acli: fix Windows TUI (run_command, wheel scroll) + bench-driven fixes

dashscope/acli/__init__.py

@@ -3,7 +3,7 @@
 
 import uuid
 
-__version__ = "0.6.2"
+__version__ = "0.6.4"
 
 # Per-process identifier sent as x-dashscope-sdk-session-id so the
 # backend can group multi-turn requests from one CLI run.

dashscope/acli/cli/__init__.py

@@ -86,6 +86,33 @@
 _scheduler = None
 
 
+def _apply_cli_overrides(
+    config: Config,
+    protocol_override: str | None,
+    max_turns_override: str | None,
+) -> None:
+    """Apply --protocol/--max-turns to a loaded config; exit on bad input."""
+    if protocol_override:
+        if protocol_override.lower() in ("openai", "anthropic"):
+            config.protocol = protocol_override.lower()
+        else:
+            print(
+                f"Error: unknown protocol '{protocol_override}' "
+                f"(choices: openai, anthropic)",
+            )
+            sys.exit(1)
+
+    if max_turns_override is not None:
+        try:
+            config.max_turns = int(max_turns_override)
+        except ValueError:
+            print(
+                f"Error: --max-turns must be an integer, got "
+                f"'{max_turns_override}'",
+            )
+            sys.exit(1)
+
+
 def main():
     # Parse --cli / --tui / --dry-run flags (can appear anywhere in argv)
     use_cli = "--cli" in sys.argv
@@ -146,7 +173,7 @@ def main():
             )
             print(
                 "  --max-turns <n>         Override max conversation "
-                "turns (default 1000)",
+                "turns (default from config: 50)",
             )
             print(
                 "  --dry-run               Preview current config (loaded "
@@ -175,6 +202,13 @@ def main():
                 print('Usage: acli -c "your request"')
                 sys.exit(1)
             config = Config.load()
+            # This branch returns before main()'s override tail, so the
+            # flags must be applied here or oneshot silently ignores them.
+            _apply_cli_overrides(
+                config,
+                protocol_override,
+                max_turns_override,
+            )
             asyncio.run(_run_oneshot(config, prompt))
             return
         elif arg in ("example", "examples"):
@@ -186,27 +220,7 @@ def main():
             _mcp_server_main()
             return
     config = Config.load()
-    # Apply --protocol override
-    if protocol_override:
-        if protocol_override.lower() in ("openai", "anthropic"):
-            config.protocol = protocol_override.lower()
-        else:
-            print(
-                f"Error: unknown protocol '{protocol_override}' "
-                f"(choices: openai, anthropic)",
-            )
-            sys.exit(1)
-
-    # Apply --max-turns override
-    if max_turns_override is not None:
-        try:
-            config.max_turns = int(max_turns_override)
-        except ValueError:
-            print(
-                f"Error: --max-turns must be an integer, got "
-                f"'{max_turns_override}'",
-            )
-            sys.exit(1)
+    _apply_cli_overrides(config, protocol_override, max_turns_override)
 
     # Handle --dry-run: preview configuration without starting the agent
     if use_dry_run:

dashscope/acli/cli/handlers_setup.py

@@ -192,7 +192,7 @@ def _setup_finalize(config: Config, agent) -> None:
 
 def _apply_preset_bailian(config: Config) -> None:
     config.provider = "tongyi"
-    config.model = "qwen3.7-plus"
+    config.model = "qwen3.8-max"
     config.enabled_capabilities = [
         "bailian.mcp",
         "bailian.cli",
@@ -284,7 +284,7 @@ async def _handle_setup(config: Config, agent) -> None:
     console.print("\nSelect a configuration mode:")
     console.print(
         "  [cyan][1][/cyan] [bold]Alibaba/Bailian[/bold] (default) — "
-        "tongyi/qwen3.7-plus + bailian.mcp/cli",
+        "tongyi/qwen3.8-max + bailian.mcp/cli",
     )
     console.print(
         "  [cyan][2][/cyan] China general        — "

dashscope/acli/cli/repl.py

@@ -102,7 +102,10 @@ async def _run_loop(config: Config):
     sync_extensions_into_catalog(_ext)
     # _load_plugins() — deprecated; plugins load via the hooks mechanism
     provider = get_provider_chain(config)
-    executor = Executor(auto_approve=config.auto_approve)
+    executor = Executor(
+        auto_approve=config.auto_approve,
+        confirm_mode=config.confirm_mode,
+    )
 
     # Initialize memory
     # Memory client managed by handlers_profile
@@ -248,7 +251,7 @@ async def _run_loop(config: Config):
     if not WORKSPACE_CONFIG_FILE.exists():
         has_api_key = bool(config.api_key)
         using_defaults = (
-            config.provider == "tongyi" and config.model == "qwen3.7-plus"
+            config.provider == "tongyi" and config.model == "qwen3.8-max"
         )
         if not has_api_key or using_defaults:
             await _handle_setup(config, agent)

dashscope/acli/cli/runners.py

@@ -67,13 +67,17 @@ async def _run_oneshot(config: Config, prompt: str):
         sys.exit(1)
 
     provider = get_provider_chain(config)
-    executor = Executor()
+    executor = Executor(
+        auto_approve=config.auto_approve,
+        confirm_mode=config.confirm_mode,
+    )
     agent = Agent(
         provider=provider,
         executor=executor,
         max_turns=config.max_turns,
         provider_name=config.provider,
         model_name=config.model,
+        oneshot=True,
     )
 
     # Pin parent agent ref for local.subagent / local.delegate BEFORE platform
@@ -278,7 +282,10 @@ def _run_tui_mode(config: Config):
     # CLI mode uses the full provider chain so temporary failures can fall back
     # to configured alternatives instead of dying immediately.
     provider = get_provider_chain(config)
-    executor = Executor(auto_approve=config.auto_approve)
+    executor = Executor(
+        auto_approve=config.auto_approve,
+        confirm_mode=config.confirm_mode,
+    )
 
     from dashscope.acli.agents.delegate import (
         set_config as set_delegate_config,

dashscope/acli/config.py

@@ -23,6 +23,7 @@
 
 PROVIDER_MODELS = {
     "tongyi": [
+        "qwen3.8-max",
         "qwen3.7-max",
         "qwen3.7-plus",
         "qwen3.5-plus",
@@ -226,7 +227,7 @@ def from_dict(cls, d: dict) -> DelegationConfig:
 @dataclass
 class Config:
     provider: str = "tongyi"
-    model: str = "qwen3.7-plus"
+    model: str = "qwen3.8-max"
     # Dual-LLM: when thinking_model is set, Plan/Thinking loop phases route to
     # the thinking model; Execute uses the execution model (provider/model
     # above).
@@ -241,6 +242,10 @@ class Config:
     openai_api_key: str = ""
     base_url: str = ""
     auto_approve: bool = False
+    # Which tool executions prompt for confirmation:
+    #   "all"       - CONFIRM and DANGEROUS both prompt (classic)
+    #   "dangerous" - only DANGEROUS prompts; CONFIRM auto-passes
+    confirm_mode: str = "dangerous"
     max_turns: int = 50
     timeout: int = 30
     mcp_servers: list[MCPServerConfig] = field(default_factory=list)
@@ -277,9 +282,11 @@ class Config:
     protocol: str = "openai"  # "openai" | "anthropic"
     tui: bool = True
     # False = let the terminal handle the mouse (native selection).
-    # Default off on Windows: Textual's mouse-capture path is crash-prone
-    # in PowerShell/conhost; native terminal mouse still allows selection.
-    tui_mouse: bool = field(default_factory=lambda: os.name != "nt")
+    # Capture is on everywhere, including Windows: without it the wheel goes
+    # to the console host, which scrolls the whole alternate screen — the
+    # input box moves with the output and repaints slowly. Set false only if
+    # your terminal mishandles Textual's mouse reporting.
+    tui_mouse: bool = True
     privacy_mode: bool = (
         False  # When True, all data stays local, no cloud capabilities
     )
@@ -314,7 +321,7 @@ def api_key(self, value: str) -> None:
     def load(
         cls,
         default_provider: str = "tongyi",
-        default_model: str = "qwen3.7-plus",
+        default_model: str = "qwen3.8-max",
     ) -> Config:
         config = cls()
         # Caller's defaults become the initial values; config files and env
@@ -542,6 +549,20 @@ def _load_workspace_from(self, path: Path):
                 self.max_turns = int(data["max_turns"])
             except (ValueError, TypeError):
                 pass
+        if "auto_approve" in data:
+            val = str(data["auto_approve"]).lower()
+            self.auto_approve = val in ("true", "1", "yes")
+        if "confirm_mode" in data:
+            val = str(data["confirm_mode"]).lower()
+            if val in ("all", "dangerous"):
+                self.confirm_mode = val
+        # Env override for headless/benchmark use
+        if os.environ.get("ACLI_AUTO_APPROVE", "").lower() in (
+            "1",
+            "true",
+            "yes",
+        ):
+            self.auto_approve = True
         if "session_persist" in data:
             val = str(data["session_persist"]).lower()
             self.session_persist = val not in ("false", "0", "no")

dashscope/acli/examples/basic-chat/.acli/config.toml

@@ -1,4 +1,4 @@
 user_name = "dashscope"
 provider = "tongyi"
-model = "qwen3.7-max"
+model = "qwen3.8-max"
 memory_user_id = "acli-basic"

dashscope/acli/examples/basic-chat/.acli/custom-extensions.toml

@@ -19,8 +19,9 @@
 name = "tongyi"
 base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 api_key_env = "DASHSCOPE_API_KEY"
-default_model = "qwen3.7-max"
+default_model = "qwen3.8-max"
 models = [
+    "qwen3.8-max",
     "qwen3.7-max",
     "qwen3.7-plus",
     "qwen-turbo",

dashscope/acli/examples/basic-chat/README.md

@@ -51,8 +51,8 @@ Declares which LLM providers acli can use. A minimal config needs just one `[[pr
 name = "tongyi"
 base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 api_key_env = "DASHSCOPE_API_KEY"      # ← stores only the env var name; the shell provides sk-xxx
-default_model = "qwen3.7-max"
-models = ["qwen3.7-max", "qwen3.7-plus", "qwen-turbo", "qwen-vl-max"]
+default_model = "qwen3.8-max"
+models = ["qwen3.8-max", "qwen3.7-max", "qwen3.7-plus", "qwen-turbo", "qwen-vl-max"]
 vision_models = ["qwen-vl-max"]        # ← tells acli these models accept image input
 protocol = "openai"                     # ← openai / anthropic / dashscope
 ```
@@ -95,7 +95,7 @@ How to invoke:
 ```toml
 user_name = "dashscope"
 provider = "tongyi"
-model = "qwen3.7-max"
+model = "qwen3.8-max"
 memory_user_id = "acli-basic"
 ```
 

dashscope/acli/examples/dashscope-sdk-expert/.acli/custom-extensions.toml

@@ -10,7 +10,7 @@
 name = "tongyi"
 base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 api_key_env = "DASHSCOPE_API_KEY"
-default_model = "qwen3.7-plus"
+default_model = "qwen3.8-max"
 models = [
     "qwen3.8-max",
     "qwen3.7-max",

dashscope/acli/examples/dashscope-sdk-expert/README.md

@@ -86,7 +86,7 @@ Before generating code, first verify the user's installed SDK version and API si
 ```toml
 user_name = "dashscope"
 provider = "tongyi"
-model = "qwen3.7-plus"
+model = "qwen3.8-max"
 memory_user_id = "acli-dashscope"
 ```
 

dashscope/acli/executor.py

@@ -35,8 +35,15 @@
 
 
 class Executor:
-    def __init__(self, auto_approve: bool = False):
+    def __init__(
+        self,
+        auto_approve: bool = False,
+        confirm_mode: str = "dangerous",
+    ):
         self.auto_approve = auto_approve
+        # "all": CONFIRM and DANGEROUS both prompt; "dangerous": only
+        # DANGEROUS prompts (CONFIRM-level tools auto-pass).
+        self.confirm_mode = confirm_mode
         # Trust cache scoped to ONE conversation turn (a single user prompt
         # plus the agent loop that answers it). Agent.run / run_stream clear
         # these in a finally block on completion or abort. DANGEROUS tools
@@ -214,6 +221,14 @@ async def _async_check_permission(
                     and policy.check_command(cmd) == "deny"
                 ):
                     return False
+            # confirm_mode="dangerous": only risky (DANGEROUS) operations
+            # prompt; CONFIRM-level tools auto-pass. Policy deny rules
+            # above have already been honored.
+            if (
+                self.confirm_mode == "dangerous"
+                and tool_def.permission == PermissionLevel.CONFIRM
+            ):
+                return True
             # Auto-pass read-only shell commands
             if tool_def.name == "run_command":
                 cmd = arguments.get("command", "")
@@ -284,6 +299,15 @@ def _check_permission(
                 console.print("[dim red]✗ command denied by policy[/dim red]")
                 return False
 
+        # confirm_mode="dangerous": only risky (DANGEROUS) operations
+        # prompt; CONFIRM-level tools auto-pass. Policy deny rules
+        # above have already been honored.
+        if (
+            self.confirm_mode == "dangerous"
+            and tool_def.permission == PermissionLevel.CONFIRM
+        ):
+            return True
+
         # Auto-pass read-only shell commands (grep, ls, find, git status, …).
         # The classifier lives in shell.py since it owns the shell semantics;
         # see is_safe_readonly for the allow/deny rules. Anything not

dashscope/acli/memory/manager.py

@@ -20,7 +20,10 @@
 
 from dashscope.acli.memory.experience import ExperienceTracker
 from dashscope.acli.memory.plan import PlanTracker
-from dashscope.acli.memory.reflection import ReflectionTracker
+from dashscope.acli.memory.reflection import (
+    ReflectionTracker,
+    StagnationTracker,
+)
 from dashscope.acli.memory.tool_chains import ToolChainLibrary
 from dashscope.acli.memory.trace import TraceLogger
 
@@ -38,12 +41,14 @@ def __init__(self, workspace_dir: Path):
 
         self.plan = PlanTracker()
         self.reflection = ReflectionTracker()
+        self.stagnation = StagnationTracker()
         self.tool_chains = ToolChainLibrary(self.session_dir)
 
     def reset(self) -> None:
         """Reset session memory (called at session start)."""
         self.plan.clear_plan()
         self.reflection.reset()
+        self.stagnation.reset()
 
 
 class PersistentMemory: