plugin-sensitive-filter-xr 0.1.6 → 0.1.8

package/README.md

@@ -1,6 +1,6 @@
 # Sensitive Filter Middleware
 
-`plugin-sensitive-filter-xr` filters sensitive content for both input and output in two mutually exclusive modes:
+`@xpert-ai/plugin-sensitive-filter` filters sensitive content for both input and output in two mutually exclusive modes:
 
 - `rule`: deterministic rules (`keyword` / `regex`)
 - `llm`: natural-language policy evaluation with rewrite-only enforcement
@@ -94,8 +94,4 @@ Current behavior:
 
 ## Validation Commands
 
-```bash
-/Users/xr/Documents/code/xpert-plugins/xpertai/node_modules/.bin/tsc -p /Users/xr/Documents/code/xpert-plugins/xpertai/middlewares/sensitive-filter/tsconfig.lib.json --noEmit
-npx jest --runInBand src/lib/sensitiveFilter.spec.ts
-node /Users/xr/Documents/code/xpert-plugins/plugin-dev-harness/dist/index.js --workspace ./xpertai --plugin ./middlewares/sensitive-filter
-```
+

package/dist/lib/sensitiveFilter.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"sensitiveFilter.d.ts","sourceRoot":"","sources":["../../src/lib/sensitiveFilter.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EAAa,oBAAoB,EAA8B,MAAM,kBAAkB,CAAA;AAGnG,OAAO,EACL,eAAe,EAGf,uBAAuB,EACvB,wBAAwB,EAEzB,MAAM,sBAAsB,CAAA;AAC7B,OAAO,EAML,qBAAqB,EAKtB,MAAM,YAAY,CAAA;AAulBnB,qBAEa,yBAA0B,YAAW,wBAAwB,CAAC,qBAAqB,CAAC;IAE/F,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAY;IAEvC,QAAQ,CAAC,IAAI,EAAE,oBAAoB,CA6NlC;IAEK,gBAAgB,CACpB,OAAO,EAAE,qBAAqB,EAC9B,OAAO,EAAE,uBAAuB,GAC/B,OAAO,CAAC,eAAe,CAAC;IAa3B,OAAO,CAAC,wBAAwB;IA6OhC,OAAO,CAAC,uBAAuB;CAkbhC;AAED,YAAY,EAAE,qBAAqB,EAAE,CAAA"}
+{"version":3,"file":"sensitiveFilter.d.ts","sourceRoot":"","sources":["../../src/lib/sensitiveFilter.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAAa,oBAAoB,EAA8B,MAAM,kBAAkB,CAAA;AAGnG,OAAO,EACL,eAAe,EAGf,uBAAuB,EACvB,wBAAwB,EAEzB,MAAM,sBAAsB,CAAA;AAC7B,OAAO,EAML,qBAAqB,EAKtB,MAAM,YAAY,CAAA;AAoxBnB,qBAEa,yBAA0B,YAAW,wBAAwB,CAAC,qBAAqB,CAAC;IAE/F,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAY;IAEvC,QAAQ,CAAC,IAAI,EAAE,oBAAoB,CA6NlC;IAEK,gBAAgB,CACpB,OAAO,EAAE,qBAAqB,EAC9B,OAAO,EAAE,uBAAuB,GAC/B,OAAO,CAAC,eAAe,CAAC;IAa3B,OAAO,CAAC,wBAAwB;IAgThC,OAAO,CAAC,uBAAuB;CA2fhC;AAED,YAAY,EAAE,qBAAqB,EAAE,CAAA"}

package/dist/lib/sensitiveFilter.js

@@ -1,6 +1,8 @@
 import { __decorate, __metadata } from "tslib";
 import { z as z4 } from 'zod/v4';
-import { AIMessage, HumanMessage } from '@langchain/core/messages';
+import { AIMessage, AIMessageChunk, HumanMessage } from '@langchain/core/messages';
+import { BaseChatModel } from '@langchain/core/language_models/chat_models';
+import { ChatGenerationChunk } from '@langchain/core/outputs';
 import { Inject, Injectable } from '@nestjs/common';
 import { CommandBus } from '@nestjs/cqrs';
 import { AgentMiddlewareStrategy, CreateModelClientCommand, WrapWorkflowNodeExecutionCommand, } from '@xpert-ai/plugin-sdk';
@@ -13,6 +15,7 @@ const CONFIG_PARSE_ERROR = '敏感词过滤配置格式不正确，请检查填
 const BUSINESS_RULES_VALIDATION_ERROR = '请至少配置 1 条有效业务规则（pattern/type/action/scope/severity）。';
 const LLM_MODE_VALIDATION_ERROR = '请完善 LLM 过滤配置：需填写过滤模型、生效范围、审核规则说明。';
 const INTERNAL_LLM_INVOKE_TAG = 'sensitive-filter/internal-eval';
+const INTERNAL_SOURCE_STREAM_TAG = 'sensitive-filter/internal-source-stream';
 const INTERNAL_LLM_INVOKE_OPTIONS = {
     tags: [INTERNAL_LLM_INVOKE_TAG],
     metadata: {
@@ -141,6 +144,147 @@ function replaceModelResponseText(response, text) {
     }
     return new AIMessage(text);
 }
+function cloneAiMessage(source) {
+    return new AIMessage({
+        content: source.content,
+        additional_kwargs: source.additional_kwargs,
+        response_metadata: source.response_metadata,
+        tool_calls: source.tool_calls,
+        invalid_tool_calls: source.invalid_tool_calls,
+        usage_metadata: source.usage_metadata,
+        id: source.id,
+        name: source.name,
+    });
+}
+function cloneAiMessageWithText(source, text) {
+    const cloned = cloneAiMessage(source);
+    cloned.content = text;
+    return cloned;
+}
+function toAiMessageChunk(value) {
+    if (value instanceof AIMessageChunk) {
+        return value;
+    }
+    if (!isRecord(value) || !('content' in value)) {
+        return null;
+    }
+    return new AIMessageChunk({
+        content: value['content'],
+        additional_kwargs: isRecord(value['additional_kwargs']) ? value['additional_kwargs'] : {},
+        response_metadata: isRecord(value['response_metadata']) ? value['response_metadata'] : {},
+        tool_call_chunks: Array.isArray(value['tool_call_chunks']) ? value['tool_call_chunks'] : [],
+        tool_calls: Array.isArray(value['tool_calls']) ? value['tool_calls'] : [],
+        invalid_tool_calls: Array.isArray(value['invalid_tool_calls']) ? value['invalid_tool_calls'] : [],
+        usage_metadata: isRecord(value['usage_metadata']) ? value['usage_metadata'] : undefined,
+        id: typeof value['id'] === 'string' ? value['id'] : undefined,
+    });
+}
+function toAiMessage(value) {
+    if (value instanceof AIMessage) {
+        return value;
+    }
+    if (value instanceof AIMessageChunk) {
+        return new AIMessage({
+            content: value.content,
+            additional_kwargs: value.additional_kwargs,
+            response_metadata: value.response_metadata,
+            tool_calls: value.tool_calls,
+            invalid_tool_calls: value.invalid_tool_calls,
+            usage_metadata: value.usage_metadata,
+            id: value.id,
+            name: value.name,
+        });
+    }
+    if (isRecord(value) && 'content' in value) {
+        return new AIMessage({
+            content: value['content'],
+            additional_kwargs: isRecord(value['additional_kwargs']) ? value['additional_kwargs'] : {},
+            response_metadata: isRecord(value['response_metadata']) ? value['response_metadata'] : {},
+            tool_calls: Array.isArray(value['tool_calls']) ? value['tool_calls'] : [],
+            invalid_tool_calls: Array.isArray(value['invalid_tool_calls']) ? value['invalid_tool_calls'] : [],
+            usage_metadata: isRecord(value['usage_metadata']) ? value['usage_metadata'] : undefined,
+            id: typeof value['id'] === 'string' ? value['id'] : undefined,
+            name: typeof value['name'] === 'string' ? value['name'] : undefined,
+        });
+    }
+    return new AIMessage(extractPrimitiveText(value));
+}
+function buildInternalSourceOptions(options) {
+    const tags = Array.isArray(options?.tags) ? options.tags : [];
+    const metadata = isRecord(options?.metadata) ? options.metadata : {};
+    return {
+        ...(options ?? {}),
+        tags: [...tags, INTERNAL_SOURCE_STREAM_TAG],
+        metadata: {
+            ...metadata,
+            internal: true,
+        },
+    };
+}
+class BufferedOutputProxyChatModel extends BaseChatModel {
+    constructor(innerModel, resolveOutput) {
+        super({});
+        this.innerModel = innerModel;
+        this.resolveOutput = resolveOutput;
+    }
+    _llmType() {
+        return 'sensitive-filter-output-proxy';
+    }
+    async collectInnerMessage(messages, options) {
+        const internalOptions = buildInternalSourceOptions(options);
+        const streamFn = this.innerModel?.stream;
+        if (typeof streamFn === 'function') {
+            const stream = await streamFn.call(this.innerModel, messages, internalOptions);
+            if (stream && typeof stream[Symbol.asyncIterator] === 'function') {
+                let mergedChunk = null;
+                for await (const rawChunk of stream) {
+                    const chunk = toAiMessageChunk(rawChunk);
+                    if (!chunk) {
+                        continue;
+                    }
+                    mergedChunk = mergedChunk ? mergedChunk.concat(chunk) : chunk;
+                }
+                if (mergedChunk) {
+                    return toAiMessage(mergedChunk);
+                }
+            }
+        }
+        return toAiMessage(await this.innerModel.invoke(messages, internalOptions));
+    }
+    async finalizeMessage(messages, options) {
+        const sourceMessage = await this.collectInnerMessage(messages, options);
+        return this.resolveOutput(sourceMessage, extractPrimitiveText(sourceMessage.content));
+    }
+    async _generate(messages, options, _runManager) {
+        const resolved = await this.finalizeMessage(messages, options);
+        return {
+            generations: [
+                {
+                    text: extractPrimitiveText(resolved.finalMessage.content),
+                    message: resolved.finalMessage,
+                },
+            ],
+        };
+    }
+    async *_streamResponseChunks(messages, options, runManager) {
+        const resolved = await this.finalizeMessage(messages, options);
+        const finalText = extractPrimitiveText(resolved.finalMessage.content);
+        if (!finalText) {
+            return;
+        }
+        const generationChunk = new ChatGenerationChunk({
+            message: new AIMessageChunk({
+                content: finalText,
+                id: resolved.finalMessage.id,
+            }),
+            text: finalText,
+        });
+        yield generationChunk;
+        await runManager?.handleLLMNewToken(finalText, undefined, undefined, undefined, undefined, {
+            chunk: generationChunk,
+        });
+    }
+}
 function rewriteModelRequestInput(request, rewrittenText) {
     if (!Array.isArray(request?.messages) || request.messages.length === 0) {
         return request;
@@ -720,12 +864,14 @@ let SensitiveFilterMiddleware = class SensitiveFilterMiddleware {
         };
         let inputBlockedMessage = null;
         let pendingInputRewrite = null;
+        let bufferedOutputResolution = null;
         let finalAction = 'pass';
         let auditEntries = [];
         let runtimeConfigurable = null;
         const resetRunState = () => {
             inputBlockedMessage = null;
             pendingInputRewrite = null;
+            bufferedOutputResolution = null;
             finalAction = 'pass';
             auditEntries = [];
         };
@@ -847,7 +993,63 @@ let SensitiveFilterMiddleware = class SensitiveFilterMiddleware {
                 }
                 const modelRequest = pendingInputRewrite ? rewriteModelRequestInput(request, pendingInputRewrite) : request;
                 pendingInputRewrite = null;
-                const response = await handler(modelRequest);
+                bufferedOutputResolution = null;
+                const shouldBufferOutput = compiledRules.some((rule) => rule.scope === 'output' || rule.scope === 'both');
+                const effectiveRequest = shouldBufferOutput
+                    ? {
+                        ...modelRequest,
+                        model: new BufferedOutputProxyChatModel(modelRequest.model, async (message, outputText) => {
+                            if (message.tool_calls?.length || message.invalid_tool_calls?.length) {
+                                bufferedOutputResolution = {
+                                    finalMessage: cloneAiMessage(message),
+                                    matched: false,
+                                    source: 'rule',
+                                    reason: 'tool-call-skip',
+                                    errorPolicyTriggered: false,
+                                };
+                                return bufferedOutputResolution;
+                            }
+                            const outputMatches = findMatches(outputText, 'output', compiledRules, normalize, caseSensitive);
+                            const winner = pickWinningRule(outputMatches);
+                            if (!winner) {
+                                bufferedOutputResolution = {
+                                    finalMessage: cloneAiMessage(message),
+                                    matched: false,
+                                    source: 'rule',
+                                    errorPolicyTriggered: false,
+                                };
+                                return bufferedOutputResolution;
+                            }
+                            const finalText = winner.action === 'block'
+                                ? winner.replacementText?.trim() || DEFAULT_OUTPUT_BLOCK_MESSAGE
+                                : rewriteTextByRule(outputText, winner, caseSensitive);
+                            bufferedOutputResolution = {
+                                finalMessage: cloneAiMessageWithText(message, finalText),
+                                matched: true,
+                                source: 'rule',
+                                action: winner.action,
+                                reason: `rule:${winner.id}`,
+                                errorPolicyTriggered: false,
+                            };
+                            return bufferedOutputResolution;
+                        }),
+                    }
+                    : modelRequest;
+                const response = await handler(effectiveRequest);
+                if (bufferedOutputResolution) {
+                    pushAudit({
+                        phase: 'output',
+                        matched: bufferedOutputResolution.matched,
+                        source: bufferedOutputResolution.source,
+                        action: bufferedOutputResolution.action,
+                        reason: bufferedOutputResolution.reason,
+                        errorPolicyTriggered: bufferedOutputResolution.errorPolicyTriggered,
+                    });
+                    if (bufferedOutputResolution.matched && bufferedOutputResolution.action) {
+                        finalAction = bufferedOutputResolution.action === 'block' ? 'block' : 'rewrite';
+                    }
+                    return response;
+                }
                 const outputText = extractModelResponseText(response);
                 const outputMatches = findMatches(outputText, 'output', compiledRules, normalize, caseSensitive);
                 const winner = pickWinningRule(outputMatches);
@@ -915,6 +1117,7 @@ let SensitiveFilterMiddleware = class SensitiveFilterMiddleware {
             return structuredModelPromises.get(method);
         };
         let pendingInputRewrite = null;
+        let bufferedOutputResolution = null;
         let finalAction = 'pass';
         let auditEntries = [];
         let runtimeConfigurable = null;
@@ -923,6 +1126,7 @@ let SensitiveFilterMiddleware = class SensitiveFilterMiddleware {
         let methodAttempts = [];
         const resetRunState = () => {
             pendingInputRewrite = null;
+            bufferedOutputResolution = null;
             finalAction = 'pass';
             auditEntries = [];
             resolvedOutputMethod = undefined;
@@ -1183,7 +1387,70 @@ let SensitiveFilterMiddleware = class SensitiveFilterMiddleware {
                 const llmConfig = getLlmConfig();
                 const modelRequest = pendingInputRewrite ? rewriteModelRequestInput(request, pendingInputRewrite) : request;
                 pendingInputRewrite = null;
-                const response = await handler(modelRequest);
+                bufferedOutputResolution = null;
+                const effectiveRequest = modeIncludesScope(llmConfig.scope, 'output')
+                    ? {
+                        ...modelRequest,
+                        model: new BufferedOutputProxyChatModel(modelRequest.model, async (message, outputText) => {
+                            if (message.tool_calls?.length || message.invalid_tool_calls?.length) {
+                                bufferedOutputResolution = {
+                                    finalMessage: cloneAiMessage(message),
+                                    matched: false,
+                                    source: 'llm',
+                                    reason: 'tool-call-skip',
+                                    errorPolicyTriggered: false,
+                                };
+                                return bufferedOutputResolution;
+                            }
+                            if (!outputText) {
+                                bufferedOutputResolution = {
+                                    finalMessage: cloneAiMessage(message),
+                                    matched: false,
+                                    source: 'llm',
+                                    reason: 'empty-output',
+                                    errorPolicyTriggered: false,
+                                };
+                                return bufferedOutputResolution;
+                            }
+                            let decision;
+                            let fromErrorPolicy = false;
+                            try {
+                                decision = await invokeAndTrack('output', outputText, request?.runtime, llmConfig);
+                            }
+                            catch (error) {
+                                decision = resolveOnErrorDecision(llmConfig, error);
+                                fromErrorPolicy = true;
+                            }
+                            const finalText = decision.matched && decision.action
+                                ? toNonEmptyString(decision.replacementText) ?? llmConfig.rewriteFallbackText
+                                : outputText;
+                            bufferedOutputResolution = {
+                                finalMessage: cloneAiMessageWithText(message, finalText),
+                                matched: decision.matched,
+                                source: fromErrorPolicy ? 'error-policy' : 'llm',
+                                action: decision.action,
+                                reason: decision.reason,
+                                errorPolicyTriggered: fromErrorPolicy,
+                            };
+                            return bufferedOutputResolution;
+                        }),
+                    }
+                    : modelRequest;
+                const response = await handler(effectiveRequest);
+                if (bufferedOutputResolution) {
+                    pushAudit({
+                        phase: 'output',
+                        matched: bufferedOutputResolution.matched,
+                        source: bufferedOutputResolution.source,
+                        action: bufferedOutputResolution.action,
+                        reason: bufferedOutputResolution.reason,
+                        errorPolicyTriggered: bufferedOutputResolution.errorPolicyTriggered,
+                    });
+                    if (bufferedOutputResolution.matched && bufferedOutputResolution.action) {
+                        finalAction = 'rewrite';
+                    }
+                    return response;
+                }
                 if (!modeIncludesScope(llmConfig.scope, 'output')) {
                     pushAudit({
                         phase: 'output',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "plugin-sensitive-filter-xr",
-  "version": "0.1.6",
+  "version": "0.1.8",
   "author": {
     "name": "XpertAI",
     "url": "https://xpertai.cn"