RomanNum3ral
/
honcho
mirror of https://github.com/plastic-labs/honcho.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge 93e3f2941a into 2ad56a4d71

This commit is contained in:
Willow Lopez 2026-09-03 09:04:09 -04:00 committed by GitHub
parent 2ad56a4d71 93e3f2941a
commit a7a84eced4
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 171 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
sdks/python/src/honcho/api_types.py
View File

@ -45,6 +45,7 @@ class SummaryConfiguration(BaseModel):
enabled: bool | None = None
messages_per_short_summary: int | None = None
messages_per_long_summary: int | None = None
custom_instructions: str | None = None
class DreamConfiguration(BaseModel):

21
src/schemas/configuration.py
View File

@ -65,6 +65,15 @@ class SummaryConfiguration(BaseModel):
ge=20,
description="Number of messages per long summary. Must be positive, greater than or equal to 20, and greater than messages_per_short_summary.",
)
custom_instructions: str | None = Field(
default=None,
description="Optional custom instructions for session summaries. Validated against DERIVER.MAX_CUSTOM_INSTRUCTIONS_TOKENS.",
)
@field_validator("custom_instructions")
@classmethod
def validate_custom_instructions(cls, value: str | None) -> str | None:
return _validate_custom_instructions_budget(value)
@model_validator(mode="after")
def validate_summary_thresholds(self) -> Self:
@ -130,7 +139,7 @@ class WorkspaceConfiguration(BaseModel):
)
dream: DreamConfiguration | None = Field(
default=None,
description="Configuration for dream functionality. If reasoning is disabled, dreams will also be disabled and these settings will be ignored.",
description="Configuration for dream functionality. If reasoning is disabled, dreams will also be disabled and this setting will be ignored.",
)
@ -155,6 +164,10 @@ class MessageConfiguration(BaseModel):
default=None,
description="Configuration for reasoning functionality.",
)
summary: SummaryConfiguration | None = Field(
default=None,
description="Configuration for summary functionality.",
)
class ResolvedReasoningConfiguration(BaseModel):
@ -176,6 +189,12 @@ class ResolvedSummaryConfiguration(BaseModel):
enabled: bool
messages_per_short_summary: int
messages_per_long_summary: int
custom_instructions: str | None = None
@field_validator("custom_instructions")
@classmethod
def validate_custom_instructions(cls, value: str | None) -> str | None:
return _validate_custom_instructions_budget(value)
class ResolvedDreamConfiguration(BaseModel):

25
src/utils/config_helpers.py
View File

@ -12,18 +12,34 @@ from src.schemas import (
logger = logging.getLogger(__name__)
_NONE_OVERRIDE_PATHS: set[tuple[str, ...]] = {
("summary", "custom_instructions"),
}
def deep_update(base: dict[str, Any], update: dict[str, Any]) -> None:
def deep_update(
base: dict[str, Any],
update: dict[str, Any],
path: tuple[str, ...] = (),
) -> None:
"""
Recursive update of a dictionary.
Skips None values in the update dictionary.
Skips None values unless None explicitly clears a nullable field.
"""
for key, value in update.items():
current_path = (*path, key)
if value is None:
if current_path in _NONE_OVERRIDE_PATHS:
base[key] = None
continue
if isinstance(value, dict) and key in base and isinstance(base[key], dict):
deep_update(cast(dict[str, Any], base[key]), cast(dict[str, Any], value))
deep_update(
cast(dict[str, Any], base[key]),
cast(dict[str, Any], value),
current_path,
)
else:
base[key] = value
@ -113,6 +129,7 @@ def get_configuration(
"enabled": settings.SUMMARY.ENABLED,
"messages_per_short_summary": settings.SUMMARY.MESSAGES_PER_SHORT_SUMMARY,
"messages_per_long_summary": settings.SUMMARY.MESSAGES_PER_LONG_SUMMARY,
"custom_instructions": None,
},
"dream": {"enabled": settings.DREAM.ENABLED},
}
@ -130,7 +147,7 @@ def get_configuration(
deep_update(
config_dict,
normalize_configuration_dict(
message_configuration.model_dump(exclude_none=True)
message_configuration.model_dump(exclude_unset=True)
),
)

91
src/utils/summarizer.py
View File

@ -2,7 +2,7 @@ import asyncio
import logging
import time
from enum import Enum
from functools import cache
from functools import cache, lru_cache
from inspect import cleandoc as c
from typing import TypedDict
@ -15,6 +15,8 @@ from src.cache.client import cache as cache_client
from src.config import ConfiguredModelSettings, settings
from src.crud.session import session_cache_key
from src.dependencies import tracked_db
# TODO: move _custom_instructions_section to shared utility
from src.deriver.prompts import _custom_instructions_section
from src.exceptions import ResourceNotFoundException
from src.llm import HonchoLLMCallResponse, honcho_llm_call
from src.llm.types import LLMTelemetryContext
@ -58,6 +60,7 @@ class Summary(TypedDict):
def to_schema_summary(s: Summary) -> schemas.Summary:
"""Convert a Summary TypedDict to a Pydantic Summary schema object."""
return schemas.Summary(
content=s["content"],
message_id=s["message_id"],
@ -82,6 +85,7 @@ __all__ = [
def _get_summary_model_config() -> ConfiguredModelSettings:
"""Return the configured model settings for summary generation."""
return settings.SUMMARY.MODEL_CONFIG
@ -102,8 +106,10 @@ def short_summary_prompt(
formatted_messages: str,
output_words: int,
previous_summary_text: str,
custom_instructions: str | None = None,
) -> str:
"""Generate the short summary prompt."""
custom_instructions_section = _custom_instructions_section(custom_instructions)
return c(f"""
You are a system that summarizes parts of a conversation to create a concise and accurate summary. Focus on capturing:
@ -118,6 +124,7 @@ Provide a concise, factual summary that captures the essence of the conversation
Return only the summary without any explanation or meta-commentary.
{custom_instructions_section}
<previous_summary>
{previous_summary_text}
</previous_summary>
@ -134,8 +141,10 @@ def long_summary_prompt(
formatted_messages: str,
output_words: int,
previous_summary_text: str,
custom_instructions: str | None = None,
) -> str:
"""Generate the long summary prompt."""
custom_instructions_section = _custom_instructions_section(custom_instructions)
return c(f"""
You are a system that creates thorough, comprehensive summaries of conversations. Focus on capturing:
@ -152,6 +161,7 @@ Provide a thorough and detailed summary that captures the essence of the convers
Return only the summary without any explanation or meta-commentary.
{custom_instructions_section}
<previous_summary>
{previous_summary_text}
</previous_summary>
@ -164,46 +174,65 @@ Hard limit: {output_words} words maximum. If needed, drop lower-priority detail
""")
@cache
def estimate_short_summary_prompt_tokens() -> int:
"""Estimate tokens for the short summary prompt (without messages/previous_summary)."""
@lru_cache(maxsize=128)
def estimate_short_summary_prompt_tokens(
custom_instructions: str | None = None,
) -> int:
"""Estimate tokens for the short summary prompt, optionally including custom instructions."""
try:
return estimate_tokens(
short_summary_prompt(
formatted_messages="",
output_words=0,
previous_summary_text="",
custom_instructions=custom_instructions,
)
)
except Exception:
# Return a rough estimate if estimation fails
return 200
@cache
def estimate_long_summary_prompt_tokens() -> int:
"""Estimate tokens for the long summary prompt (without messages/previous_summary)."""
@lru_cache(maxsize=128)
def estimate_long_summary_prompt_tokens(
custom_instructions: str | None = None,
) -> int:
"""Estimate tokens for the long summary prompt, optionally including custom instructions."""
try:
return estimate_tokens(
long_summary_prompt(
formatted_messages="",
output_words=0,
previous_summary_text="",
custom_instructions=custom_instructions,
)
)
except Exception:
# Return a rough estimate if estimation fails
return 200
@conditional_observe(name="Create Short Summary")
async def create_short_summary(
formatted_messages: str,
input_tokens: int,
previous_summary: str | None = None,
custom_instructions: str | None = None,
*,
workspace_name: str | None = None,
) -> HonchoLLMCallResponse[str]:
"""
Generate a short summary via an LLM call.
Args:
formatted_messages: Pre-formatted conversation messages.
input_tokens: Token count of the input (messages + previous summary).
previous_summary: Previous summary text for continuity, if any.
custom_instructions: Optional custom instructions from configuration.
workspace_name: Workspace name for telemetry attribution.
Returns:
The LLM response containing the short summary text and token counts.
"""
# input_tokens indicates how many tokens the message list + previous summary take up
# we want to optimize short summaries to be smaller than the actual content being summarized
# so we ask the agent to produce a word count roughly equal to either the input, or the max
@ -217,7 +246,10 @@ async def create_short_summary(
previous_summary_text = "There is no previous summary -- the messages are the beginning of the conversation."
prompt = short_summary_prompt(
formatted_messages, output_words, previous_summary_text
formatted_messages,
output_words,
previous_summary_text,
custom_instructions=custom_instructions,
)
# Mint a root span id.
@ -242,9 +274,22 @@ async def create_short_summary(
async def create_long_summary(
formatted_messages: str,
previous_summary: str | None = None,
custom_instructions: str | None = None,
*,
workspace_name: str | None = None,
) -> HonchoLLMCallResponse[str]:
"""
Generate a comprehensive long summary via an LLM call.
Args:
formatted_messages: Pre-formatted conversation messages.
previous_summary: Previous summary text for continuity, if any.
custom_instructions: Optional custom instructions from configuration.
workspace_name: Workspace name for telemetry attribution.
Returns:
The LLM response containing the long summary text and token counts.
"""
# the word/token ratio is roughly 4:3 so we multiply by 0.75.
# LLMs *seem* to respond better to getting asked for a word count but should workshop this.
output_words = int(settings.SUMMARY.MAX_TOKENS_LONG * 0.75)
@ -255,7 +300,10 @@ async def create_long_summary(
previous_summary_text = "There is no previous summary -- the messages are the beginning of the conversation."
prompt = long_summary_prompt(
formatted_messages, output_words, previous_summary_text
formatted_messages,
output_words,
previous_summary_text,
custom_instructions=custom_instructions,
)
# Mint a root span id.
@ -452,6 +500,13 @@ async def _create_and_save_summary(
previous_summary_tokens = latest_summary["token_count"] if latest_summary else 0
input_tokens = messages_tokens + previous_summary_tokens
# Extract custom_instructions from the summarizer's own configuration.
# This is separate from reasoning custom_instructions — workspace
# operators may want summaries in a different style than deriver output.
custom_instructions: str | None = None
if configuration.summary and configuration.summary.custom_instructions is not None:
custom_instructions = configuration.summary.custom_instructions
(
new_summary,
is_fallback,
@ -466,6 +521,7 @@ async def _create_and_save_summary(
last_message_id=last_message_id,
last_message_content_preview=last_message_content_preview,
message_count=message_count,
custom_instructions=custom_instructions,
workspace_name=workspace_name,
)
@ -473,9 +529,13 @@ async def _create_and_save_summary(
# save-summary path and the telemetry emit below can use it
# without basedpyright tripping on a possibly-unbound name.
if summary_type == SummaryType.SHORT:
prompt_tokens = estimate_short_summary_prompt_tokens()
prompt_tokens = estimate_short_summary_prompt_tokens(
custom_instructions
)
else:
prompt_tokens = estimate_long_summary_prompt_tokens()
prompt_tokens = estimate_long_summary_prompt_tokens(
custom_instructions
)
# Step 3: Save to database with new transaction
if not is_fallback:
@ -561,6 +621,7 @@ async def _create_summary(
last_message_id: int,
last_message_content_preview: str,
message_count: int,
custom_instructions: str | None = None,
*,
workspace_name: str | None = None,
) -> tuple[Summary, bool, int, int]:
@ -576,6 +637,8 @@ async def _create_summary(
last_message_id: ID of the last message
last_message_content_preview: Preview of last message content for fallback
message_count: Number of messages for fallback
custom_instructions: Optional workspace-level custom instructions for prompt
workspace_name: Optional workspace name for telemetry
Returns:
A tuple of (Summary, is_fallback, llm_input_tokens, llm_output_tokens)
@ -594,12 +657,14 @@ async def _create_summary(
formatted_messages,
input_tokens,
previous_summary_text,
custom_instructions=custom_instructions,
workspace_name=workspace_name,
)
else:
response = await create_long_summary(
formatted_messages,
previous_summary_text,
custom_instructions=custom_instructions,
workspace_name=workspace_name,
)

51
tests/test_config_helpers.py Normal file
View File

@ -0,0 +1,51 @@
from types import SimpleNamespace
from typing import Any
from src.schemas import MessageConfiguration, SummaryConfiguration
from src.utils.config_helpers import deep_update, get_configuration
def _configured_node(configuration: dict[str, Any]) -> Any:
return SimpleNamespace(configuration=configuration)
class TestDeepUpdate:
def test_summary_custom_instructions_none_clears_inherited_value(self) -> None:
base = {
"summary": {
"enabled": True,
"custom_instructions": "Write summaries in German.",
}
}
deep_update(
base,
{"summary": {"enabled": None, "custom_instructions": None}},
)
assert base["summary"]["enabled"] is True
assert base["summary"]["custom_instructions"] is None
class TestGetConfiguration:
def test_session_can_clear_workspace_summary_custom_instructions(self) -> None:
workspace = _configured_node(
{"summary": {"custom_instructions": "Write summaries in German."}}
)
session = _configured_node({"summary": {"custom_instructions": None}})
configuration = get_configuration(None, session, workspace)
assert configuration.summary.custom_instructions is None
def test_message_can_clear_session_summary_custom_instructions(self) -> None:
session = _configured_node(
{"summary": {"custom_instructions": "Write summaries in German."}}
)
message_configuration = MessageConfiguration(
summary=SummaryConfiguration(custom_instructions=None)
)
configuration = get_configuration(message_configuration, session)
assert configuration.summary.custom_instructions is None