dsh-reasoning-distill 0.1.7 → 0.1.9

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 本插件的版本变更记录。
 
+## 0.1.9
+
+- README 开头补充「适用场景」说明：本插件只在单轮次、多步工具调用的超长单轮对话里划算；多轮问答中每轮步骤少、压缩收益有限，且 DeepSeek 的 API 默认不会参考之前轮次的 `reasoning_content`（仅请求携带 tools 时拼接进上下文），改写还会让缓存前缀失效、后续请求按未命中计费，整体往往难以抵消。插件面向「说完任务就不管」的许愿式编程，在根据用户反馈反复精修的阶段建议关闭。
+- 本次没有代码改动。
+
+## 0.1.8
+
+- **调整初始配置**：刚安装时默认规则里的「对可续接子代理启用」「附带工具调用结果」「压缩最终回答轮」现在默认关闭。新建的逐模型规则同样以一个全关的初始层开始（三种会话类型全关，上述三项也显式关闭），不会因为现有默认规则把它们开着而自动继承为开。
+- 面板提示与 README 的默认值说明同步更新；已有存储配置不迁移。
+
 ## 0.1.7
 
 - **调整内置蒸馏提示词**：明确要求保留原思考中正确且有效的推理过程，而不是直接跳到结论。可以压缩措辞、合并重复、删除噪音，但不能省略有效的推导步骤；只保留结论的短版本视为错误。自我纠正时保留纠正后的整条推理线（结论、证据与有效步骤），只丢弃被放弃的尝试。

package/README.md

@@ -2,6 +2,14 @@
 
 把过长的模型思考压缩成核心内容，之后每次请求都重放压缩版，省下重复发送的 token；同时给出每个会话的开销与节省报告。
 
+> **⚠️ 适用场景**
+>
+> 本插件只在 **「单轮次、多步工具调用的超长单轮对话」** 里划算。模型连续调用很多步工具时，同一段历史在每一步都要重新发送，压缩后的思考会在同一轮里被反复重放，省下的输入 token 才能累积到超过蒸馏开销。
+>
+> **「多轮问答」** 中每轮通常只有少量步骤，压缩省下的 token 有限。更关键的是，**DeepSeek 的 API 在多轮对话中不会参考之前轮次的 `reasoning_content`**（只有请求携带 tools 时才会拼接进上下文），压缩后的思考通常不会被模型读到；改写却会改变历史内容，而上下文缓存按前缀精确匹配，改写点之后的内容会变成未命中输入，再加上每次蒸馏的费用，整体往往难以抵消。
+>
+> 可以说这个插件的设计目标就是 **「说完任务就什么都不管了」的「许愿式编程」** ，也就是一次说清任务、让 agent 自行跑完很长的工具调用链。在 **「根据用户反馈反复精修」** 的阶段，对话形态是多轮问答，**建议关闭插件**（关掉设置里的总开关即可）。
+
 > 插件**默认关闭**，装好后需要在「设置 → 思考压缩」里选中一个改写后端并手动打开总开关。
 
 > **版本与安装方式（先看这里）**
@@ -154,14 +162,14 @@ node scripts/dsh-inplace-patch.mjs --check          # 打全了输出 PATCHED
 | 选项 | 含义 |
 | --- | --- |
 | 对主对话启用（`enabled`） | 该模型在主对话里是否压缩。 |
-| 对可续接子代理启用（`continuableSubagents`） | 是否也压缩可续聊子代理（如队友）。默认开，因为它们会持续重发思考。 |
+| 对可续接子代理启用（`continuableSubagents`） | 是否也压缩可续聊子代理（如队友）。默认关；需要时再打开。 |
 | 对一次性子代理启用（`oneShotSubagents`） | 是否压缩跑完即结束的一次性委派。默认关，因为它的最终回答之后没有任何请求会重放这次压缩。 |
 | 最大生效深度（`continuableMaxDepth` / `oneShotMaxDepth`） | 两个子代理开关各自带一个深度上限：比该深度更深的子孙代理不再压缩，0 表示不限制。 |
 | 触发阈值（`thresholdTokens`） | 一步思考估算超过多少 token 才触发，默认 4000。 |
 | 按上下文占用限制触发（`contextCap` / `maxContextPercent`） | 默认开、60%。该模型的上下文占用超过这个比例时，即使单步思考达到阈值也不再压缩——否则压缩刚生效、没几轮就被宿主内置的上下文整体压缩收走，白花一次蒸馏的钱。占用比例用宿主自己的 token 计量与上下文窗口计算，和内置上下文压缩触发读的是同一个值，整段压缩后会自动变短。关闭时不看占用比例（与旧行为一致），比例输入框会变灰。 |
 | 最低节省（`minSavedTokens`） | 压缩结果至少要比原文省这么多 token 才采用，默认 600；未达到时原文保持不变，但这次蒸馏的开销照样计入。 |
-| 附带工具调用结果（`includeToolResults`） | 交给蒸馏模型的上下文里是否附上该步工具调用的结果；工具调用本身始终附带。 |
-| 压缩最终回答轮（`includeFinalTurns`） | 是否压缩没有工具调用的收尾步（最终回答）。 |
+| 附带工具调用结果（`includeToolResults`） | 交给蒸馏模型的上下文里是否附上该步工具调用的结果；工具调用本身始终附带。默认关（刚安装的默认规则与新建的逐模型规则都是关的）。 |
+| 压缩最终回答轮（`includeFinalTurns`） | 是否压缩没有工具调用的收尾步（最终回答）。默认关（同上）。 |
 | 超时（`timeoutMs`） | 蒸馏必须在这个时间内完成；到点就放弃这次压缩（原文不动、不记账），下一步只等这么久。0 = 不设限，一直等到出结果。 |
 | 背景块长度上限（`maxInputChars`） | 交给蒸馏模型的单个上下文块（例如一条工具结果）的字符上限，0 = 不限。 |
 | 蒸馏结果长度上限（`maxOutputChars`） | 实际采用并重放的压缩结果字符上限，0 = 不限。 |

package/lib/client/index.js

@@ -341,7 +341,7 @@ window.__ModuleLoader__.load({
 			"settings.maxDepth": "最大生效深度",
 			"settings.maxDepthHint": "两个子代理开关各自带一个深度上限：超过该深度的子孙代理不再压缩（填 0 = 不限制）。深度由宿主持久化在会话 header 上，重启后续接仍按同一深度判定。",
 			"settings.subagentsLocked": "当前后端（另开分支会话）无法承接委派：交接会打断子代理的委派关系，因此这两条子代理开关已被关闭并锁定。换成不做分支的后端即可重新开启。",
-			"settings.kindsHint": "三种会话各自独立开关；宿主用 header.origin 与 subagent/descriptor 的 mode 区分它们。可续接子代理（如队友）会持续重发思考，默认开；一次性委派跑完即返回，最终回答的压缩没有任何请求会重放它，默认关。",
+			"settings.kindsHint": "三种会话各自独立开关；宿主用 header.origin 与 subagent/descriptor 的 mode 区分它们。可续接子代理（如队友）会持续重发思考，需要时再打开，默认关；一次性委派跑完即返回，最终回答的压缩没有任何请求会重放它，默认也关。",
 			"settings.ruleIdle": "三种会话都未启用时，该规则的其他设置不会生效（届时会锁定）。",
 			"settings.master": "启用思考压缩",
 			"settings.default": "默认规则",
@@ -560,7 +560,7 @@ window.__ModuleLoader__.load({
 			"settings.maxDepth": "Max depth",
 			"settings.maxDepthHint": "Each subagent switch carries its own depth cap: children nested deeper than this are left alone (0 = no limit). The depth is persisted on the session header, so a resumed child is judged at the same depth.",
 			"settings.subagentsLocked": "The selected backend forks the conversation to apply a rewrite, which a delegation cannot follow: a hand-off inside a subagent would end its turn without an answer, so both subagent switches are pinned off. Select a backend that does not branch to enable them again.",
-			"settings.kindsHint": "The three session kinds have independent switches; the host separates them by header.origin and the subagent/descriptor mode. Continuable children (teammates) keep replaying their reasoning, so they are on by default; a terminal one-shot delegation is never resumed, so its final answer pays for a distillation nothing replays, and it is off by default.",
+			"settings.kindsHint": "The three session kinds have independent switches; the host separates them by header.origin and the subagent/descriptor mode. Continuable children (teammates) keep replaying their reasoning, so turn them on when needed (off by default); a terminal one-shot delegation is never resumed, so its final answer pays for a distillation nothing replays, and it is off by default.",
 			"settings.ruleIdle": "With all three session kinds off, the rest of this rule cannot take effect and is locked.",
 			"settings.default": "Default rule",
 			"settings.models": "Per-model rules",
@@ -759,7 +759,7 @@ window.__ModuleLoader__.load({
 		*/
 		const FALLBACK_RULE = {
 			enabled: true,
-			continuableSubagents: true,
+			continuableSubagents: false,
 			continuableMaxDepth: 0,
 			oneShotSubagents: false,
 			oneShotMaxDepth: 0,
@@ -767,8 +767,8 @@ window.__ModuleLoader__.load({
 			contextCap: true,
 			maxContextPercent: 60,
 			minSavedTokens: 600,
-			includeToolResults: true,
-			includeFinalTurns: true,
+			includeToolResults: false,
+			includeFinalTurns: false,
 			timeoutMs: 0,
 			maxInputChars: 0,
 			maxOutputChars: 0,
@@ -806,6 +806,22 @@ window.__ModuleLoader__.load({
 		//#endregion
 		//#region src/client/ops.ts
 		/**
+		* Initial stored layer for a model the user just added.
+		*
+		* The entry starts inert: no session kind is enabled, and the optional extras —
+		* subagent distillation, attaching tool-call results, and compressing final
+		* answers — are stated off rather than inherited, so adding a model can never
+		* start paying for work the user did not ask for. Every field stays editable and
+		* the remaining fields still inherit from the default rule.
+		*/
+		const NEW_MODEL_RULE = {
+			enabled: false,
+			continuableSubagents: false,
+			oneShotSubagents: false,
+			includeToolResults: false,
+			includeFinalTurns: false
+		};
+		/**
 		* Whether a rule decides nothing because all three session kinds are off, in
 		* which case the rest of its fields are locked in the panel.
 		* @param rule - the rule being edited.
@@ -1343,49 +1359,49 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var Settings_module_css_default = {
-			"secondary": "jP3NOa_secondary",
-			"backendOptions": "jP3NOa_backendOptions",
-			"installProgress": "jP3NOa_installProgress",
-			"itemBody": "jP3NOa_itemBody",
-			"preview": "jP3NOa_preview",
-			"modalHead": "jP3NOa_modalHead",
-			"installBlock": "jP3NOa_installBlock",
-			"priceGrid": "jP3NOa_priceGrid",
-			"muted": "jP3NOa_muted",
-			"error": "jP3NOa_error",
-			"modal": "jP3NOa_modal",
-			"modalBody": "jP3NOa_modalBody",
-			"steps": "jP3NOa_steps",
 			"logBox": "jP3NOa_logBox",
-			"search": "jP3NOa_search",
+			"modalHead": "jP3NOa_modalHead",
+			"holidayBox": "jP3NOa_holidayBox",
 			"pickerItem": "jP3NOa_pickerItem",
-			"log": "jP3NOa_log",
-			"pickerPanel": "jP3NOa_pickerPanel",
+			"section": "jP3NOa_section",
 			"iconButton": "jP3NOa_iconButton",
-			"rule": "jP3NOa_rule",
-			"holidayBox": "jP3NOa_holidayBox",
+			"peakBox": "jP3NOa_peakBox",
+			"picker": "jP3NOa_picker",
+			"muted": "jP3NOa_muted",
+			"modalBackdrop": "jP3NOa_modalBackdrop",
 			"checks": "jP3NOa_checks",
+			"installProgress": "jP3NOa_installProgress",
+			"itemBody": "jP3NOa_itemBody",
+			"search": "jP3NOa_search",
+			"error": "jP3NOa_error",
 			"capBox": "jP3NOa_capBox",
-			"field": "jP3NOa_field",
 			"item": "jP3NOa_item",
 			"primary": "jP3NOa_primary",
-			"installEntry": "jP3NOa_installEntry",
-			"prompt": "jP3NOa_prompt",
+			"secondary": "jP3NOa_secondary",
+			"header": "jP3NOa_header",
+			"installBlock": "jP3NOa_installBlock",
 			"toggle": "jP3NOa_toggle",
-			"sectionHeading": "jP3NOa_sectionHeading",
-			"modalBackdrop": "jP3NOa_modalBackdrop",
 			"modeCard": "jP3NOa_modeCard",
-			"picker": "jP3NOa_picker",
+			"log": "jP3NOa_log",
+			"installEntry": "jP3NOa_installEntry",
+			"actions": "jP3NOa_actions",
+			"pickerPanel": "jP3NOa_pickerPanel",
 			"badge": "jP3NOa_badge",
-			"page": "jP3NOa_page",
+			"sectionHeading": "jP3NOa_sectionHeading",
+			"priceGrid": "jP3NOa_priceGrid",
+			"field": "jP3NOa_field",
+			"preview": "jP3NOa_preview",
+			"backendOptions": "jP3NOa_backendOptions",
+			"rule": "jP3NOa_rule",
+			"modal": "jP3NOa_modal",
 			"pickerList": "jP3NOa_pickerList",
-			"header": "jP3NOa_header",
-			"actions": "jP3NOa_actions",
-			"peakBox": "jP3NOa_peakBox",
+			"modalBody": "jP3NOa_modalBody",
+			"steps": "jP3NOa_steps",
+			"note": "jP3NOa_note",
 			"modeCards": "jP3NOa_modeCards",
+			"prompt": "jP3NOa_prompt",
 			"route": "jP3NOa_route",
-			"section": "jP3NOa_section",
-			"note": "jP3NOa_note"
+			"page": "jP3NOa_page"
 		};
 		//#endregion
 		//#region src/client/InstallBackend.tsx
@@ -2559,11 +2575,7 @@ window.__ModuleLoader__.load({
 									...current,
 									models: {
 										...current.models,
-										[model.id]: {
-											enabled: false,
-											continuableSubagents: false,
-											oneShotSubagents: false
-										}
+										[model.id]: { ...NEW_MODEL_RULE }
 									}
 								}))
 							}),
@@ -2822,41 +2834,41 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var View_module_css_default = {
+			"facts": "_7CZWHG_facts",
+			"popupHead": "_7CZWHG_popupHead",
+			"header": "_7CZWHG_header",
 			"entry": "_7CZWHG_entry",
+			"actions": "_7CZWHG_actions",
+			"blockedBanner": "_7CZWHG_blockedBanner",
+			"popupFacts": "_7CZWHG_popupFacts",
 			"metric": "_7CZWHG_metric",
+			"iconButton": "_7CZWHG_iconButton",
+			"activityDotBlocked": "_7CZWHG_activityDotBlocked",
+			"secondary": "_7CZWHG_secondary",
+			"note": "_7CZWHG_note",
+			"popupList": "_7CZWHG_popupList",
+			"summary": "_7CZWHG_summary",
 			"activityDot": "_7CZWHG_activityDot",
-			"popupActions": "_7CZWHG_popupActions",
-			"activeRow": "_7CZWHG_activeRow",
+			"sessionInfo": "_7CZWHG_sessionInfo",
 			"activityLabel": "_7CZWHG_activityLabel",
-			"facts": "_7CZWHG_facts",
+			"activityWhy": "_7CZWHG_activityWhy",
 			"error": "_7CZWHG_error",
-			"entryBody": "_7CZWHG_entryBody",
+			"positive": "_7CZWHG_positive",
 			"section": "_7CZWHG_section",
-			"negative": "_7CZWHG_negative",
-			"popupList": "_7CZWHG_popupList",
-			"sessionInfo": "_7CZWHG_sessionInfo",
 			"rdPulse": "_7CZWHG_rdPulse",
-			"activityWhy": "_7CZWHG_activityWhy",
-			"popupRow": "_7CZWHG_popupRow",
-			"positive": "_7CZWHG_positive",
+			"entryBody": "_7CZWHG_entryBody",
+			"muted": "_7CZWHG_muted",
+			"page": "_7CZWHG_page",
 			"sessionRow": "_7CZWHG_sessionRow",
-			"summary": "_7CZWHG_summary",
+			"negative": "_7CZWHG_negative",
 			"headerTrigger": "_7CZWHG_headerTrigger",
-			"popupFacts": "_7CZWHG_popupFacts",
-			"iconButton": "_7CZWHG_iconButton",
-			"muted": "_7CZWHG_muted",
-			"activity": "_7CZWHG_activity",
-			"secondary": "_7CZWHG_secondary",
-			"popupHead": "_7CZWHG_popupHead",
-			"note": "_7CZWHG_note",
-			"actions": "_7CZWHG_actions",
-			"blockedBanner": "_7CZWHG_blockedBanner",
-			"entryTitle": "_7CZWHG_entryTitle",
-			"header": "_7CZWHG_header",
-			"textBlock": "_7CZWHG_textBlock",
 			"popup": "_7CZWHG_popup",
-			"activityDotBlocked": "_7CZWHG_activityDotBlocked",
-			"page": "_7CZWHG_page"
+			"popupRow": "_7CZWHG_popupRow",
+			"popupActions": "_7CZWHG_popupActions",
+			"textBlock": "_7CZWHG_textBlock",
+			"entryTitle": "_7CZWHG_entryTitle",
+			"activity": "_7CZWHG_activity",
+			"activeRow": "_7CZWHG_activeRow"
 		};
 		//#endregion
 		//#region src/client/ActivityChip.tsx