Skip to content

Self-improving AI Employees for everyone

refactor(agent): centralize runtime defaults in llm_defaults.json

trohitg9c26d92f2d71
diff --git a/server/config/llm_defaults.json b/server/config/llm_defaults.json
index ebc41516..76a76dea 100644
--- a/server/config/llm_defaults.json
+++ b/server/config/llm_defaults.json
@@ -3,6 +3,15 @@
"last_updated": "2026-04-03",
"_description": "LLM provider defaults. Connection config (base_url, models_endpoint) drives both native SDK and LangChain paths. Model specs auto-enriched by model_registry.json (OpenRouter cache).",
+ "agent": {
+ "_description": "Agent execution defaults shared across the LangGraph + compaction layers.",
+ "recursion_limit": 50,
+ "default_temperature": 0.7,
+ "compaction": {
+ "ratio": 0.5
+ }
+ },
+
"providers": {
"openai": {
"default_model": "gpt-5.4",
diff --git a/server/nodes/agent/_specialized.py b/server/nodes/agent/_specialized.py
index fc8fe663..e499fff5 100644
--- a/server/nodes/agent/_specialized.py
+++ b/server/nodes/agent/_specialized.py
@@ -48,12 +48,17 @@ class SpecializedAgentParams(BaseModel):
)
# ---- "Options" group (collapsed by default in the parameter panel) ----
- temperature: float = Field(
- default=0.7, ge=0.0, le=2.0,
+ # default=None so an unset value falls through to ``agent.default_temperature``
+ # in server/config/llm_defaults.json (resolved by _resolve_temperature).
+ temperature: Optional[float] = Field(
+ default=None, ge=0.0, le=2.0,
json_schema_extra={"group": "options"},
)
+ # default=None so an unset value is absent from the dumped dict and the
+ # backend (_resolve_max_tokens) falls through to the per-model default
+ # in server/config/llm_defaults.json instead of being silently capped.
max_tokens: Optional[int] = Field(
- default=1000, ge=1, le=200000,
+ default=None, ge=1, le=200000,
json_schema_extra={"group": "options"},
)
diff --git a/server/nodes/agent/ai_agent.py b/server/nodes/agent/ai_agent.py
index b080fc47..c5e3724a 100644
--- a/server/nodes/agent/ai_agent.py
+++ b/server/nodes/agent/ai_agent.py
@@ -51,12 +51,17 @@ class AIAgentParams(BaseModel):
)
# ---- "Options" group (collapsed by default in the parameter panel) ----
- temperature: float = Field(
- default=0.7, ge=0.0, le=2.0,
+ # default=None so an unset value falls through to ``agent.default_temperature``
+ # in server/config/llm_defaults.json (resolved by _resolve_temperature).
+ temperature: Optional[float] = Field(
+ default=None, ge=0.0, le=2.0,
json_schema_extra={"numberStepSize": 0.1, "group": "options"},
)
+ # default=None so an unset value is absent from the dumped dict and the
+ # backend (_resolve_max_tokens) falls through to the per-model default
+ # in server/config/llm_defaults.json instead of being silently capped.
max_tokens: Optional[int] = Field(
- default=1000, ge=1, le=200000,
+ default=None, ge=1, le=200000,
json_schema_extra={"group": "options"},
)
diff --git a/server/nodes/agent/chat_agent.py b/server/nodes/agent/chat_agent.py
index 1e58e8ce..f8295744 100644
--- a/server/nodes/agent/chat_agent.py
+++ b/server/nodes/agent/chat_agent.py
@@ -44,12 +44,17 @@ class ChatAgentParams(BaseModel):
)
# ---- "Options" group (collapsed by default in the parameter panel) ----
- temperature: float = Field(
- default=0.7, ge=0.0, le=2.0,
+ # default=None so an unset value falls through to ``agent.default_temperature``
+ # in server/config/llm_defaults.json (resolved by _resolve_temperature).
+ temperature: Optional[float] = Field(
+ default=None, ge=0.0, le=2.0,
json_schema_extra={"group": "options"},
)
+ # default=None so an unset value is absent from the dumped dict and the
+ # backend (_resolve_max_tokens) falls through to the per-model default
+ # in server/config/llm_defaults.json instead of being silently capped.
max_tokens: Optional[int] = Field(
- default=1000, ge=1, le=200000,
+ default=None, ge=1, le=200000,
json_schema_extra={"group": "options"},
)
diff --git a/server/nodes/model/_base.py b/server/nodes/model/_base.py
index 5195e283..bd4aa099 100644
--- a/server/nodes/model/_base.py
+++ b/server/nodes/model/_base.py
@@ -39,11 +39,16 @@ class ChatModelParams(BaseModel):
default="",
json_schema_extra={"placeholder": "Select a model..."},
)
- temperature: float = Field(
- default=0.7, ge=0.0, le=2.0,
+ # default=None so an unset value falls through to ``agent.default_temperature``
+ # in server/config/llm_defaults.json (resolved by _resolve_temperature).
+ temperature: Optional[float] = Field(
+ default=None, ge=0.0, le=2.0,
json_schema_extra={"numberStepSize": 0.1},
)
- max_tokens: Optional[int] = Field(default=1000, ge=1, le=200000)
+ # default=None so an unset value is absent from the dumped dict and the
+ # backend (_resolve_max_tokens) falls through to the per-model default
+ # in server/config/llm_defaults.json instead of being silently capped.
+ max_tokens: Optional[int] = Field(default=None, ge=1, le=200000)
system_prompt: Optional[str] = Field(
default="",
json_schema_extra={"rows": 3, "placeholder": "You are a helpful assistant."},
diff --git a/server/services/ai.py b/server/services/ai.py
index a59df210..3e1c934f 100644
--- a/server/services/ai.py
+++ b/server/services/ai.py
@@ -44,6 +44,7 @@ from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, Base
_ailog("importing langgraph + pydantic...")
from langchain_core.tools import StructuredTool
from langgraph.graph import StateGraph, END
+from langgraph.errors import GraphRecursionError
from pydantic import BaseModel, Field, create_model
_ailog("eager LangChain imports done")
import json
@@ -375,7 +376,10 @@ def _resolve_temperature(flattened: dict, model: str, provider: str, thinking_en
from services.model_registry import get_model_registry
registry = get_model_registry()
- user_temp = float(flattened.get('temperature', 0.7))
+ user_val = flattened.get('temperature')
+ if user_val is None:
+ user_val = registry.get_agent_defaults()["default_temperature"]
+ user_temp = float(user_val)
# O-series reasoning models always require temperature=1
if registry.is_reasoning_model(model, provider):
@@ -847,14 +851,18 @@ def create_tool_node(tool_executor: Callable):
def should_continue(state: AgentState) -> str:
- """Determine if the agent should continue or end.
-
- This is the conditional edge function for LangGraph.
- Returns "tools" to execute pending tool calls, or "end" to finish.
+ """Conditional edge: route to ``tools`` when the LLM emitted tool calls,
+ otherwise terminate the graph.
+
+ The previous implementation tracked a custom ``iteration`` counter to cap
+ the loop. That counter shared a name with LangGraph internals and was
+ being mutated unexpectedly, causing the edge to return ``"end"`` on the
+ first tool call. The standard LangGraph pattern is to route purely on
+ pending tool calls and rely on ``recursion_limit`` (passed via
+ ``ainvoke(config=...)``) as the safety net.
"""
- if state.get("should_continue", False):
- if state.get("iteration", 0) < state.get("max_iterations", 10):
- return "tools"
+ if state.get("pending_tool_calls"):
+ return "tools"
return "end"
@@ -1325,11 +1333,14 @@ class AIService:
from services.model_registry import get_model_registry
registry = get_model_registry()
- # Force temperature=1 for reasoning models (regardless of thinking mode)
+ # Reasoning models (listed in ``reasoning_models`` per provider in
+ # llm_defaults.json) do not support the ``temperature`` parameter —
+ # the API rejects requests that send any value, including 1. Drop
+ # it entirely rather than pass an explicit value.
if registry.is_reasoning_model(model, provider):
- if kwargs.get('temperature', 1) != 1:
- logger.info(f"[AI] Reasoning model '{model}': forcing temperature to 1 (was {kwargs.get('temperature')})")
- kwargs['temperature'] = 1
+ if 'temperature' in kwargs:
+ logger.info(f"[AI] Reasoning model '{model}': omitting temperature param")
+ kwargs.pop('temperature', None)
if thinking and thinking.enabled:
thinking_type = registry.get_thinking_type(model, provider)
@@ -1907,9 +1918,29 @@ class AIService:
"history_count": history_count
})
- # Execute the graph using ainvoke for proper async support
- # This allows async tool nodes and WebSocket broadcasts to work correctly
- final_state = await agent_graph.ainvoke(initial_state)
+ # Execute the graph. ``recursion_limit`` (cap on agent/tool
+ # supersteps) is sourced from llm_defaults.json so deployments
+ # tune it without code changes. On hit, surface what the agent
+ # produced before stopping rather than the raw library error.
+ from services.model_registry import get_model_registry
+ recursion_limit = int(get_model_registry().get_agent_defaults()["recursion_limit"])
+ final_state = None
+ try:
+ async for snapshot in agent_graph.astream(
+ initial_state,
+ config={"recursion_limit": recursion_limit},
+ stream_mode="values",
+ ):
+ final_state = snapshot
+ except GraphRecursionError:
+ msgs = (final_state or {}).get("messages") or []
+ raise ValueError(
+ f"Agent reached the {recursion_limit}-superstep recursion "
+ f"limit without producing a final answer "
+ f"({len(msgs)} message(s) exchanged). "
+ f"Adjust agent.recursion_limit in llm_defaults.json or "
+ f"simplify the task."
+ )
# Extract the AI response (last message in the accumulated messages)
all_messages = final_state["messages"]
@@ -2345,8 +2376,29 @@ class AIService:
"thinking_content": None
}
- # Execute the graph
- final_state = await agent_graph.ainvoke(initial_state)
+ # Execute the graph. ``recursion_limit`` is sourced from
+ # llm_defaults.json; see ``execute_agent`` for the rationale.
+ from services.model_registry import get_model_registry
+ recursion_limit = int(
+ get_model_registry().get_agent_defaults()["recursion_limit"]
+ )
+ final_state = None
+ try:
+ async for snapshot in agent_graph.astream(
+ initial_state,
+ config={"recursion_limit": recursion_limit},
+ stream_mode="values",
+ ):
+ final_state = snapshot
+ except GraphRecursionError:
+ msgs = (final_state or {}).get("messages") or []
+ raise ValueError(
+ f"Agent reached the {recursion_limit}-superstep "
+ f"recursion limit without producing a final answer "
+ f"({len(msgs)} message(s) exchanged). "
+ f"Adjust agent.recursion_limit in llm_defaults.json "
+ f"or simplify the task."
+ )
# Extract response
all_messages = final_state["messages"]
diff --git a/server/services/compaction.py b/server/services/compaction.py
index 10f9eb55..30f293f4 100644
--- a/server/services/compaction.py
+++ b/server/services/compaction.py
@@ -21,9 +21,9 @@ if TYPE_CHECKING:
logger = get_logger(__name__)
-# Default fraction of context window used as compaction threshold
-# Overridden by user_settings.compaction_ratio if configured.
-DEFAULT_CONTEXT_THRESHOLD_RATIO = 0.5
+# Compaction ratio is sourced from server/config/llm_defaults.json
+# (``agent.compaction.ratio``). Per-user override in ``user_settings`` still
+# wins via :meth:`CompactionService._get_compaction_ratio`.
def _extract_text_from_response(content) -> str:
@@ -77,7 +77,9 @@ class CompactionService:
self._ai_service = ai_service
async def _get_compaction_ratio(self) -> float:
- """Get compaction ratio from user settings, falling back to default."""
+ """Per-user override from ``user_settings.compaction_ratio``, falling
+ back to ``agent.compaction.ratio`` in ``llm_defaults.json``."""
+ from services.model_registry import get_model_registry
try:
settings = await self._db.get_user_settings("default")
if settings and "compaction_ratio" in settings:
@@ -86,17 +88,20 @@ class CompactionService:
return ratio
except Exception:
pass
- return DEFAULT_CONTEXT_THRESHOLD_RATIO
+ return float(get_model_registry().get_agent_defaults()["compaction"]["ratio"])
- def get_model_threshold(self, model: str, provider: str, ratio: float = DEFAULT_CONTEXT_THRESHOLD_RATIO) -> int:
+ def get_model_threshold(self, model: str, provider: str, ratio: Optional[float] = None) -> int:
"""Compute compaction threshold based on model's context window.
- Returns ratio * model's context length, floored to the global
- default if the registry has no data. Per-session custom_threshold
- (checked in track()) still takes priority over this value.
+ ``ratio`` defaults to ``agent.compaction.ratio`` from
+ ``llm_defaults.json``; pass an explicit value to override per call.
+ Per-session ``custom_threshold`` (checked in :meth:`track`) still
+ wins over the value returned here.
"""
from services.model_registry import get_model_registry
registry = get_model_registry()
+ if ratio is None:
+ ratio = float(registry.get_agent_defaults()["compaction"]["ratio"])
context_length = registry.get_context_length(model, provider)
model_threshold = int(context_length * ratio)
# Never go below the configured minimum (ge=10000 in Settings)
diff --git a/server/services/model_registry.py b/server/services/model_registry.py
index 38304c71..c0a2c1da 100644
--- a/server/services/model_registry.py
+++ b/server/services/model_registry.py
@@ -483,6 +483,16 @@ class ModelRegistryService:
logger.warning(f"Could not load llm_defaults.json: {e}")
self._llm_defaults = {"providers": {}}
+ def get_agent_defaults(self) -> Dict[str, Any]:
+ """Return the ``agent`` block from ``llm_defaults.json``.
+
+ Holds runtime defaults shared across the LangGraph executor and
+ compaction layer (e.g. ``recursion_limit``, ``compaction.ratio``).
+ Callers should ``.get(...)`` with their own fallback so the system
+ still boots if the JSON is missing or partial.
+ """
+ return self._llm_defaults.get("agent", {})
+
def _get_default_max_output_tokens(self, provider: str, model: str) -> int:
"""Fallback: get max output tokens from llm_defaults.json."""
providers = self._llm_defaults.get("providers", {})