Stabilize live chat tool loop

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

View File

@ -21,6 +21,7 @@ DUCK_MAX_INPUT_TOKENS=49152
DUCK_MAX_RECENT_EVENTS_TOKENS=12000
DUCK_MAX_MEMORY_TOKENS=8000
DUCK_MAX_SKILL_TOKENS=6000
DUCK_ENABLE_REFLECTION=0
QDRANT_URL=http://127.0.0.1:6333

View File

@ -48,9 +48,11 @@ WebChat доступен через FastAPI на `http://127.0.0.1:8000/`.
- MemoryStore в SQLite.
- MemoryPolicy через LLM role `memory_policy` с fallback в безопасный no-store режим.
- Structured JSON validation для `action` и `memory_policy`: невалидный JSON/schema violation не запускает tools и уходит в безопасный fallback.
- Tool observations компактируются перед повторной подачей в model context; полные outputs остаются в event/audit log.
- Duplicate tool actions в рамках одной задачи пропускаются, чтобы модель не выполняла одну и ту же команду повторно.
- VectorMemory adapter для Qdrant с локальной embedding-моделью или remote embeddings endpoint.
- Recall-фильтрация памяти через `recall` role.
- Reflection через `critic` role.
- Reflection через `critic` role доступна, но выключена по умолчанию в API (`DUCK_ENABLE_REFLECTION=0`), чтобы не забивать single-slot llama во время интерактивного чата.
- ExperienceRecorder и skill update proposals.
- Scripts для llama-server, verification и benchmark.
- Docker compose для Qdrant.
@ -69,6 +71,8 @@ WebChat доступен через FastAPI на `http://127.0.0.1:8000/`.
- Skill candidate selection теперь используется в обычном и streaming chat.
- `scripts/duck.sh status --probe` и `scripts/duck-mtp.sh status --probe` показывают live-состояние DuckLM runtime, model backend и vector memory.
- Structured utility-outputs валидируются локально по JSON schema; это защищает tool loop и memory writes от мусора модели.
- Live E2E выявил и исправил два runtime-дефекта: большие stdout больше не раздувают следующий planning prompt, повторяющиеся identical actions больше не исполняются повторно.
- Critic reflection по умолчанию выключена для API/WebChat, потому что на одном `llama-server --parallel 1` она конкурировала с пользовательскими запросами и вызывала timeouts.
## Соответствие этапам из Ducklm.md
@ -88,6 +92,7 @@ WebChat доступен через FastAPI на `http://127.0.0.1:8000/`.
## Остаточные ограничения
- Qdrant и локальная embedding-модель должны быть доступны отдельно; при ошибках vector memory деградирует без падения runtime.
- `DUCK_ENABLE_REFLECTION=1` включает critic reflection после задач, но для single-slot llama это может заметно тормозить последующие запросы.
- Token speed считается приближённо по текущему estimator, а не по tokenizer конкретной модели.
- Skill selection сейчас keyword-based. LLM selection можно добавить позже, если понадобится.
- WebChat остаётся lightweight vanilla JS UI; это не production frontend framework.
@ -151,5 +156,6 @@ bash scripts/duck-mtp.sh logs --follow
1. Пройти live E2E checklist в WebChat на реальной модели.
2. Вынести runtime/model role routing в явный конфиг с fallback-политикой, оставив Qwen основным backend для всех ролей.
3. Расширить strict validation/fallback на `recall` и будущие structured utility-roles.
4. При необходимости заменить keyword skill selection на LLM-based selection.
5. Позже мигрировать FastAPI startup на lifespan.
4. Добавить WebChat runtime/status panel поверх `/v1/status?probe=true`.
5. При необходимости заменить keyword skill selection на LLM-based selection.
6. Позже мигрировать FastAPI startup на lifespan.

View File

@ -16,6 +16,10 @@ cp .env.example .env
The default `DUCK_MAIN_MODEL_PATH` points to `./models/Qwen3.6/nonMTP/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf`.
`DUCK_ENABLE_REFLECTION=0` is the recommended default for the local single-slot
stack. Set it to `1` only when you explicitly want critic reflection after each
chat and accept that it can slow down the next request.
3. Start DuckLM:
```bash

View File

@ -51,6 +51,9 @@ Structured utility roles are validated locally before side effects:
- `action` output must match `duck_core/schemas/action_directive.schema.json`; invalid directives are logged as `action_directive_failed` and no tool runs.
- `memory_policy` output must match its JSON schema; invalid decisions fall back to `should_store=false`.
- Large tool outputs are compacted before being sent back to the model; full outputs remain in task events and command audit.
- Duplicate identical tool actions in one task are skipped and logged as `tool_call_skipped`.
- Critic reflection is controlled by `DUCK_ENABLE_REFLECTION`; the default is off for responsive single-slot local chat.
Chat requests accept optional `reasoning`:

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",
{

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")),
)

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(

View File

@ -113,6 +113,7 @@ def test_stream_chat_forwards_reasoning_toggle_to_thinker(tmp_path, monkeypatch)
def test_stream_chat_runs_memory_policy_and_reflection_after_completion(tmp_path, monkeypatch):
monkeypatch.setenv("DUCK_DB_PATH", str(tmp_path / "duck.sqlite3"))
monkeypatch.setenv("DUCK_ENABLE_REFLECTION", "1")
async def fake_chat(self, role, messages, temperature=None, max_output_tokens=None, response_format=None):
if role == "action":

View File

@ -176,6 +176,41 @@ class FakeMalformedActionModelClient:
)
class FakeRepeatingActionModelClient:
async def chat(self, role, messages):
if role == "action":
return ModelResponse(
role=role,
model="local-main",
content=json.dumps(
{
"kind": "action_directive",
"intent": "repeat same action",
"risk_level": "low",
"actions": [
{
"tool": "file_read",
"args": {"path": "note.txt"},
"reason": "Read requested file",
}
],
}
),
reasoning_content=None,
raw={},
latency_ms=5.0,
)
assert role == "thinker"
return ModelResponse(
role=role,
model="local-main",
content="Final answer from first observation.",
reasoning_content=None,
raw={},
latency_ms=12.0,
)
@pytest.mark.asyncio
async def test_runtime_executes_action_directive_tool_and_finishes_with_observation(tmp_path):
(tmp_path / "note.txt").write_text("hello from tool")
@ -253,6 +288,48 @@ async def test_runtime_rejects_malformed_action_directive_before_tools(tmp_path)
assert not any(event.event_type == "tool_call_started" for event in events)
def test_runtime_compacts_large_tool_observations_for_model_context(tmp_path):
db_path = str(tmp_path / "duck.sqlite3")
task_store = TaskStore(db_path)
event_store = EventStore(db_path)
loop = RuntimeLoop(task_store, event_store, FakeToolModelClient())
compact = loop.format_tool_observations_for_model([
{
"tool": "shell_exec_safe",
"result": {
"ok": True,
"output": "A" * 2500 + "KEEP_TAIL",
"metadata": {"command": "ls /tmp"},
},
}
])
assert "tool_observations" in compact
assert "truncated" in compact
assert "KEEP_TAIL" in compact
assert len(compact) < 2300
@pytest.mark.asyncio
async def test_runtime_skips_duplicate_action_within_same_task(tmp_path):
(tmp_path / "note.txt").write_text("hello once")
db_path = str(tmp_path / "duck.sqlite3")
task_store = TaskStore(db_path)
event_store = EventStore(db_path)
loop = RuntimeLoop(task_store, event_store, FakeRepeatingActionModelClient())
result = await loop.run_chat("read note.txt", str(tmp_path), debug=True)
events = await event_store.list_events(result.task_id)
finished_tools = [event for event in events if event.event_type == "tool_call_finished"]
skipped_tools = [event for event in events if event.event_type == "tool_call_skipped"]
assert result.status == "completed"
assert len(finished_tools) == 1
assert len(skipped_tools) == 1
assert skipped_tools[0].payload["reason"] == "duplicate_action"
class FakeApprovalModelClient:
async def chat(self, role, messages):
if role == "action":