mirivlad/ducklm
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Stabilize live chat tool loop

Browse Source
This commit is contained in:
mirivlad
2026-05-22 21:13:58 +08:00
parent eef010f227
commit 1291281d7e
9 changed files with 182 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
duck_core/api.py
+25 -9
View File
@@ -261,6 +261,7 @@ def create_app() -> FastAPI:
memory_records=memory_records,
skill_summary=await selected_skill_summary(body.message),
reasoning=body.reasoning,
reflect=bool(settings.enable_reflection),
)
await conversations.add_message(
conversation.conversation_id,
@@ -454,8 +455,7 @@ def create_app() -> FastAPI:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(tool_observations, ensure_ascii=False, indent=2),
"content": runtime.format_tool_observations_for_model(tool_observations),
},
]
yield runtime_status(
@@ -531,7 +531,13 @@ def create_app() -> FastAPI:
"generation_stats": generation_stats.summary(),
},
)
asyncio.create_task(runtime.complete_postprocessing(task.task_id, content))
asyncio.create_task(
runtime.complete_postprocessing(
task.task_id,
content,
reflect=bool(settings.enable_reflection),
)
)
yield sse(
"done",
{
@@ -707,8 +713,7 @@ def create_app() -> FastAPI:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(tool_observations, ensure_ascii=False, indent=2),
"content": runtime.format_tool_observations_for_model(tool_observations),
},
]
yield runtime_status(task_id, "answering", "Формирую ответ...")
@@ -763,7 +768,13 @@ def create_app() -> FastAPI:
"generation_stats": generation_stats.summary(),
},
)
asyncio.create_task(runtime.complete_postprocessing(task_id, content))
asyncio.create_task(
runtime.complete_postprocessing(
task_id,
content,
reflect=bool(settings.enable_reflection),
)
)
yield sse(
"done",
{
@@ -878,8 +889,7 @@ def create_app() -> FastAPI:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(tool_observations, ensure_ascii=False, indent=2),
"content": runtime.format_tool_observations_for_model(tool_observations),
},
]
yield runtime_status(task_id, "answering", "Формирую ответ...")
@@ -934,7 +944,13 @@ def create_app() -> FastAPI:
"generation_stats": generation_stats.summary(),
},
)
asyncio.create_task(runtime.complete_postprocessing(task_id, content))
asyncio.create_task(
runtime.complete_postprocessing(
task_id,
content,
reflect=bool(settings.enable_reflection),
)
)
yield sse(
"done",
{
duck_core/config.py
+2
View File
@@ -23,6 +23,7 @@ class Settings:
max_memory_tokens: int = 8000
max_skill_tokens: int = 6000
qdrant_url: str = "http://127.0.0.1:6333"
enable_reflection: int = 0
skip_live_llm_tests: int = 0
@property
@@ -52,5 +53,6 @@ def get_settings() -> Settings:
max_memory_tokens=int(os.getenv("DUCK_MAX_MEMORY_TOKENS", "8000")),
max_skill_tokens=int(os.getenv("DUCK_MAX_SKILL_TOKENS", "6000")),
qdrant_url=os.getenv("QDRANT_URL", "http://127.0.0.1:6333"),
enable_reflection=int(os.getenv("DUCK_ENABLE_REFLECTION", "0")),
skip_live_llm_tests=int(os.getenv("DUCK_SKIP_LIVE_LLM_TESTS", "0")),
)
duck_core/runtime_loop.py
+60 -8
View File
@@ -19,6 +19,7 @@ from duck_core.tools.gateway import ToolGateway
logger = logging.getLogger(__name__)
ACTION_DIRECTIVE_SCHEMA = load_json_schema("duck_core/schemas/action_directive.schema.json")
MAX_TOOL_OBSERVATION_TEXT_CHARS = 2000
@dataclass
@@ -96,8 +97,7 @@ class RuntimeLoop:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(tool_observations, ensure_ascii=False, indent=2),
"content": self.format_tool_observations_for_model(tool_observations),
},
]
await self.event_store.append(
@@ -232,8 +232,7 @@ class RuntimeLoop:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(tool_observations, ensure_ascii=False, indent=2),
"content": self.format_tool_observations_for_model(tool_observations),
},
]
await self.event_store.append(task_id, "model_call_started", {"role": "thinker"})
@@ -394,9 +393,13 @@ class RuntimeLoop:
"tool": tool_name,
"reason": action.get("reason"),
"decision": decision,
"action": action,
"result": result_payload,
}
def _action_key(self, action: dict[str, Any]) -> str:
return json.dumps(action, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
async def _approval_action_index(self, task_id: str, action: dict[str, Any]) -> int:
events = await self.event_store.list_events(task_id)
for event in reversed(events):
@@ -413,6 +416,7 @@ class RuntimeLoop:
messages: list[dict[str, str]],
workspace: str | None,
start_index: int = 1,
seen_action_keys: set[str] | None = None,
) -> list[dict[str, Any]]:
try:
await self.event_store.append(task_id, "model_call_started", {"role": "action"})
@@ -443,6 +447,21 @@ class RuntimeLoop:
{"index": index, "ok": False, "error": "Action must be an object"}
)
continue
action_key = self._action_key(action)
if seen_action_keys is not None and action_key in seen_action_keys:
await self.event_store.append(
task_id,
"tool_call_skipped",
{
"index": index,
"tool": str(action.get("tool", "")),
"reason": "duplicate_action",
"action": action,
},
)
continue
if seen_action_keys is not None:
seen_action_keys.add(action_key)
tool_name = str(action.get("tool", ""))
await self.event_store.append(
task_id,
@@ -478,6 +497,7 @@ class RuntimeLoop:
"index": index,
"tool": tool_name,
"reason": action.get("reason"),
"action": action,
"requires_approval": True,
"approval_id": approval.approval_id if approval else None,
"result": result_payload,
@@ -494,11 +514,39 @@ class RuntimeLoop:
"index": index,
"tool": tool_name,
"reason": action.get("reason"),
"action": action,
"result": result_payload,
}
)
return observations
def format_tool_observations_for_model(self, observations: list[dict[str, Any]]) -> str:
return "tool_observations:\n" + json.dumps(
self.compact_tool_observations_for_model(observations),
ensure_ascii=False,
indent=2,
)
def compact_tool_observations_for_model(
self, observations: list[dict[str, Any]]
) -> list[dict[str, Any]]:
return [self._compact_observation_for_model(observation) for observation in observations]
def _compact_observation_for_model(self, value: Any) -> Any:
if isinstance(value, dict):
return {key: self._compact_observation_for_model(item) for key, item in value.items()}
if isinstance(value, list):
return [self._compact_observation_for_model(item) for item in value]
if isinstance(value, str) and len(value) > MAX_TOOL_OBSERVATION_TEXT_CHARS:
half = MAX_TOOL_OBSERVATION_TEXT_CHARS // 2
omitted = len(value) - MAX_TOOL_OBSERVATION_TEXT_CHARS
return (
value[:half]
+ f"\n... [truncated {omitted} chars for model context] ...\n"
+ value[-half:]
)
return value
async def _append_command_audit(
self,
task_id: str,
@@ -538,14 +586,18 @@ class RuntimeLoop:
initial_observations: list[dict[str, Any]] | None = None,
) -> list[dict[str, Any]]:
all_observations = list(initial_observations or [])
seen_action_keys = {
self._action_key(observation["action"])
for observation in all_observations
if isinstance(observation.get("action"), dict)
}
current_messages = messages
if all_observations:
current_messages = [
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(all_observations, ensure_ascii=False, indent=2),
"content": self.format_tool_observations_for_model(all_observations),
},
]
for _ in range(self.max_tool_iterations):
@@ -554,6 +606,7 @@ class RuntimeLoop:
current_messages,
workspace,
start_index=len(all_observations) + 1,
seen_action_keys=seen_action_keys,
)
if not observations:
return all_observations
@@ -564,8 +617,7 @@ class RuntimeLoop:
*messages,
{
"role": "user",
"content": "tool_observations:\n"
+ json.dumps(all_observations, ensure_ascii=False, indent=2),
"content": self.format_tool_observations_for_model(all_observations),
},
]
await self.event_store.append(