Fix/general run

.gitignore

@@ -43,6 +43,7 @@ MANIFEST
 *.spec
 release.sh
 build.sh
+scripts/tb21-bench.sh
 *.html
 
 # Installer logs

ms_agent/a2a/agent_card.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import os
 from typing import Any, Dict, List
 

ms_agent/a2a/client.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import os
 from typing import Any, Dict, List, Optional
 

ms_agent/a2a/errors.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 from ms_agent.utils.logger import get_logger
 
 logger = get_logger()

ms_agent/a2a/executor.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import logging
 import sys
 from typing import Any

ms_agent/acp/client.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import asyncio
 from contextlib import asynccontextmanager
 from typing import Any, Dict, List, Optional

ms_agent/acp/errors.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 from acp import RequestError
 
 

ms_agent/acp/http_adapter.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import asyncio
 import os
 import secrets

ms_agent/acp/proxy.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import logging
 import os
 import sys

ms_agent/acp/registry.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import os
 import sys
 from typing import Any, Dict

ms_agent/acp/server.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import logging
 import sys
 from typing import Any

ms_agent/acp/session_store.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 import asyncio
 import os
 import uuid

ms_agent/agent/agent.yaml

@@ -56,6 +56,21 @@ callbacks:
   - input_callback
 
 tools:
+  file_system:
+    mcp: false
+    include:
+      - write_file
+      - read_file
+      - edit_file
+      - grep
+      - glob
+  code_executor:
+    mcp: false
+    implementation: python_env
+    include:
+      - shell_executor
+      - python_executor
+      - notebook_executor
 
 help: |
   A commonly use config, try whatever you want!

ms_agent/agent/base.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import os
 from abc import ABC, abstractmethod

ms_agent/agent/llm_agent.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import importlib
@@ -15,7 +16,6 @@
 
 from ms_agent.agent.runtime import Runtime
 from ms_agent.callbacks import Callback, callbacks_mapping
-from ms_agent.knowledge_search import SirchmunkSearch
 from ms_agent.llm.llm import LLM
 from ms_agent.llm.utils import Message, ToolResult
 from ms_agent.memory import Memory, get_memory_meta_safe, memory_mapping

ms_agent/callbacks/repetition_guard.py

@@ -0,0 +1,189 @@
+"""Callback that detects stuck loops and injects strategy-variation prompts.
+
+When the agent repeats the same tool call (same name + same arguments) multiple
+times without making progress, this callback injects a user message encouraging
+a different approach.  After ``max_warnings`` injections, the callback forces
+the agent to stop to avoid wasting remaining API budget.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from collections import Counter
+from dataclasses import dataclass
+from typing import List
+
+from omegaconf import DictConfig
+
+from ms_agent.agent.runtime import Runtime
+from ms_agent.callbacks.base import Callback
+from ms_agent.llm.utils import Message
+from ms_agent.utils.logger import get_logger
+
+logger = get_logger()
+
+_DEFAULT_THRESHOLD = 3
+_DEFAULT_LOOKBACK = 8
+_DEFAULT_MAX_WARNINGS = 2
+
+_STRATEGY_PROMPT = (
+    "IMPORTANT: You have been repeating the same approach multiple times "
+    "without making progress. The exact same tool call or command has been "
+    "attempted {count} times with the same arguments.\n\n"
+    "Repeated action: {description}\n\n"
+    "Please try a fundamentally different strategy. Consider:\n"
+    "1. Check whether the prerequisites for your approach are actually met.\n"
+    "2. Break the problem into smaller, verifiable sub-steps.\n"
+    "3. Use a different tool or a different command to achieve the same goal.\n"
+    "4. Read error messages carefully — they often suggest specific fixes.\n"
+    "5. If a dependency is missing, install it before retrying.\n\n"
+    "Do NOT repeat the same command again."
+)
+
+_FORCE_STOP_PROMPT = (
+    "You have been warned multiple times about repeating the same approach, "
+    "but continue to retry the same failing action. "
+    "To conserve resources, execution is being stopped. "
+    "Please review the errors above and formulate a new plan before "
+    "the next attempt."
+)
+
+
+@dataclass(frozen=True)
+class _Repetition:
+    key: str
+    tool_name: str
+    count: int
+    description: str
+
+
+def _args_hash(arguments: str) -> str:
+    try:
+        parsed = json.loads(arguments) if isinstance(arguments, str) else arguments
+    except (json.JSONDecodeError, TypeError):
+        parsed = arguments
+    canonical = json.dumps(parsed, sort_keys=True, ensure_ascii=False)
+    return hashlib.md5(canonical.encode("utf-8")).hexdigest()[:12]
+
+
+class RepetitionGuardCallback(Callback):
+    """Detects stuck loops and injects strategy-variation prompts."""
+
+    def __init__(self, config: DictConfig) -> None:
+        super().__init__(config)
+        guard_cfg = getattr(config, "repetition_guard", None) or {}
+        self.threshold: int = int(guard_cfg.get("threshold", _DEFAULT_THRESHOLD))
+        self.lookback: int = int(guard_cfg.get("lookback_rounds", _DEFAULT_LOOKBACK))
+        self.max_warnings: int = int(guard_cfg.get("max_warnings", _DEFAULT_MAX_WARNINGS))
+        self._warnings_given: int = 0
+        self._warned_keys: set[str] = set()
+
+    async def after_tool_call(self, runtime: Runtime, messages: List[Message]) -> None:
+        if self._warnings_given >= self.max_warnings:
+            logger.info(
+                "[RepetitionGuard] Max warnings (%d) reached — forcing stop.",
+                self.max_warnings,
+            )
+            messages.append(Message(role="user", content=_FORCE_STOP_PROMPT))
+            runtime.should_stop = True
+            return
+
+        recent = _extract_recent_tool_calls(messages, self.lookback)
+        repetition = _detect_repetition(recent, self.threshold)
+
+        if repetition is None:
+            return
+        if repetition.key in self._warned_keys:
+            return
+
+        self._warned_keys.add(repetition.key)
+        self._warnings_given += 1
+        logger.info(
+            "[RepetitionGuard] Stuck loop detected (%s, %dx). Injecting strategy prompt. "
+            "Warning %d/%d.",
+            repetition.tool_name,
+            repetition.count,
+            self._warnings_given,
+            self.max_warnings,
+        )
+        runtime.should_stop = False
+        prompt = _STRATEGY_PROMPT.format(
+            count=repetition.count,
+            description=repetition.description,
+        )
+        messages.append(Message(role="user", content=prompt))
+
+
+def _extract_recent_tool_calls(
+    messages: List[Message],
+    lookback: int,
+) -> list[tuple[str, str, str]]:
+    """Return ``(tool_name, args_hash, description)`` for recent rounds.
+
+    Walks backwards through *messages* collecting assistant tool-call entries.
+    Stops after scanning *lookback* assistant-with-tool-calls messages.
+    """
+    calls: list[tuple[str, str, str]] = []
+    rounds_seen = 0
+
+    for msg in reversed(messages):
+        if rounds_seen >= lookback:
+            break
+        if msg.role == "assistant" and msg.tool_calls:
+            rounds_seen += 1
+            for tc in msg.tool_calls:
+                name = tc.get("tool_name", "")
+                raw_args = tc.get("arguments", "{}")
+                ah = _args_hash(raw_args)
+                desc = _summarize_call(name, raw_args)
+                calls.append((name, ah, desc))
+
+    return calls
+
+
+def _detect_repetition(
+    calls: list[tuple[str, str, str]],
+    threshold: int,
+) -> _Repetition | None:
+    if not calls:
+        return None
+
+    key_counts: Counter[str] = Counter()
+    key_to_info: dict[str, tuple[str, str]] = {}
+
+    for name, ah, desc in calls:
+        key = f"{name}:{ah}"
+        key_counts[key] += 1
+        if key not in key_to_info:
+            key_to_info[key] = (name, desc)
+
+    most_common_key, count = key_counts.most_common(1)[0]
+    if count < threshold:
+        return None
+
+    tool_name, description = key_to_info[most_common_key]
+    return _Repetition(
+        key=most_common_key,
+        tool_name=tool_name,
+        count=count,
+        description=description,
+    )
+
+
+def _summarize_call(tool_name: str, raw_args: str) -> str:
+    try:
+        args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
+    except (json.JSONDecodeError, TypeError):
+        args = raw_args
+
+    if isinstance(args, dict) and "command" in args:
+        cmd = str(args["command"])
+        if len(cmd) > 120:
+            cmd = cmd[:120] + "..."
+        return f"{tool_name}(command={cmd})"
+
+    summary = json.dumps(args, ensure_ascii=False)
+    if len(summary) > 120:
+        summary = summary[:120] + "..."
+    return f"{tool_name}({summary})"

ms_agent/callbacks/utils.py

@@ -1,4 +1,8 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
 from ms_agent.callbacks.input_callback import InputCallback
+from ms_agent.callbacks.repetition_guard import RepetitionGuardCallback
 
-callbacks_mapping = {'input_callback': InputCallback}
+callbacks_mapping = {
+    'input_callback': InputCallback,
+    'repetition_guard': RepetitionGuardCallback,
+}

ms_agent/capabilities/async_task.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import logging

ms_agent/capabilities/descriptor.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import os
 import shutil

ms_agent/capabilities/mcp_server.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import argparse
 import json

ms_agent/capabilities/registry.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import logging
 from typing import Any, Awaitable, Callable

ms_agent/capabilities/wrappers/agent_delegate.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import contextlib
 import logging

ms_agent/capabilities/wrappers/code_genesis.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import os

ms_agent/capabilities/wrappers/deep_research.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import os

ms_agent/capabilities/wrappers/doc_research.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import logging

ms_agent/capabilities/wrappers/filesystem.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import os
 from typing import Any

ms_agent/capabilities/wrappers/fin_research.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import os

ms_agent/capabilities/wrappers/singularity_cinema.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import asyncio
 import os

ms_agent/command/builtin/context_cmds.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 from ms_agent.command.router import CommandRouter
 from ms_agent.command.types import (
     CommandContext,

ms_agent/command/types.py

@@ -1,3 +1,4 @@
+from __future__ import annotations
 from dataclasses import dataclass, field
 from enum import Enum
 from typing import Any, Awaitable, Callable, Optional

ms_agent/llm/anthropic_llm.py

@@ -390,7 +390,7 @@ def _call_llm(self,
         else:
             return self.client.messages.create(**params)
 
-    @retry(max_attempts=LLM.retry_count, delay=1.0)
+    @retry(max_attempts=LLM.retry_count, delay=3.0)
     def generate(self,
                  messages: List[Message],
                  tools: Optional[List[Tool]] = None,

ms_agent/llm/openai_llm.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import httpx
 import inspect
@@ -211,7 +213,7 @@ def format_tools(self,
             tools = None
         return tools
 
-    @retry(max_attempts=LLM.retry_count, delay=1.0)
+    @retry(max_attempts=LLM.retry_count, delay=3.0)
     def generate(self,
                  messages: List[Message],
                  tools: Optional[List[Tool]] = None,