to REVIEW. The execution remains paused until a human user either approves
or rejects the data.
This is useful for workflows that require human validation or intervention before
proceeding to the next steps.
**How it works:**
- The input data is presented to a human reviewer
- The reviewer can approve or reject (and optionally modify the data if editable)
- On approval: the data flows out through the `approved_data` output pin
- On rejection: the data flows out through the `rejected_data` output pin
**Important:** The output pins yield the actual data itself, NOT status strings.
The approval/rejection decision determines WHICH output pin fires, not the value.
You do NOT need to compare the output to "APPROVED" or "REJECTED" - simply connect
downstream blocks to the appropriate output pin for each case.
**Example usage:**
- Connect `approved_data` → next step in your workflow (data was approved)
- Connect `rejected_data` → error handling or notification (data was rejected)
"""
class Input(BlockSchemaInput):
data: Any = SchemaField(description="The data to be reviewed by a human user")
data: Any = SchemaField(
description="The data to be reviewed by a human user. "
"This exact data will be passed through to either approved_data or "
"rejected_data output based on the reviewer's decision."
)
name: str = SchemaField(
description="A descriptive name for what this data represents",
description="A descriptive name for what this data represents. "
"This helps the reviewer understand what they are reviewing.",
)
editable: bool = SchemaField(
description="Whether the human reviewer can edit the data",
description="Whether the human reviewer can edit the data before "
"approving or rejecting it",
default=True,
advanced=True,
)
class Output(BlockSchemaOutput):
approved_data: Any = SchemaField(
description="The data when approved (may be modified by reviewer)"
description="Outputs the input data when the reviewer APPROVES it. "
"The value is the actual data itself (not a status string like 'APPROVED'). "
"If the reviewer edited the data, this contains the modified version. "
"Connect downstream blocks here for the 'approved' workflow path."
)
rejected_data: Any = SchemaField(
description="The data when rejected (may be modified by reviewer)"
description="Outputs the input data when the reviewer REJECTS it. "
"The value is the actual data itself (not a status string like 'REJECTED'). "
"If the reviewer edited the data, this contains the modified version. "
"Connect downstream blocks here for the 'rejected' workflow path."
)
review_message: str = SchemaField(
description="Any message provided by the reviewer", default=""
description="Optional message provided by the reviewer explaining their "
"decision. Only outputs when the reviewer provides a message; "
"this pin does not fire if no message was given.",
default="",
)
def __init__(self):
super().__init__(
id="8b2a7b3c-6e9d-4a5f-8c1b-2e3f4a5b6c7d",
description="Pause execution and wait for human approval or modification of data",
description="Pause execution for human review. Data flows through "
"approved_data or rejected_data output based on the reviewer's decision. "
"Outputs contain the actual data, not status strings.",
categories={BlockCategory.BASIC},
input_schema=HumanInTheLoopBlock.Input,
output_schema=HumanInTheLoopBlock.Output,

View File

@@ -61,7 +61,7 @@ Below is a comprehensive list of all available blocks, categorized by their prim
| [Get List Item](block-integrations/basic.md#get-list-item) | Returns the element at the given index |
| [Get Store Agent Details](block-integrations/system/store_operations.md#get-store-agent-details) | Get detailed information about an agent from the store |
| [Get Weather Information](block-integrations/basic.md#get-weather-information) | Retrieves weather information for a specified location using OpenWeatherMap API |
| [Human In The Loop](block-integrations/basic.md#human-in-the-loop) | Pause execution and wait for human approval or modification of data |
| [Human In The Loop](block-integrations/basic.md#human-in-the-loop) | Pause execution for human review |
| [List Is Empty](block-integrations/basic.md#list-is-empty) | Checks if a list is empty |
| [List Library Agents](block-integrations/system/library_operations.md#list-library-agents) | List all agents in your personal library |
| [Note](block-integrations/basic.md#note) | A visual annotation block that displays a sticky note in the workflow editor for documentation and organization purposes |

View File

@@ -975,7 +975,7 @@ A travel planning application could use this block to provide users with current
## Human In The Loop
### What it is
Pause execution and wait for human approval or modification of data
Pause execution for human review. Data flows through approved_data or rejected_data output based on the reviewer's decision. Outputs contain the actual data, not status strings.
### How it works
<!-- MANUAL: how_it_works -->
@@ -988,18 +988,18 @@ This enables human oversight at critical points in automated workflows, ensuring
| Input | Description | Type | Required |
|-------|-------------|------|----------|
| data | The data to be reviewed by a human user | Data | Yes |
| name | A descriptive name for what this data represents | str | Yes |
| editable | Whether the human reviewer can edit the data | bool | No |
| data | The data to be reviewed by a human user. This exact data will be passed through to either approved_data or rejected_data output based on the reviewer's decision. | Data | Yes |
| name | A descriptive name for what this data represents. This helps the reviewer understand what they are reviewing. | str | Yes |
| editable | Whether the human reviewer can edit the data before approving or rejecting it | bool | No |
### Outputs
| Output | Description | Type |
|--------|-------------|------|
| error | Error message if the operation failed | str |
| approved_data | The data when approved (may be modified by reviewer) | Approved Data |
| rejected_data | The data when rejected (may be modified by reviewer) | Rejected Data |
| review_message | Any message provided by the reviewer | str |
| approved_data | Outputs the input data when the reviewer APPROVES it. The value is the actual data itself (not a status string like 'APPROVED'). If the reviewer edited the data, this contains the modified version. Connect downstream blocks here for the 'approved' workflow path. | Approved Data |
| rejected_data | Outputs the input data when the reviewer REJECTS it. The value is the actual data itself (not a status string like 'REJECTED'). If the reviewer edited the data, this contains the modified version. Connect downstream blocks here for the 'rejected' workflow path. | Rejected Data |
| review_message | Optional message provided by the reviewer explaining their decision. Only outputs when the reviewer provides a message; this pin does not fire if no message was given. | str |
### Possible use case
<!-- MANUAL: use_case -->