feat: HomeFormatter parses inline entity tags instead of tool_end blocks
The supervisor LLM now embeds <type>[id1,id2]</type> entity tags in its response text. The HomeFormatter buffers streamed tokens, detects complete tags across chunk boundaries, and emits WsStreamBlock with entity type + specific IDs. This replaces the old approach of emitting blocks for every tool_end event, which dumped ALL entities regardless of relevance. Also fixes: - NoneType guard on metadata in _run_graph_stream (metadata can be None) - Updated _HOME_SYSTEM prompt with entity tag instructions - Updated all affected tests
This commit is contained in:
@@ -236,7 +236,19 @@ _HOME_SYSTEM = (
|
||||
"multiple sub-agents if needed.\n\n"
|
||||
"You also have an update_core_memory tool — use it when the user states "
|
||||
"a preference or important fact worth remembering long-term.\n\n"
|
||||
"After gathering data, synthesize a clear, helpful response for the user.\n\n"
|
||||
"## Entity References\n"
|
||||
"When your response mentions specific workspace entities, embed them "
|
||||
"inline using entity tags so the UI can render interactive components.\n"
|
||||
"Format: <type>[comma-separated UUIDs]</type>\n"
|
||||
"Supported types: task, project, note, timeline\n\n"
|
||||
"Example response:\n"
|
||||
" Here is your project:\n"
|
||||
" <project>[abc-123-def]</project>\n"
|
||||
" It has these pending tasks:\n"
|
||||
" <task>[def-456,ghi-789]</task>\n\n"
|
||||
"IMPORTANT: Only include IDs of entities that are directly relevant to "
|
||||
"the user's question. Do NOT dump all entity IDs returned by a tool — "
|
||||
"filter to only the ones the user asked about or that matter for the answer.\n\n"
|
||||
"Memory context:\n{memory_context}"
|
||||
)
|
||||
|
||||
@@ -360,6 +372,7 @@ async def _run_graph_stream(
|
||||
isinstance(msg, AIMessageChunk)
|
||||
and msg.content
|
||||
and not msg.tool_calls
|
||||
and isinstance(metadata, dict)
|
||||
and metadata.get("langgraph_node") == "agent"
|
||||
):
|
||||
yield ("token", str(msg.content))
|
||||
|
||||
@@ -6,9 +6,10 @@ Consumes ``(event_type, data)`` tuples yielded by ``deep_agent.run_*_stream()``:
|
||||
* ``("mutations", list)`` — collected CRUD mutations for ``stream_end``
|
||||
|
||||
HomeFormatter:
|
||||
* Sniffs ``tool_end`` events → emits ``WsStreamBlock`` (entity_ref with raw data)
|
||||
* Streams text tokens → emits ``WsStreamText``
|
||||
* Attaches mutations → injects into ``WsStreamEnd``
|
||||
* Buffers text tokens and parses inline entity tags
|
||||
``<type>[id1,id2]</type>`` → emits ``WsStreamBlock`` (entity_ref with IDs)
|
||||
* Streams surrounding text → emits ``WsStreamText``
|
||||
* Attaches mutations → injects into ``WsStreamEnd``
|
||||
|
||||
FloatingFormatter:
|
||||
* Sniffs first ``tool_end`` name → emits ``WsFloatingDomain``
|
||||
@@ -19,6 +20,7 @@ FloatingFormatter:
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
import re
|
||||
from collections.abc import AsyncGenerator
|
||||
from typing import Any
|
||||
|
||||
@@ -40,20 +42,91 @@ _AGENT_DOMAIN: dict[str, str] = {
|
||||
"project_agent": "projects",
|
||||
}
|
||||
|
||||
# Regex for complete inline entity tags: <task>[id1,id2]</task>
|
||||
_ENTITY_TAG_RE = re.compile(
|
||||
r"<(task|project|note|timeline)>\[([^\]]+)\]</\1>"
|
||||
)
|
||||
|
||||
# Tag name → plural entity type for the WsStreamBlock data
|
||||
_TAG_ENTITY: dict[str, str] = {
|
||||
"task": "tasks",
|
||||
"project": "projects",
|
||||
"note": "notes",
|
||||
"timeline": "timelines",
|
||||
}
|
||||
|
||||
WsFrame = WsStreamStart | WsStreamText | WsStreamBlock | WsStreamEnd | WsFloatingDomain
|
||||
|
||||
|
||||
class HomeFormatter:
|
||||
"""Consumes a deep-agent event stream and yields WS frames for the Home view.
|
||||
|
||||
``tool_end`` events from sub-agents are emitted as ``WsStreamBlock``
|
||||
(entity_ref) so the client can render structured data. Text tokens are
|
||||
forwarded as ``WsStreamText``. Mutations are attached to ``WsStreamEnd``.
|
||||
The supervisor's response contains inline entity tags like
|
||||
``<project>[abc-123]</project>``. This formatter detects them,
|
||||
emits ``WsStreamBlock(block_type="entity_ref")`` with the entity
|
||||
type and IDs, and forwards surrounding text as ``WsStreamText``.
|
||||
Mutations are attached to ``WsStreamEnd``.
|
||||
"""
|
||||
|
||||
def __init__(self, request_id: str) -> None:
|
||||
self.request_id = request_id
|
||||
self._mutations: list[dict] = []
|
||||
self._buffer: str = ""
|
||||
|
||||
def _flush_buffer(self, force: bool = False):
|
||||
"""Extract complete entity tags and text from the buffer.
|
||||
|
||||
Yields (frame_type, data) pairs:
|
||||
("text", str) — plain text to send as WsStreamText
|
||||
("block", dict) — entity_ref block to send as WsStreamBlock
|
||||
|
||||
When *force* is True (end of stream), the entire buffer is flushed.
|
||||
Otherwise, text after the last unmatched ``<`` is held back in case
|
||||
it is the start of an entity tag arriving across token boundaries.
|
||||
"""
|
||||
buf = self._buffer
|
||||
|
||||
while True:
|
||||
m = _ENTITY_TAG_RE.search(buf)
|
||||
if not m:
|
||||
break
|
||||
|
||||
# Text before the tag
|
||||
before = buf[: m.start()]
|
||||
if before:
|
||||
yield ("text", before)
|
||||
|
||||
# The entity tag itself → a block
|
||||
tag_type = m.group(1)
|
||||
raw_ids = m.group(2)
|
||||
ids = [i.strip() for i in raw_ids.split(",") if i.strip()]
|
||||
yield (
|
||||
"block",
|
||||
{
|
||||
"entity": _TAG_ENTITY[tag_type],
|
||||
"ids": ids,
|
||||
},
|
||||
)
|
||||
|
||||
buf = buf[m.end() :]
|
||||
|
||||
if force:
|
||||
# End of stream — flush everything that remains
|
||||
if buf:
|
||||
yield ("text", buf)
|
||||
self._buffer = ""
|
||||
else:
|
||||
# Keep a potential partial tag (text after last '<') in the buffer
|
||||
last_lt = buf.rfind("<")
|
||||
if last_lt != -1:
|
||||
safe = buf[:last_lt]
|
||||
if safe:
|
||||
yield ("text", safe)
|
||||
self._buffer = buf[last_lt:]
|
||||
else:
|
||||
if buf:
|
||||
yield ("text", buf)
|
||||
self._buffer = ""
|
||||
|
||||
async def format(
|
||||
self,
|
||||
@@ -64,22 +137,36 @@ class HomeFormatter:
|
||||
async for event_type, data in event_stream:
|
||||
if event_type == "token":
|
||||
if data:
|
||||
yield WsStreamText(request_id=self.request_id, chunk=data)
|
||||
|
||||
elif event_type == "tool_end":
|
||||
# Sub-agent finished — emit its result as an entity_ref block
|
||||
name = data.get("name", "")
|
||||
entity = _AGENT_DOMAIN.get(name)
|
||||
if entity:
|
||||
yield WsStreamBlock(
|
||||
request_id=self.request_id,
|
||||
block_type="entity_ref",
|
||||
data={"entity": entity, "result": data.get("result", "")},
|
||||
)
|
||||
self._buffer += data
|
||||
for ftype, fdata in self._flush_buffer():
|
||||
if ftype == "text":
|
||||
yield WsStreamText(
|
||||
request_id=self.request_id, chunk=fdata
|
||||
)
|
||||
elif ftype == "block":
|
||||
yield WsStreamBlock(
|
||||
request_id=self.request_id,
|
||||
block_type="entity_ref",
|
||||
data=fdata,
|
||||
)
|
||||
|
||||
elif event_type == "mutations":
|
||||
self._mutations = data or []
|
||||
|
||||
# tool_end events are intentionally ignored — the supervisor
|
||||
# embeds relevant entity IDs inline via <type>[ids]</type> tags.
|
||||
|
||||
# Flush any remaining buffer content
|
||||
for ftype, fdata in self._flush_buffer(force=True):
|
||||
if ftype == "text":
|
||||
yield WsStreamText(request_id=self.request_id, chunk=fdata)
|
||||
elif ftype == "block":
|
||||
yield WsStreamBlock(
|
||||
request_id=self.request_id,
|
||||
block_type="entity_ref",
|
||||
data=fdata,
|
||||
)
|
||||
|
||||
yield WsStreamEnd(
|
||||
request_id=self.request_id,
|
||||
mutations=[
|
||||
|
||||
Reference in New Issue
Block a user