dsh-llm-verifier 0.6.9 → 0.8.0-beta.0

package/README.md

@@ -89,7 +89,16 @@ pnpm test           # vitest 单元测试
 pnpm run typecheck:local   # 依据 tsconfig.local.json，把 @deepseek-ai/dsh-* 解析到本地 checkout 的 lib/types
 ```
 
-两套类型定义可能不同步：例如 `Session.events` 在 DSH 0.1.5 已被 `snapshotEvents()` 取代，`tool/code-dispatch` 也已改名 `tool/ptc-dispatch`。插件内部的 `sessionEvents()` 同时兼容两种会话形态，两类派发事件都会计入证据，因此 0.1.1 与 0.1.5 宿主都可以运行。
+两套类型定义可能不同步：例如 `Session.events` 在 DSH 0.1.5 已被 `snapshotEvents()` 取代，`tool/code-dispatch` 也已改名 `tool/ptc-dispatch`。插件内部的 `sessionEvents()` 同时兼容两种会话形态，两类派发事件都会计入证据，因此 0.1.1 与 0.1.7 宿主都可以运行。
+
+DSH 0.1.7 又搬动了两处判定地基，两侧形态插件都读：
+
+- **工具结果从内容块变成独立消息**。0.1.6 及以前，一条工具结果是一个 `tool-result` 内容块，真正的输出套在里面、失败标志 `isError` 挂在块上；0.1.7 删除了该块类型，工具结果改为 `role: 'tool'` 的独立消息，`isError` 上移到消息本身。`session.ts` 的 `toolResultBlocks()` / `toolResultFailed()` 是这两个差异的唯一读取点，旧宿主的嵌套输出仍会被渲染成 `[Tool Result]` 证据。
+- **注入消息的来源改为生产者自报**。0.1.7 删除了 `MessageSourceMap` 的统一包装 `{ kind: 'plugin', plugin }`（持久化准入还会直接拒绝该包装），改为每个生产者声明自己的 `kind`。插件在 `src/message-source.ts` 声明并使用 `kind: 'llm-verifier'`；宿主读取旧日志时会把历史包装行改写成 `plugin:<名称>`，那只是历史行的显示名，不影响判定。
+
+客户端面（Web 设置页与看板）另有两处只随新版本线变化：提供 `ctx.slots` 的包是 `@deepseek-ai/dsh-client-ui-renderer`（`@deepseek-ai/dsh-client-runtime` 在 0.1.2 就已删除，插件类型一律取自 `@deepseek-ai/cordis` 的 `Context`）；宿主图标名在 0.1.7 整体由尺寸后缀改为笔画后缀。图标是具名导入、且该模块对客户端 bundle 是 external，一份源码无法同时满足两套名字，因此插件改为按名探测两种拼写、都缺失时渲染空——旧宿主不会因此白屏。
+
+typecheck 门禁按 `package.json` 的 `devDependencies` 解析，本轮已从 `0.1.1-rc.2` 上移到 `0.1.7-alpha.1`：客户端面的图标名同样在 0.1.7 才改名，一份源码只能对齐一侧，宿主面的旧形态兼容由运行时代码与回归测试保证。
 
 其中一处更隐蔽的差异是 `deepFreeze`：0.1.1 的 `@deepseek-ai/dsh-llm` 会重新导出它，0.1.5 已把它移到 `@deepseek-ai/dsh-util-values`。插件因此自带一份等价的兜底实现，并同样**放过 `AbortSignal`**——冻结那个还在重试循环里使用、尚未被订阅的信号，会让传输层首次 `addEventListener`（Node ≥26.5）或超时/取消时的 `controller.abort()`（所有 Node 版本）抛 `Cannot assign to read only property`。
 
@@ -99,7 +108,7 @@ pnpm run typecheck:local   # 依据 tsconfig.local.json，把 @deepseek-ai/dsh-*
 
 发布内容由 `package.json` 的 `files` 字段决定：`lib`（构建产物、source map 与 `lib/types` 类型声明）、`src`、`cordis.patch.yml`、`README.md`。`tsconfig.local.json` 之类的本机文件不会进入 tarball。
 
-`peerDependencies` 中逐个列出的 `^0.1.5-alpha.1`、`^0.1.5-rc.1` 等预发布范围是必需的：按 semver 规则，预发布版本只有在同一 `x.y.z` 段存在带预发布的比较符时才算满足，因此不能简化成 `>=0.1.1-rc.2 <0.2.0`（那样会漏掉 `0.1.2-alpha.*`、`0.1.5-rc.*` 等宿主）。
+`peerDependencies` 中逐个列出的 `^0.1.5-alpha.1`、`^0.1.7-alpha.1` 等预发布范围是必需的：按 semver 规则，预发布版本只有在同一 `x.y.z` 段存在带预发布的比较符时才算满足，因此不能简化成 `>=0.1.1-rc.2 <0.2.0`（那样会漏掉 `0.1.2-alpha.*`、`0.1.5-rc.*`、`0.1.7-alpha.*` 等宿主）。
 
 统计数据有两条传输路径：优先走 `/api/llm-verifier/statistics`（Connection 的 exact Fetch 路由，带 Host/Origin 栅栏与浏览器鉴权），宿主较旧时回落到插件自己的 `/llm-verifier` RPC 通道（同样经过鉴权）。裸的 webServer 路由已被移除。
 

package/lib/{auto-C9YG0E22.js → auto-CGbckoaX.js}

@@ -184,6 +184,34 @@ function sessionEvents(session) {
 	if (typeof candidate?.snapshotEvents === "function") return candidate.snapshotEvents();
 	return candidate?.events ?? [];
 }
+/**
+* Inner blocks of a pre-0.1.7 nested `tool-result` block.
+*
+* DSH 0.1.6 and earlier modelled a tool result as a `tool-result` content block wrapping the
+* result's own blocks inside the answering message. DSH 0.1.7 deleted that block type: the tool
+* result is now a tool-role message whose `content` IS those blocks. Both shapes are still read,
+* because the plugin declares support for hosts on either side of the change.
+* @param block - one content block, possibly the legacy wrapper.
+* @returns The wrapped blocks, or undefined when this is not a legacy wrapper.
+*/
+function toolResultBlocks(block) {
+	if (block.type !== "tool-result") return void 0;
+	const inner = block.content;
+	return Array.isArray(inner) ? inner : void 0;
+}
+/**
+* Whether one tool result reports a failed invocation.
+*
+* DSH 0.1.7 moved the flag from the nested `tool-result` block onto the tool-role message itself
+* (`isError`); earlier hosts carry it on the block. Both are read so a failure is never scored as
+* a success on either host.
+* @param message - the `tool/result` event's message.
+* @returns True when either host's shape marks the invocation as failed.
+*/
+function toolResultFailed(message) {
+	if (message.isError === true) return true;
+	return message.content.some((block) => block.isError === true);
+}
 function textOf(blocks) {
 	const parts = [];
 	for (const block of blocks) {
@@ -191,8 +219,10 @@ function textOf(blocks) {
 		if (blockType === "text") parts.push(block.text);
 		else if (blockType === "reasoning") parts.push("[Reasoning] " + block.text);
 		else if (blockType === "tool-call") parts.push("[Tool Call] " + block.name + " " + block.arguments);
-		else if (blockType === "tool-result") parts.push("[Tool Result] " + textOf(block.content));
-		else if (blockType === "file") {
+		else if (blockType === "tool-result") {
+			const nested = toolResultBlocks(block);
+			if (nested) parts.push("[Tool Result] " + textOf(nested));
+		} else if (blockType === "file") {
 			const fileData = block;
 			parts.push("[File] " + (fileData.path ?? fileData.filename ?? fileData.title ?? "attachment"));
 		}
@@ -346,7 +376,10 @@ function blockText$1(blocks) {
 	const parts = [];
 	const visit = (items) => {
 		for (const block of items) if (block.type === "text" || block.type === "reasoning") parts.push(block.text);
-		else if (block.type === "tool-result") visit(block.content);
+		else {
+			const nested = toolResultBlocks(block);
+			if (nested) visit(nested);
+		}
 	};
 	visit(blocks);
 	return parts.join("\n").trim();
@@ -365,13 +398,16 @@ function narrativeText$1(blocks) {
 	const parts = [];
 	const visit = (items) => {
 		for (const block of items) if (block.type === "text") parts.push(block.text);
-		else if (block.type === "tool-result") visit(block.content);
+		else {
+			const nested = toolResultBlocks(block);
+			if (nested) visit(nested);
+		}
 	};
 	visit(blocks);
 	return parts.join("\n").trim();
 }
 function successful(event) {
-	return event.data.error === void 0 && event.data.message.content.every((block) => block.isError !== true);
+	return event.data.error === void 0 && !toolResultFailed(event.data.message);
 }
 function strictJson(text) {
 	const trimmed = text.trim();
@@ -1770,14 +1806,22 @@ const PASSIVE_TOOLS = /* @__PURE__ */ new Set([
 	"glob",
 	"grep",
 	"web_search",
+	"web_fetch",
 	"ssh_list",
 	"job_list",
 	"job_output",
 	"list_agents",
+	"list_subagent_models",
+	"list_mcp_resources",
+	"list_mcp_resource_templates",
+	"read_mcp_resource",
 	"get_goal",
 	"skill",
 	"mcp__codegraph__codegraph_explore",
-	"ask_user_question"
+	"ask_user_question",
+	"todo_write",
+	"present",
+	"run_code"
 ]);
 const VERIFIER_TOOLS = /* @__PURE__ */ new Set([
 	"verifier_compare",
@@ -1791,7 +1835,6 @@ const CONSEQUENTIAL_TOOLS = /* @__PURE__ */ new Set([
 	"write",
 	"pwsh",
 	"bash",
-	"run_code",
 	"codex_image_generate",
 	"ssh_exec",
 	"ssh_upload",
@@ -1911,7 +1954,7 @@ function hasPendingSubagents(events, taskStartSeq) {
 		if (event.type === "tool/result") {
 			const call = calls.get(String(event.data.message.source.callId));
 			if (call && SUBAGENT_TOOLS.has(call.data.name) && event.data.error === void 0) {
-				if (event.data.message.content.every((b) => b.isError !== true)) {
+				if (!toolResultFailed(event.data.message)) {
 					const text = blockText(event.data.message.content);
 					const contMatch = text.match(CONTINUABLE_SUBAGENT_START);
 					const jobMatch = text.match(BACKGROUND_SUBAGENT_JOB_START);
@@ -2136,7 +2179,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		if (event.type !== "tool/result") continue;
 		allResults.set(String(event.data.message.source.callId), event);
 		if (event.data.error !== void 0) continue;
-		if (!event.data.message.content.every((block) => block.isError !== true)) continue;
+		if (toolResultFailed(event.data.message)) continue;
 		results.set(String(event.data.message.source.callId), event);
 	}
 	const pairedCalls = calls.filter((event) => results.has(String(event.data.callId)));
@@ -2446,4 +2489,4 @@ function selectRouteFeedbackDetail(candidates, result, maxChars = MAX_ROUTE_FEED
 //#endregion
 export { semanticDecision as A, resolveCacheFile as B, inspectDeliveryPhase as C, nextDiagnosticCycleId as D, latestDirectUserSeq as E, extractSession as F, sanitizeVerifierText as I, sessionEvents as L, semanticRouteHint as M, verificationFailed as N, parseSemanticRoute as O, verificationVerdict as P, ScoreCache as R, estimateRoutedCalls as S, itemBudget as T, stableHash as V, analyzeStructuredRoute as _, compareRouteFeedbackDetail as a, buildSemanticRoutePrompt as b, inspectUserInteractionPause as c, selectRouteFeedbackDetail as d, sessionAccepted as f, RECOVERY_FAILURE_CONTEXT_CHARS as g, MAX_ROUTED_CHECKPOINTS as h, automaticFeedback as i, semanticReferencesVisible as j, routedRepeats as k, isAwaitingUserText as l, AutoVerifierRouter as m, PROPOSAL_FEEDBACK_NOTE as n, failedAcceptanceCriteria as o, topScoreIndices as p, analyzeAutoTask as r, hasPendingSubagents as s, MAX_ROUTE_FEEDBACK_CHARS as t, isSubagentSession as u, boundDecision as v, inspectRecoverySignal as w, buildSemanticRouteView as x, buildEvidenceIndex as y, SingleFlight as z };
 
-//# sourceMappingURL=auto-C9YG0E22.js.map
+//# sourceMappingURL=auto-CGbckoaX.js.map