apunkt/SurfSense
1
0
Fork 0
mirror of https://github.com/MODSetter/SurfSense.git synced 2026-07-14 22:52:15 +02:00
Code Issues Projects Releases Packages Wiki Activity

feat: moved most things behind correct feature flag

Browse source
This commit is contained in:
DESKTOP-RTLN3BA\$punk 2026-05-02 23:10:48 -07:00
parent bdb97a0888
commit c938d39277
13 changed files with 237 additions and 85 deletions
Download patch file Download diff file Expand all files Collapse all files

106
surfsense_backend/app/agents/new_chat/feature_flags.py
View file

@ -3,8 +3,10 @@ Feature flags for the SurfSense new_chat agent stack.
These flags gate the newer agent middleware (some ported from OpenCode,
some sourced from ``langchain.agents.middleware`` / ``deepagents``, some
SurfSense-native). They follow a "default-OFF for risky things,
default-ON for safe upgrades, master kill-switch for everything new" model.
SurfSense-native). Most shipped agent-stack upgrades default ON so Docker
image updates work even when older installs do not have newly introduced
environment variables. Risky/experimental integrations stay default OFF,
and the master kill-switch can still disable everything new.
All new middleware checks its flag at agent build time. If the master
kill-switch ``SURFSENSE_DISABLE_NEW_AGENT_STACK`` is set, every new
@ -14,16 +16,19 @@ operators a single switch to revert to pre-port behavior.
Examples
--------
Local development (recommended for trying everything except doom-loop / selector):
Defaults:
SURFSENSE_ENABLE_CONTEXT_EDITING=true
SURFSENSE_ENABLE_COMPACTION_V2=true
SURFSENSE_ENABLE_RETRY_AFTER=true
SURFSENSE_ENABLE_MODEL_FALLBACK=false
SURFSENSE_ENABLE_MODEL_CALL_LIMIT=true
SURFSENSE_ENABLE_TOOL_CALL_LIMIT=true
SURFSENSE_ENABLE_TOOL_CALL_REPAIR=true
SURFSENSE_ENABLE_PERMISSION=false # default off, opt-in per deploy
SURFSENSE_ENABLE_DOOM_LOOP=false # default off until UI ships
SURFSENSE_ENABLE_LLM_TOOL_SELECTOR=false
SURFSENSE_ENABLE_STREAM_PARITY_V2=false # structured streaming events
SURFSENSE_ENABLE_PERMISSION=true
SURFSENSE_ENABLE_DOOM_LOOP=true
SURFSENSE_ENABLE_LLM_TOOL_SELECTOR=false # adds a per-turn LLM call
SURFSENSE_ENABLE_STREAM_PARITY_V2=true
Master kill-switch (overrides everything else):
@ -60,32 +65,28 @@ class AgentFeatureFlags:
disable_new_agent_stack: bool = False
# Agent quality — context budget, retry/limits, name-repair, doom-loop
enable_context_editing: bool = False
enable_compaction_v2: bool = False
enable_retry_after: bool = False
enable_context_editing: bool = True
enable_compaction_v2: bool = True
enable_retry_after: bool = True
enable_model_fallback: bool = False
enable_model_call_limit: bool = False
enable_tool_call_limit: bool = False
enable_tool_call_repair: bool = False
enable_doom_loop: bool = (
False # Default OFF until UI handles permission='doom_loop'
)
enable_model_call_limit: bool = True
enable_tool_call_limit: bool = True
enable_tool_call_repair: bool = True
enable_doom_loop: bool = True
# Safety — permissions, concurrency, tool-set narrowing
enable_permission: bool = False # Default OFF for first deploy
enable_busy_mutex: bool = False
enable_permission: bool = True
enable_busy_mutex: bool = True
enable_llm_tool_selector: bool = False # Default OFF — adds per-turn LLM cost
# Skills + subagents
enable_skills: bool = False
enable_specialized_subagents: bool = False
enable_kb_planner_runnable: bool = False
enable_skills: bool = True
enable_specialized_subagents: bool = True
enable_kb_planner_runnable: bool = True
# Snapshot / revert
enable_action_log: bool = False
enable_revert_route: bool = (
False # Backend ships before UI; route returns 503 until this flips
)
enable_action_log: bool = True
enable_revert_route: bool = True
# Streaming parity v2 — opt in to LangChain's structured
# ``AIMessageChunk`` content (typed reasoning blocks, tool-input
@ -94,7 +95,7 @@ class AgentFeatureFlags:
# text path and the synthetic ``call_<run_id>`` tool-call id (no
# ``langchainToolCallId`` propagation). Schema migrations 135/136
# ship unconditionally because they're forward-compatible.
enable_stream_parity_v2: bool = False
enable_stream_parity_v2: bool = True
# Plugins
enable_plugin_loader: bool = False
@ -115,43 +116,64 @@ class AgentFeatureFlags:
"SURFSENSE_DISABLE_NEW_AGENT_STACK is set: every new agent "
"middleware is forced OFF for this build."
)
return cls(disable_new_agent_stack=True)
return cls(
disable_new_agent_stack=True,
enable_context_editing=False,
enable_compaction_v2=False,
enable_retry_after=False,
enable_model_fallback=False,
enable_model_call_limit=False,
enable_tool_call_limit=False,
enable_tool_call_repair=False,
enable_doom_loop=False,
enable_permission=False,
enable_busy_mutex=False,
enable_llm_tool_selector=False,
enable_skills=False,
enable_specialized_subagents=False,
enable_kb_planner_runnable=False,
enable_action_log=False,
enable_revert_route=False,
enable_stream_parity_v2=False,
enable_plugin_loader=False,
enable_otel=False,
)
return cls(
disable_new_agent_stack=False,
# Agent quality
enable_context_editing=_env_bool("SURFSENSE_ENABLE_CONTEXT_EDITING", False),
enable_compaction_v2=_env_bool("SURFSENSE_ENABLE_COMPACTION_V2", False),
enable_retry_after=_env_bool("SURFSENSE_ENABLE_RETRY_AFTER", False),
enable_context_editing=_env_bool("SURFSENSE_ENABLE_CONTEXT_EDITING", True),
enable_compaction_v2=_env_bool("SURFSENSE_ENABLE_COMPACTION_V2", True),
enable_retry_after=_env_bool("SURFSENSE_ENABLE_RETRY_AFTER", True),
enable_model_fallback=_env_bool("SURFSENSE_ENABLE_MODEL_FALLBACK", False),
enable_model_call_limit=_env_bool(
"SURFSENSE_ENABLE_MODEL_CALL_LIMIT", False
"SURFSENSE_ENABLE_MODEL_CALL_LIMIT", True
),
enable_tool_call_limit=_env_bool("SURFSENSE_ENABLE_TOOL_CALL_LIMIT", False),
enable_tool_call_limit=_env_bool("SURFSENSE_ENABLE_TOOL_CALL_LIMIT", True),
enable_tool_call_repair=_env_bool(
"SURFSENSE_ENABLE_TOOL_CALL_REPAIR", False
"SURFSENSE_ENABLE_TOOL_CALL_REPAIR", True
),
enable_doom_loop=_env_bool("SURFSENSE_ENABLE_DOOM_LOOP", False),
enable_doom_loop=_env_bool("SURFSENSE_ENABLE_DOOM_LOOP", True),
# Safety
enable_permission=_env_bool("SURFSENSE_ENABLE_PERMISSION", False),
enable_busy_mutex=_env_bool("SURFSENSE_ENABLE_BUSY_MUTEX", False),
enable_permission=_env_bool("SURFSENSE_ENABLE_PERMISSION", True),
enable_busy_mutex=_env_bool("SURFSENSE_ENABLE_BUSY_MUTEX", True),
enable_llm_tool_selector=_env_bool(
"SURFSENSE_ENABLE_LLM_TOOL_SELECTOR", False
),
# Skills + subagents
enable_skills=_env_bool("SURFSENSE_ENABLE_SKILLS", False),
enable_skills=_env_bool("SURFSENSE_ENABLE_SKILLS", True),
enable_specialized_subagents=_env_bool(
"SURFSENSE_ENABLE_SPECIALIZED_SUBAGENTS", False
"SURFSENSE_ENABLE_SPECIALIZED_SUBAGENTS", True
),
enable_kb_planner_runnable=_env_bool(
"SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE", False
"SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE", True
),
# Snapshot / revert
enable_action_log=_env_bool("SURFSENSE_ENABLE_ACTION_LOG", False),
enable_revert_route=_env_bool("SURFSENSE_ENABLE_REVERT_ROUTE", False),
enable_action_log=_env_bool("SURFSENSE_ENABLE_ACTION_LOG", True),
enable_revert_route=_env_bool("SURFSENSE_ENABLE_REVERT_ROUTE", True),
# Streaming parity v2
enable_stream_parity_v2=_env_bool(
"SURFSENSE_ENABLE_STREAM_PARITY_V2", False
"SURFSENSE_ENABLE_STREAM_PARITY_V2", True
),
# Plugins
enable_plugin_loader=_env_bool("SURFSENSE_ENABLE_PLUGIN_LOADER", False),

8
surfsense_backend/app/routes/agent_flags_route.py
View file

@ -23,6 +23,7 @@ from fastapi import APIRouter, Depends
from pydantic import BaseModel
from app.agents.new_chat.feature_flags import AgentFeatureFlags, get_flags
from app.config import config
from app.db import User
from app.users import current_active_user
@ -58,10 +59,15 @@ class AgentFeatureFlagsRead(BaseModel):
enable_otel: bool
enable_desktop_local_filesystem: bool
@classmethod
def from_flags(cls, flags: AgentFeatureFlags) -> AgentFeatureFlagsRead:
# asdict() avoids missing-field bugs when AgentFeatureFlags grows.
return cls(**asdict(flags))
return cls(
**asdict(flags),
enable_desktop_local_filesystem=config.ENABLE_DESKTOP_LOCAL_FILESYSTEM,
)
@router.get("/agent/flags", response_model=AgentFeatureFlagsRead)

2
surfsense_backend/app/services/auto_model_pin_service.py
View file

@ -399,7 +399,7 @@ async def resolve_or_get_pinned_llm_config_id(
False if force_repin_free else await _is_premium_eligible(session, user_id)
)
if premium_eligible:
eligible = candidates
eligible = [c for c in candidates if _tier_of(c) == "premium"]
else:
eligible = [c for c in candidates if _tier_of(c) != "premium"]

38
surfsense_backend/tests/unit/agents/new_chat/test_feature_flags.py
View file

@ -31,18 +31,38 @@ def _clear_all(monkeypatch: pytest.MonkeyPatch) -> None:
"SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE",
"SURFSENSE_ENABLE_ACTION_LOG",
"SURFSENSE_ENABLE_REVERT_ROUTE",
"SURFSENSE_ENABLE_STREAM_PARITY_V2",
"SURFSENSE_ENABLE_PLUGIN_LOADER",
"SURFSENSE_ENABLE_OTEL",
]:
monkeypatch.delenv(name, raising=False)
def test_defaults_all_off(monkeypatch: pytest.MonkeyPatch) -> None:
def test_defaults_match_shipped_agent_stack(monkeypatch: pytest.MonkeyPatch) -> None:
_clear_all(monkeypatch)
flags = reload_for_tests()
assert isinstance(flags, AgentFeatureFlags)
assert flags.disable_new_agent_stack is False
assert flags.any_new_middleware_enabled() is False
assert flags.enable_context_editing is True
assert flags.enable_compaction_v2 is True
assert flags.enable_retry_after is True
assert flags.enable_model_fallback is False
assert flags.enable_model_call_limit is True
assert flags.enable_tool_call_limit is True
assert flags.enable_tool_call_repair is True
assert flags.enable_doom_loop is True
assert flags.enable_permission is True
assert flags.enable_busy_mutex is True
assert flags.enable_llm_tool_selector is False
assert flags.enable_skills is True
assert flags.enable_specialized_subagents is True
assert flags.enable_kb_planner_runnable is True
assert flags.enable_action_log is True
assert flags.enable_revert_route is True
assert flags.enable_stream_parity_v2 is True
assert flags.enable_plugin_loader is False
assert flags.enable_otel is False
assert flags.any_new_middleware_enabled() is True
def test_master_kill_switch_overrides_individual_flags(
@ -100,21 +120,13 @@ def test_each_flag_can_be_set_independently(monkeypatch: pytest.MonkeyPatch) ->
"enable_kb_planner_runnable": "SURFSENSE_ENABLE_KB_PLANNER_RUNNABLE",
"enable_action_log": "SURFSENSE_ENABLE_ACTION_LOG",
"enable_revert_route": "SURFSENSE_ENABLE_REVERT_ROUTE",
"enable_stream_parity_v2": "SURFSENSE_ENABLE_STREAM_PARITY_V2",
"enable_plugin_loader": "SURFSENSE_ENABLE_PLUGIN_LOADER",
"enable_otel": "SURFSENSE_ENABLE_OTEL",
}
# `enable_otel` is intentionally orthogonal — it does NOT count toward
# ``any_new_middleware_enabled`` because OTel is observability-only and
# ships under its own ``OTEL_EXPORTER_OTLP_ENDPOINT`` requirement.
counts_toward_middleware = {k for k in flag_to_env if k != "enable_otel"}
for attr, env_name in flag_to_env.items():
_clear_all(monkeypatch)
monkeypatch.setenv(env_name, "true")
monkeypatch.setenv(env_name, "false")
flags = reload_for_tests()
assert getattr(flags, attr) is True, f"{attr} did not flip on for {env_name}"
if attr in counts_toward_middleware:
assert flags.any_new_middleware_enabled() is True
else:
assert flags.any_new_middleware_enabled() is False
assert getattr(flags, attr) is False, f"{attr} did not flip off for {env_name}"

55
surfsense_backend/tests/unit/services/test_auto_model_pin_service.py
View file

@ -101,11 +101,58 @@ async def test_auto_first_turn_pins_one_model(monkeypatch):
user_id="00000000-0000-0000-0000-000000000001",
selected_llm_config_id=0,
)
assert result.resolved_llm_config_id in {-1, -2}
assert result.resolved_llm_config_id == -1
assert session.thread.pinned_llm_config_id == result.resolved_llm_config_id
assert session.commit_count == 1
@pytest.mark.asyncio
async def test_premium_eligible_auto_prefers_premium_over_free(monkeypatch):
from app.config import config
session = _FakeSession(_thread())
monkeypatch.setattr(
config,
"GLOBAL_LLM_CONFIGS",
[
{
"id": -2,
"provider": "OPENAI",
"model_name": "gpt-free",
"api_key": "k1",
"billing_tier": "free",
"quality_score": 100,
},
{
"id": -1,
"provider": "OPENAI",
"model_name": "gpt-prem",
"api_key": "k2",
"billing_tier": "premium",
"quality_score": 10,
},
],
)
async def _allowed(*_args, **_kwargs):
return _FakeQuotaResult(allowed=True)
monkeypatch.setattr(
"app.services.auto_model_pin_service.TokenQuotaService.premium_get_usage",
_allowed,
)
result = await resolve_or_get_pinned_llm_config_id(
session,
thread_id=1,
search_space_id=10,
user_id="00000000-0000-0000-0000-000000000001",
selected_llm_config_id=0,
)
assert result.resolved_llm_config_id == -1
assert result.resolved_tier == "premium"
@pytest.mark.asyncio
async def test_next_turn_reuses_existing_pin(monkeypatch):
from app.config import config
@ -361,12 +408,12 @@ async def test_invalid_pinned_config_repairs_with_new_pin(monkeypatch):
],
)
async def _allowed(*_args, **_kwargs):
return _FakeQuotaResult(allowed=True)
async def _blocked(*_args, **_kwargs):
return _FakeQuotaResult(allowed=False)
monkeypatch.setattr(
"app.services.auto_model_pin_service.TokenQuotaService.premium_get_usage",
_allowed,
_blocked,
)
result = await resolve_or_get_pinned_llm_config_id(