feat(btw): select models independently for each loop

astrbot/core/astr_main_agent.py

@@ -186,6 +186,8 @@ class MainAgentBuildConfig:
     add_cron_tools: bool = True
     """This will add cron job management tools to the main agent for proactive cron job execution."""
     provider_settings: dict = field(default_factory=dict)
+    provider_id_override: str = ""
+    """Optional request-scoped chat provider override."""
     fallback_provider_ids: list[str] = field(default_factory=list)
     request_max_retries: int = 5
     subagent_orchestrator: dict = field(default_factory=dict)
@@ -317,10 +319,12 @@ def _set_llm_error_message(event: AstrMessageEvent, message: str) -> None:
 
 
 def _select_provider(
-    event: AstrMessageEvent, plugin_context: CoreExecutionContext
+    event: AstrMessageEvent,
+    plugin_context: CoreExecutionContext,
+    provider_id_override: str = "",
 ) -> ChatModel | None:
     """Select chat provider for the event."""
-    sel_provider = event.get_extra("selected_provider")
+    sel_provider = provider_id_override or event.get_extra("selected_provider")
     if sel_provider and isinstance(sel_provider, str):
         provider = plugin_context.get_provider_by_id(sel_provider)
         if provider is None:
@@ -1952,7 +1956,9 @@ async def build_main_agent(
 
     If apply_reset is False, will not call reset on the agent runner.
     """
-    provider = provider or _select_provider(event, plugin_context)
+    provider = provider or _select_provider(
+        event, plugin_context, config.provider_id_override
+    )
     if provider is None:
         logger.info("未找到任何对话模型（提供商），跳过 LLM 请求处理。")
         if not event.get_extra(LLM_ERROR_MESSAGE_EXTRA_KEY):

astrbot/core/config/default.py

@@ -190,7 +190,8 @@
     },
     "btw": {
         "enabled": False,
-        "work_loop": {"enabled": False, "max_concurrent": 2},
+        "conversation_loop": {"provider_id": ""},
+        "work_loop": {"enabled": False, "provider_id": "", "max_concurrent": 2},
         "work_session": {"max_age_seconds": 3600},
     },
     "provider_stt_settings": {
@@ -4699,12 +4700,26 @@
             "type": "bool",
             "hint": "实验功能，默认关闭。开启后，普通 AI 请求通过对话循环进入现有 Agent。",
         },
+        "btw.conversation_loop.provider_id": {
+            "description": "对话循环模型",
+            "type": "string",
+            "_special": "select_provider",
+            "hint": "留空时沿用当前会话的模型选择。配置后优先使用此模型。",
+            "condition": {"btw.enabled": True},
+        },
         "btw.work_loop.enabled": {
             "description": "启用工作循环",
             "type": "bool",
             "hint": "默认关闭；允许显式工作请求使用工作执行器。",
             "condition": {"btw.enabled": True},
         },
+        "btw.work_loop.provider_id": {
+            "description": "工作循环模型",
+            "type": "string",
+            "_special": "select_provider",
+            "hint": "留空时沿用当前会话的模型选择。配置后优先使用此模型。",
+            "condition": {"btw.enabled": True},
+        },
         "btw.work_loop.max_concurrent": {
             "description": "工作任务执行并发",
             "type": "int",

astrbot/core/pipeline/process_stage/method/agent_sub_stages/internal.py

@@ -268,9 +268,19 @@ async def _build_checked_agent_runner(
         streaming_response: bool,
     ) -> MainAgentBuildResult | None:
         """Build a runner and reject configured provider endpoints unsafe for use."""
+        btw = self._profile_config(event).get("btw", {})
+        provider_id_override = ""
+        if isinstance(btw, dict) and btw.get("enabled", False):
+            loop = "work" if event.get_extra("btw_loop") == "work" else "conversation"
+            loop_config = btw.get(f"{loop}_loop", {})
+            if isinstance(loop_config, dict):
+                provider_id = loop_config.get("provider_id", "")
+                if isinstance(provider_id, str):
+                    provider_id_override = provider_id.strip()
         build_cfg = replace(
             self.main_agent_cfg,
             streaming_response=streaming_response,
+            provider_id_override=provider_id_override,
         )
         build_result = await build_main_agent(
             event=event,

dashboard/src/i18n/locales/en-US/features/config-metadata.json

@@ -1190,13 +1190,23 @@
           "max_concurrent": {
             "description": "Concurrent work execution",
             "hint": "Active execution limit, default 2. This is not a waiting-queue length limit."
+          },
+          "provider_id": {
+            "description": "Work loop model",
+            "hint": "Leave empty to keep the current session model selection. When set, this model takes priority."
           }
         },
         "work_session": {
           "max_age_seconds": {
             "description": "Terminal work retention (seconds)",
             "hint": "Keep completed, failed or cancelled records for 3600 seconds by default."
           }
+        },
+        "conversation_loop": {
+          "provider_id": {
+            "description": "Conversation loop model",
+            "hint": "Leave empty to keep the current session model selection. When set, this model takes priority."
+          }
         }
       }
     }

dashboard/src/i18n/locales/zh-CN/features/config-metadata.json

@@ -1184,13 +1184,23 @@
           "max_concurrent": {
             "description": "工作任务执行并发",
             "hint": "同时执行的工作任务数，默认 2。此值不是等待队列的长度限制。"
+          },
+          "provider_id": {
+            "description": "工作循环模型",
+            "hint": "留空时沿用当前会话的模型选择。配置后优先使用此模型。"
           }
         },
         "work_session": {
           "max_age_seconds": {
             "description": "终态工作会话保留秒数",
             "hint": "已完成、失败或取消的工作会话保留时间，默认 3600 秒。"
           }
+        },
+        "conversation_loop": {
+          "provider_id": {
+            "description": "对话循环模型",
+            "hint": "留空时沿用当前会话的模型选择。配置后优先使用此模型。"
+          }
         }
       }
     }

docs/en/dev/astrbot-config.md

@@ -187,6 +187,12 @@ Local mode operates directly on the AstrBot host and belongs only in a trusted e
 
 `image_compress_enabled` and `image_compress_options.max_size/quality` control image handling in the request-preparation choke point `prepare_provider_request`. The main-agent chat path, SDK `llm_generate`, and `tool_loop_agent` share that step. Provider-bound images are converted to JPEG there; the long edge is downscaled only and never upscaled. Animated GIF/WebP sources are dhash-sampled, at most 8 frames. Disabling compression still converts to JPEG without resizing. The main agent only materializes adapter refs to local paths and does not pre-encode chat attachments to JPEG. `max_quoted_fallback_images` and `quoted_message_parser` limit quoted and forwarded-message expansion to prevent unbounded fetching. For `quoted_message_parser`, `0` is a valid boundary: depth limits keep the root level but stop child recursion, and `max_forward_fetch=0` disables recursive `get_forward_msg` calls. Negative or invalid values fall back to defaults; this setting does not globally disable a direct quoted-message `get_msg` fallback.
 
+## BTW model selection
+
+When `btw.enabled` is enabled for a local Agent profile, `btw.conversation_loop.provider_id` and `btw.work_loop.provider_id` select the chat model for each loop. A configured loop model takes priority over the event/session model selection. An empty field preserves the current selection, including the profile default. Messages without an explicit work-loop marker use the conversation model. Disabling BTW ignores both overrides.
+
+The selected provider must still be a configured chat model. An unavailable or incompatible loop provider fails through the existing model-selection error path; it does not silently switch to the other loop's model. Existing model fallback and retry settings continue to apply to the selected primary provider.
+
 ## SubAgents, speech, and knowledge base
 
 - `subagent_orchestrator.main_enable` enables handoffs.

docs/zh/dev/astrbot-config.md

@@ -189,6 +189,12 @@ API Key 属于敏感配置。不要把真实 `cmd_config.json`、截图、日志
 
 `image_compress_enabled` 和 `image_compress_options.max_size/quality` 控制请求准备卡口 `prepare_provider_request` 中的图片处理，主智能体聊天路径、SDK `llm_generate` 与 `tool_loop_agent` 共用该卡口。送给模型的图片在此处转为 JPEG，最长边只缩小、从不放大；动画 GIF/WebP 会按 dhash 抽帧，最多 8 帧。关闭压缩时仍会转 JPEG，但不缩放。主智能体只把适配器引用物化为本地路径，不在组装附件时预编码 JPEG。`max_quoted_fallback_images` 与 `quoted_message_parser` 限制引用消息和转发消息展开深度，避免无限抓取。对 `quoted_message_parser` 而言，`0` 是有效边界：深度限制会保留根层并停止子层递归，`max_forward_fetch=0` 会禁止递归调用 `get_forward_msg`。负数或无效值会回退为默认值；该设置不会全局禁止引用消息回退路径中的直接 `get_msg` 调用。
 
+## BTW 模型选择
+
+本地 Agent 配置启用 `btw.enabled` 后，`btw.conversation_loop.provider_id` 与 `btw.work_loop.provider_id` 分别选择两个循环的对话模型。已配置的循环模型优先于事件或会话的模型选择；留空则沿用当前选择，包括配置档默认模型。没有显式工作循环标记的消息使用对话循环模型。关闭 BTW 后不应用这两个覆盖项。
+
+所选提供商仍须是已配置的对话模型。不存在或类型不适用的循环提供商沿用现有模型选择错误路径，不会静默改用另一个循环的模型。已有模型回退和重试设置继续作用于所选主模型。
+
 ## 子代理、语音与知识库
 
 - `subagent_orchestrator.main_enable`：启用 handoff。

tests/unit/test_agent_internal_process.py

@@ -2,10 +2,57 @@
 
 import pytest
 
+from astrbot.core.astr_main_agent import MainAgentBuildConfig
 from astrbot.core.message.components import Json
 from tests.unit.agent_sub_stage_support import *  # noqa: F403
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("enabled", "loop", "conversation_provider", "work_provider", "expected"),
+    [
+        (True, None, "conversation-model", "work-model", "conversation-model"),
+        (
+            True,
+            "conversation",
+            "conversation-model",
+            "work-model",
+            "conversation-model",
+        ),
+        (True, "work", "conversation-model", "work-model", "work-model"),
+        (True, "invalid", "conversation-model", "work-model", "conversation-model"),
+        (True, "work", "conversation-model", "", ""),
+        (True, None, None, "work-model", ""),
+        (False, "work", "conversation-model", "work-model", ""),
+    ],
+)
+async def test_btw_loop_provider_selection_is_request_scoped(
+    monkeypatch, enabled, loop, conversation_provider, work_provider, expected
+):
+    stage = internal.InternalAgentSubStage.__new__(internal.InternalAgentSubStage)
+    stage.ctx = _pipeline_context(_internal_plugin_context())
+    stage.ctx.astrbot_config = {
+        "btw": {
+            "enabled": enabled,
+            "conversation_loop": {"provider_id": conversation_provider},
+            "work_loop": {"provider_id": work_provider},
+        }
+    }
+    stage.main_agent_cfg = MainAgentBuildConfig(tool_call_timeout=60)
+    result = SimpleNamespace(provider=SimpleNamespace(provider_config={}))
+    build = AsyncMock(return_value=result)
+    monkeypatch.setattr(internal, "build_main_agent", build)
+    event = FakeEvent(extras={"btw_loop": loop, "selected_provider": "session-model"})
+
+    assert await stage._build_checked_agent_runner(event, False) is result
+
+    config = build.await_args.kwargs["config"]
+    assert config.provider_id_override == expected
+    assert config.streaming_response is False
+    assert stage.main_agent_cfg.provider_id_override == ""
+    assert event.get_extra("selected_provider") == "session-model"
+
+
 @pytest.mark.asyncio
 async def test_internal_process_skips_empty_messages_without_provider_request(
     monkeypatch,

tests/unit/test_astr_main_agent.py

@@ -531,6 +531,31 @@ def test_config_with_custom_values(self):
 class TestSelectProvider:
     """Tests for _select_provider function."""
 
+    def test_loop_override_takes_priority_without_changing_event(
+        self, mock_event, mock_context, mock_provider
+    ):
+        mock_event.set_extra("selected_provider", "session-model")
+        mock_context.get_provider_by_id.return_value = mock_provider
+
+        assert (
+            ama._select_provider(mock_event, mock_context, "loop-model")
+            is mock_provider
+        )
+        mock_context.get_provider_by_id.assert_called_once_with("loop-model")
+        mock_context.get_using_provider.assert_not_called()
+        assert mock_event.get_extra("selected_provider") == "session-model"
+
+    @pytest.mark.parametrize("provider", [None, "not-a-chat-provider"])
+    def test_invalid_loop_override_does_not_fall_back(
+        self, mock_event, mock_context, provider
+    ):
+        mock_event.set_extra("selected_provider", "session-model")
+        mock_context.get_provider_by_id.return_value = provider
+
+        assert ama._select_provider(mock_event, mock_context, "loop-model") is None
+        assert mock_event.get_extra(ama.LLM_ERROR_MESSAGE_EXTRA_KEY)
+        mock_context.get_using_provider.assert_not_called()
+
     def test_select_provider_by_id(self, mock_event, mock_context, mock_provider):
         """Test selecting provider by ID from event extra."""
         module = ama

tests/unit/test_config_metadata_i18n.py

@@ -10,6 +10,7 @@
     CONFIG_METADATA_2,
     CONFIG_METADATA_3,
     CONFIG_METADATA_3_SYSTEM,
+    DEFAULT_CONFIG,
 )
 from astrbot.core.config.i18n_utils import ConfigMetadataI18n
 from astrbot.core.platform.sources.line.line_adapter import (
@@ -134,6 +135,13 @@ def test_config_metadata_locale_trees_match() -> None:
     assert sorted(en_keys - zh_keys) == []
 
 
+def test_every_btw_profile_field_reaches_dashboard_controls() -> None:
+    converted = ConfigMetadataI18n.convert_to_i18n_keys(CONFIG_METADATA_3)
+    items = converted["plugin_group"]["metadata"]["btw"]["items"]
+    expected_fields = {f"btw.{field}" for field in _flatten(DEFAULT_CONFIG["btw"])}
+    assert set(items) == expected_fields
+
+
 def test_btw_controls_survive_dashboard_metadata_conversion() -> None:
     converted = ConfigMetadataI18n.convert_to_i18n_keys(CONFIG_METADATA_3)
     section = converted["plugin_group"]["metadata"]["btw"]