diff --git a/.env.example b/.env.example index c4331c4..a599f59 100644 --- a/.env.example +++ b/.env.example @@ -141,6 +141,33 @@ H5_PUBLIC_BASE_URL=http://127.0.0.1:5173 # 1/on：gateway 优先读 classification.decision.route / session_hint # MEMIND_ROUTER_NORMALIZED_DECISION=0 +# H5 Chat LLM intent router（本地开发建议见下方「本地 LLM Router」块） +# 0：仅规则路由；规则未命中 fallback Agent +# 1 + SHADOW=1：调用轻量 LLM 观测分歧，行为仍 fallback Agent +# 1 + CANARY_USER_IDS：仅名单内用户在规则未命中时走 LLM 路由；留空=全员 +# MEMIND_CHAT_LLM_ROUTER_ENABLED=0 +# MEMIND_CHAT_LLM_ROUTER_SHADOW=0 +# MEMIND_CHAT_ROUTER_CANARY_USER_IDS= +# MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID= +# MEMIND_CHAT_ROUTER_MODEL= +# MEMIND_CHAT_ROUTER_MIN_CONFIDENCE=0.65 +# MEMIND_CHAT_ROUTER_TIMEOUT_MS=1200 +# MEMIND_CHAT_ROUTER_MEMORY_ENABLED=1 +# MEMIND_CHAT_ROUTER_FALLBACK_ROUTE=agent_orchestration + +# ----- 本地 LLM Router（pnpm dev，全员 Active，无 Shadow）----- +# Router 模型建议（分类 JSON，短 prompt，要快不要大）： +# 本机 Ollama：qwen2.5:3b / qwen2.5:1.5b / llama3.2:3b（见 H5_LOCAL_LLM_*） +# 云端 relay：gpt-4o-mini / deepseek-chat / qwen-turbo / glm-4-flash +# Direct Chat 仍用全局选中模型；Router 单独绑小模型即可。 +# MEMIND_CHAT_LLM_ROUTER_ENABLED=1 +# MEMIND_CHAT_LLM_ROUTER_SHADOW=0 +# MEMIND_CHAT_ROUTER_CANARY_USER_IDS= +# MEMIND_CHAT_ROUTER_MIN_CONFIDENCE=0.65 +# MEMIND_CHAT_ROUTER_TIMEOUT_MS=1200 +# MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID= +# MEMIND_CHAT_ROUTER_MODEL=qwen2.5:3b + # H5 Session stream replay（docs/h5-session-architecture-20260706.md Patch 4c） # 默认 0：session SSE 仍纯透传 goosed；设 1 时 Portal 持久化 session 事件并支持 Last-Event-ID 重连补发。 # MEMIND_SESSION_STREAM_REPLAY=0 @@ -195,7 +222,8 @@ H5_ADMIN_PASSWORD=change-me-admin # H5_ADMIN_WORKSPACE_ROOT=/Users/john/PycharmProjects/tkmind # LLM Key 数据库加密（可选；默认派生自 TKMIND_SERVER__SECRET_KEY） # H5_SETTINGS_ENCRYPTION_KEY=change-me-encryption-key -# 启动时自动导入 Relay（若不存在同名配置） +# 启动时自动导入 Relay（若不存在同名配置）；改用手动 Provider 时请设为 1 +# H5_RELAY_BOOTSTRAP_DISABLED=1 # H5_RELAY_BOOTSTRAP_NAME=Local Goose Relay DeepSeek # H5_RELAY_BOOTSTRAP_URL=http://127.0.0.1:18300/relay/buyer/v1/chat/completions # H5_RELAY_BOOTSTRAP_API_KEY=your-bearer-token @@ -305,7 +333,33 @@ VITE_TKMIND_WORKING_DIR=/Users/john/PycharmProjects/tkmind # VITE_TKMIND_MODEL=gpt-4o # 可选：harness recall 的默认查询 -# VITE_TKMIND_MEMORY_QUERY=project architecture conventions current goals decisions risks and recent work +# Memory V2 Phase A canary (local dev) +# Full template: docs/memory-v2/phase-a-canary.env.example +# Readiness check: npm run check:memory-v2-phase-a +# +# Recommended local canary users (explicit-memory testers): +# 32035858-9a20-425b-89da-c118ef0779aa +# 1c99b83b-0454-474f-a5d2-129d34506a32 +# MEMORY_ENABLED=1 +# MEMORY_CANDIDATE_ENABLED=1 +# MEMORY_CANDIDATE_MODE=canary +# MEMORY_CANDIDATE_PERSISTENCE_ENABLED=1 +# MEMORY_CANDIDATE_AUTO_ACCEPT_ALL=0 +# MEMORY_AGENT_RESOLVE_ENABLED=1 +# MEMORY_AGENT_INJECTION_MODE=canary +# MEMORY_AGENT_CANARY_USER_IDS=32035858-9a20-425b-89da-c118ef0779aa,1c99b83b-0454-474f-a5d2-129d34506a32 +# MEMORY_LIFECYCLE_ENABLED=1 +# MEMORY_LIFECYCLE_ROLLOUT_MODE=canary +# MEMORY_LIFECYCLE_ROLLOUT_USER_IDS=32035858-9a20-425b-89da-c118ef0779aa,1c99b83b-0454-474f-a5d2-129d34506a32 +# MEMORY_PROMOTION_ENABLED=1 +# MEMORY_LIFECYCLE_FORGETTING_ENABLED=0 + +# Goal Run MVP（默认关闭；canary 用户名单见 docs/architecture/goal-run-data-model-draft.md） +# GOAL_RUN_ENABLED=0 +# GOAL_RUN_CANARY_USER_IDS= +# GOAL_RUN_REQUIRE_APPROVAL=0 +# MEMORY_LIFECYCLE_DECAY_ENABLED=0 + # 历史列表每次展示/加载条数（默认 25，与 desktop 一致） # VITE_SESSION_PAGE_SIZE=25 diff --git a/chat-intent-router-rules.mjs b/chat-intent-router-rules.mjs new file mode 100644 index 0000000..2575b07 --- /dev/null +++ b/chat-intent-router-rules.mjs @@ -0,0 +1,124 @@ +/** + * H5 Chat Intent Router — FAQ / fast-path rule library. + * + * Add patterns here for messages that should bypass LLM routing latency. + * Consumed by chat-intent-router.mjs → classifyWithRules(). + * + * Keep rules conservative: task execution, page generation, and realtime lookup + * must stay on Agent paths (checked via FAQ_EXCLUSION_PATTERNS). + */ + +/** Block FAQ match when the message clearly needs Agent / tools. */ +export const FAQ_EXCLUSION_PATTERNS = [ + /\bpublic\/[^\s"'<>]+\.html\b/i, + /MindSpace\/[^/\s]+\/public\/[^\s"'<>]+\.html/i, + /(?:生成|创建|制作|做|写|设计|发布|搭建).{0,24}(?:H5|h5|HTML|html|网页|页面|活动页|宣传页|落地页|分享页|链接)/u, + /(?:H5|h5|HTML|html|网页|页面).{0,24}(?:生成|创建|制作|做|写|设计|发布|链接)/u, + /(?:写|做|生成|制作|设计).{0,16}主题(?:页面|文章)/u, + /(?:服务号|公众号).{0,16}(?:页面|文章|推文|图文)/u, + /(?:今天|今日|现在|最新|当前|实时).{0,12}(?:新闻|赛况|比分|天气|行情|股价|汇率|热点)/u, + /(?:查(?:一下|查)?|搜索|打听).{0,16}(?:今天|今日|现在|最新|当前)/u, + /(?:世界杯|欧冠|NBA|英超).{0,16}(?:赛况|赛程|比分)/u, + /(?:load_skill|page-data|sqlite|express|后端服务)/iu, + /(?:修|改|测试|回归|verify).{0,12}(?:bug|问卷|policy|public)/iu, +]; + +/** + * @typedef {{ id: string, reason: string, patterns: RegExp[] }} DirectChatFaqRule + */ + +/** @type {DirectChatFaqRule[]} */ +export const DIRECT_CHAT_FAQ_RULES = [ + { + id: 'concept_what_is', + reason: '概念解释类问答（FAQ）', + patterns: [ + /^什么是[\s\S]{1,48}[？?]?$/u, + /^何为[\s\S]{1,48}[？?]?$/u, + /^什么叫[\s\S]{1,48}[？?]?$/u, + /^[\s\S]{1,40}是什么[？?]?$/u, + /^[\s\S]{1,40}是什么意思[？?]?$/u, + /^[\s\S]{1,40}(?:是什么东西|指什么|啥意思)[？?]?$/u, + ], + }, + { + id: 'concept_explain', + reason: '解释说明类问答（FAQ）', + patterns: [ + /^(?:请)?(?:解释|说明|介绍|讲解|阐述)(?:一下|下)?[\s\S]{1,48}[？?]?$/u, + /^帮我(?:解释|说明|介绍|讲解)(?:一下|下)?[\s\S]{1,48}[？?]?$/u, + /^能(?:不能|否)(?:解释|说明|介绍)(?:一下|下)?[\s\S]{1,40}[？?]?$/u, + ], + }, + { + id: 'concept_compare', + reason: '概念对比类问答（FAQ）', + patterns: [ + /^[\s\S]{1,32}(?:和|与|跟)[\s\S]{1,32}(?:的)?(?:区别|差异|不同|对比|比较)[是什么]?[？?]?$/u, + /^(?:对比|比较)(?:一下|下)?[\s\S]{1,48}[？?]?$/u, + /^[\s\S]{1,24}(?:vs|VS|Vs)[\s\S]{1,24}(?:区别|差异|不同)?[？?]?$/u, + ], + }, + { + id: 'concept_how_works', + reason: '原理/机制类问答（FAQ）', + patterns: [ + /^[\s\S]{1,40}(?:是怎么|是如何)(?:工作|运行|实现|运作)的[？?]?$/u, + /^[\s\S]{1,40}(?:原理|机制|流程)(?:是什么|是怎样的)[？?]?$/u, + /^为什么[\s\S]{1,48}[？?]?$/u, + ], + }, + { + id: 'text_transform', + reason: '纯文字处理（总结/翻译/改写）', + patterns: [ + /^(?:请)?(?:帮我|帮忙)?(?:总结|概括|归纳|提炼)(?:一下|下)?[\s\S]{1,80}[？?]?$/u, + /^(?:请)?(?:帮我|帮忙)?(?:翻译|改写|润色|扩写|缩写|校对)(?:一下|下)?[\s\S]{1,80}[？?]?$/u, + ], + }, + { + id: 'tech_acronym', + reason: '常见技术名词解释（FAQ）', + patterns: [ + /^(?:什么是|解释一下|介绍一下)?\s*(?:SSE|WebSocket|REST(?:ful)?|GraphQL|JWT|OAuth|RAG|Embedding|向量数据库|Server-Sent Events)[\s\S]{0,24}[？?]?$/iu, + /^(?:SSE|WebSocket|REST(?:ful)?|GraphQL|JWT|OAuth|RAG)(?:是什么|什么意思|怎么用)[？?]?$/iu, + ], + }, + { + id: 'learning_advice', + reason: '学习/选型建议（纯文字）', + patterns: [ + /^(?:学|学习|入门).{0,24}(?:应该|该|推荐|建议).{0,32}(?:先|从哪|哪个)[？?]?$/u, + /^(?:初学者|新手|入门).{0,24}(?:应该|该|推荐|建议).{0,32}[？?]?$/u, + /^.{0,24}(?:先学|从哪学|怎么学).{0,32}[？?]?$/u, + ], + }, + { + id: 'casual_opinion', + reason: '主观建议/闲聊（无任务执行）', + patterns: [ + /^你觉得.{1,48}[？?]?$/u, + /^你认为.{1,48}[？?]?$/u, + /^(?:随便|随意)(?:聊|说说|谈谈).{0,24}[？?]?$/u, + /^.{0,16}(?:有什么建议|给点建议|怎么看)[？?]?$/u, + ], + }, +]; + +const MAX_FAQ_TEXT_LENGTH = 200; + +/** + * @param {string} text + * @returns {{ id: string, reason: string } | null} + */ +export function matchDirectChatFaqRule(text) { + const normalized = String(text ?? '').trim(); + if (!normalized || normalized.length > MAX_FAQ_TEXT_LENGTH) return null; + if (FAQ_EXCLUSION_PATTERNS.some((pattern) => pattern.test(normalized))) return null; + for (const rule of DIRECT_CHAT_FAQ_RULES) { + if (rule.patterns.some((pattern) => pattern.test(normalized))) { + return { id: rule.id, reason: rule.reason }; + } + } + return null; +} diff --git a/chat-intent-router.mjs b/chat-intent-router.mjs index 4f0a392..7459134 100644 --- a/chat-intent-router.mjs +++ b/chat-intent-router.mjs @@ -15,6 +15,10 @@ import { resolveMemoryInterventionMode, } from './memory-intervention.mjs'; import { filterMemoriesByQuery } from './memory-legacy-fallback.mjs'; +import { matchDirectChatFaqRule } from './chat-intent-router-rules.mjs'; +import { isGoalRunIntent } from './goal-run-intent.mjs'; + +export { matchDirectChatFaqRule, DIRECT_CHAT_FAQ_RULES, FAQ_EXCLUSION_PATTERNS } from './chat-intent-router-rules.mjs'; export const CHAT_INTENT_ROUTE = { DIRECT_CHAT: 'direct_chat', @@ -42,6 +46,7 @@ const ALLOWED_ROUTER_DECISION_FLAGS = new Set([ 'force_deep_reasoning', 'code_task', 'long_running', + 'goal_run', ]); const AGENT_ORCHESTRATION_HEADER = '【Memind 任务编排】'; @@ -304,9 +309,9 @@ function resolveRouterFallbackRoute(env, overrides = {}) { return raw === CHAT_INTENT_ROUTE.DIRECT_CHAT ? CHAT_INTENT_ROUTE.AGENT : raw; } -const DEFAULT_ROUTER_TIMEOUT_MS = 2500; +const DEFAULT_ROUTER_TIMEOUT_MS = 1200; const DEFAULT_ROUTER_MEMORY_LIMIT = 8; -const DEFAULT_ROUTER_MIN_CONFIDENCE = 0.55; +const DEFAULT_ROUTER_MIN_CONFIDENCE = 0.65; const REALTIME_WEB_AGENT_BRIEF = '先 load_skill → web；获取实时信息时同一轮并行调用 tkmind_search（专用联网搜索）和 web_search（内置联网搜索），合并去重并保留来源。向用户只称「联网搜索」，不要提具体搜索引擎或中间件名称。任一侧不可用时继续使用另一侧，必要时再用 fetch_url 读取可靠来源；禁止使用 search 技能查工作区，不要反复 scrape 同一站点。'; @@ -623,22 +628,17 @@ export function buildRouterNormalizedDecision( const mode = ROUTER_DECISION_MODE.SSE; let sessionHint = ROUTER_DECISION_SESSION_HINT.NEW; if (sessionId) { + // Portal /agent/start already binds a Goose session before the first user turn. + // Reuse it on agent fallback (e.g. direct_chat_failed) instead of orphaning it. sessionHint = ROUTER_DECISION_SESSION_HINT.REUSE; } - if ( - sessionId && - !isDirectChatSessionId(sessionId) && - sessionMessageCount != null && - Number(sessionMessageCount) === 0 - ) { - sessionHint = ROUTER_DECISION_SESSION_HINT.NEW; - } const flags = []; if (isMemoryRecallQuestion(text)) flags.push('memory_recall'); if (forceDeepReasoning) flags.push('force_deep_reasoning'); if (toolMode === 'code') flags.push('code_task'); if (classification?.suggestedSkill) flags.push('selected_skill'); + if (isGoalRunIntent(text)) flags.push('goal_run'); const normalizedFlags = [...new Set(flags)].filter((flag) => ALLOWED_ROUTER_DECISION_FLAGS.has(flag)); return { @@ -693,8 +693,9 @@ export function logRouterDecisionShadow( } /** - * Apply router session_hint for goosed agent path without overriding hard rules. - * Returns null when gateway should start a fresh goosed session. + * Resolve the Goose session id for agent-run gateway execution. + * When Portal already bound a session, always reuse it; session_hint only + * requests a fresh Goose session when no agentSessionId is present yet. */ export function resolveGatewayAgentSessionId({ agentSessionId = null, @@ -725,16 +726,63 @@ export function resolveGatewayAgentSessionId({ return sessionId; } - if (decision.session_hint === ROUTER_DECISION_SESSION_HINT.NEW) { - return null; - } - return sessionId; } +export function normalizeChatRouterCanaryUserIds(value) { + return normalizeAgentCanaryUserIds(value); +} + +export function isChatLlmRouterEligible({ + enabled = false, + canaryUserIds = [], + userId = null, +} = {}) { + if (!enabled) return false; + if (!Array.isArray(canaryUserIds) || canaryUserIds.length === 0) return true; + const normalizedUserId = String(userId ?? '').trim(); + if (!normalizedUserId) return false; + return canaryUserIds.includes(normalizedUserId); +} + +export function logChatLlmRouterShadow( + { + baseline = null, + llmResult = null, + userId = null, + sessionId = null, + routerMemoryItemsUsed = null, + } = {}, + { logger = console } = {}, +) { + const payload = { + userId: userId ?? null, + sessionId: sessionId ?? null, + baselineRoute: baseline?.route ?? null, + baselineSource: baseline?.source ?? null, + llmRoute: llmResult?.route ?? null, + llmConfidence: llmResult?.confidence ?? null, + llmReason: llmResult?.reason ?? null, + routerMemoryItemsUsed, + wouldChangeRoute: Boolean( + llmResult?.route != null + && baseline?.route != null + && llmResult.route !== baseline.route, + ), + }; + const line = `[chat-llm-router-shadow] ${JSON.stringify(payload)}`; + if (typeof logger.warn === 'function') logger.warn(line); + else if (typeof logger.info === 'function') logger.info(line); + else logger.log(line); +} + export function resolveChatIntentRouterPolicy({ env = process.env, overrides = {} } = {}) { return { enabled: envFlag(env?.MEMIND_CHAT_LLM_ROUTER_ENABLED, false), + shadowMode: envFlag(env?.MEMIND_CHAT_LLM_ROUTER_SHADOW, false), + canaryUserIds: normalizeChatRouterCanaryUserIds( + overrides.canaryUserIds ?? env?.MEMIND_CHAT_ROUTER_CANARY_USER_IDS, + ), modelProviderKeyId: String(env?.MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID ?? '').trim() || null, model: String(env?.MEMIND_CHAT_ROUTER_MODEL ?? '').trim() || null, modelApiType: String(env?.MEMIND_CHAT_ROUTER_MODEL_API ?? '').trim() || 'chat', @@ -811,6 +859,7 @@ export function buildAgentOrchestrationAgentText({ classification, skillPrompt = '', memoryContext = null, + goalContext = null, }) { const taskBody = String(displayText ?? '').trim(); const memoryLines = memoryContext?.injectionEnabled @@ -823,6 +872,9 @@ export function buildAgentOrchestrationAgentText({ .filter(Boolean) .slice(0, 8) : []; + const goalEnvelope = goalContext?.injectionEnabled + ? String(goalContext.envelope ?? '').trim() + : ''; const lines = [ `${AGENT_ORCHESTRATION_HEADER}以下为用户任务，请使用工具与技能实际执行并产出结果，不要只做文字描述。`, `路由判定：${classification.reason}`, @@ -830,6 +882,7 @@ export function buildAgentOrchestrationAgentText({ classification.suggestedSkill ? `建议 skill：${classification.suggestedSkill}` : '', buildImageGenerationInstruction(classification.imageGeneration), skillPrompt, + goalEnvelope, memoryLines.length ? [ '[Memory Context]', @@ -847,7 +900,7 @@ export function buildAgentOrchestrationAgentText({ export function applyAgentOrchestrationToUserMessage( userMessage, classification, - { grantedSkills = [], memoryContext = null } = {}, + { grantedSkills = [], memoryContext = null, goalContext = null } = {}, ) { const displayText = messageDisplayText(userMessage); const skillPrompt = resolveSkillPrompt(classification?.suggestedSkill, grantedSkills, displayText); @@ -856,6 +909,7 @@ export function applyAgentOrchestrationToUserMessage( classification, skillPrompt, memoryContext, + goalContext, }); const content = Array.isArray(userMessage?.content) ? userMessage.content.map((item, index) => { @@ -894,7 +948,6 @@ export function classifyWithRules({ sessionMessageCount = null, userMessage = null, includeIntentPatterns = true, - llmRouterEnabled = false, } = {}) { const decisionContext = { text, @@ -992,6 +1045,16 @@ export function classifyWithRules({ if (includeIntentPatterns && normalized && isRealtimeInfoQuestion(normalized)) { return finalizeRouterClassification(buildRealtimeInfoClassification(), decisionContext); } + if (includeIntentPatterns && normalized) { + const faqMatch = matchDirectChatFaqRule(normalized); + if (faqMatch) { + return finalizeRouterClassification(normalizeClassification({ + route: CHAT_INTENT_ROUTE.DIRECT_CHAT, + confidence: 0.94, + reason: faqMatch.reason, + }, { source: 'rule' }), decisionContext); + } + } if ( includeIntentPatterns && normalized && @@ -1030,6 +1093,8 @@ export function createChatIntentRouter(options = {}) { env, overrides: pickDefined(options, [ 'enabled', + 'shadowMode', + 'canaryUserIds', 'modelProviderKeyId', 'model', 'modelApiType', @@ -1041,6 +1106,68 @@ export function createChatIntentRouter(options = {}) { 'fallbackRoute', ]), }); + + function isLlmRouterEligibleForUser(userId) { + return isChatLlmRouterEligible({ + enabled: policy.enabled, + canaryUserIds: policy.canaryUserIds, + userId, + }); + } + + function buildFallbackClassification(reason = '规则未命中，走 skill 默认 Agent 通道') { + return normalizeClassification({ + route: policy.fallbackRoute, + confidence: 0.5, + reason, + }, { source: 'fallback', fallbackRoute: policy.fallbackRoute }); + } + + async function classifyWithLlm({ + text, + grantedSkills = [], + routerContext = null, + } = {}) { + if (typeof llmProviderService?.createChatCompletion !== 'function') return null; + try { + const completion = await withTimeout( + llmProviderService.createChatCompletion({ + providerKeyId: policy.modelProviderKeyId || undefined, + model: policy.model || undefined, + temperature: policy.temperature ?? 0, + messages: [ + { role: 'system', content: buildRouterSystemPrompt(grantedSkills) }, + { + role: 'user', + content: buildRouterUserPrompt({ + text, + routerContext: routerContext?.content ?? null, + }), + }, + ], + }), + policy.timeoutMs, + 'Chat intent router', + ); + if (!completion?.ok) return null; + const parsed = parseRouterJson(completion.reply); + if (!parsed) return null; + return normalizeClassification({ + route: parsed.route, + confidence: parsed.confidence, + reason: parsed.reason, + suggested_skill: parsed.suggested_skill ?? parsed.suggestedSkill, + agent_brief: parsed.agent_brief ?? parsed.agentBrief, + providerKeyId: completion.providerKeyId ?? policy.modelProviderKeyId, + model: completion.model ?? policy.model, + }, { source: 'llm', fallbackRoute: policy.fallbackRoute }); + } catch (err) { + logger?.warn?.( + `[chat-intent-router] LLM classify skipped: ${err instanceof Error ? err.message : err}`, + ); + return null; + } + } const agentMemoryPolicy = { enabled: envFlag(env?.MEMORY_AGENT_RESOLVE_ENABLED, false), mode: normalizeAgentInjectionMode(env?.MEMORY_AGENT_INJECTION_MODE), @@ -1062,8 +1189,10 @@ export function createChatIntentRouter(options = {}) { return { enabled: true, ruleRoutingEnabled: true, - llmRoutingEnabled: false, + llmRoutingEnabled: Boolean(policy.enabled && !policy.shadowMode), + llmRoutingShadow: Boolean(policy.enabled && policy.shadowMode), configuredLlmEnabled: Boolean(policy.enabled), + llmRoutingCanaryUserCount: policy.canaryUserIds.length, modelProviderKeyId: policy.modelProviderKeyId ?? null, model: policy.model ?? null, modelApiType: policy.modelApiType ?? 'chat', @@ -1083,8 +1212,8 @@ export function createChatIntentRouter(options = {}) { } function isEnabled() { - // Rule + skill routing is always active. The LLM classifier stays disabled - // independently; returning false here would make the gateway skip rules too. + // Rule + skill routing is always active. LLM routing is separately gated by + // MEMIND_CHAT_LLM_ROUTER_ENABLED; returning false here would skip rules too. return true; } @@ -1296,6 +1425,7 @@ export function createChatIntentRouter(options = {}) { }), }, decisionContext); }; + const llmRouterEligible = isLlmRouterEligibleForUser(userId); const ruleResult = classifyWithRules({ text, forceDeepReasoning, @@ -1304,20 +1434,81 @@ export function createChatIntentRouter(options = {}) { sessionMessageCount, userMessage, includeIntentPatterns: true, - llmRouterEnabled: false, }); if (ruleResult) return finalizeWithCoercion(ruleResult); - return finalizeWithCoercion(normalizeClassification({ - route: policy.fallbackRoute, - confidence: 0.5, - reason: '规则未命中，走 skill 默认 Agent 通道', - }, { source: 'fallback' })); + + const baseline = buildFallbackClassification(); + if (!llmRouterEligible) { + return finalizeWithCoercion(baseline); + } + + const memoryPromise = policy.memoryResolveEnabled && userId + ? resolveRouterContext({ + userId, + sessionId, + text, + forceDeepReasoning, + }) + .then((result) => buildRouterContext(result)) + .catch(() => buildRouterContext(null)) + : Promise.resolve(buildRouterContext(null)); + + // Memory is optional routing hint; run LLM in parallel to cut serial latency. + const [routerContext, llmResult] = await Promise.all([ + memoryPromise, + classifyWithLlm({ + text, + grantedSkills, + routerContext: null, + }), + ]); + + if (policy.shadowMode) { + logChatLlmRouterShadow({ + baseline, + llmResult, + userId, + sessionId, + routerMemoryItemsUsed: routerContext?.itemsUsed ?? 0, + }, { logger }); + return finalizeWithCoercion({ + ...baseline, + llmShadow: llmResult + ? { + route: llmResult.route, + confidence: llmResult.confidence, + reason: llmResult.reason, + source: llmResult.source ?? 'llm', + wouldChangeRoute: llmResult.route !== baseline.route, + } + : { + skipped: true, + reason: 'llm_unavailable_or_low_signal', + wouldChangeRoute: false, + }, + }); + } + + if (!llmResult || llmResult.confidence < policy.minConfidence) { + return finalizeWithCoercion({ + ...baseline, + llmSuggestion: llmResult + ? { + route: llmResult.route, + confidence: llmResult.confidence, + reason: llmResult.reason, + source: llmResult.source ?? 'llm', + } + : null, + }); + } + + return finalizeWithCoercion(llmResult); } async function classifySessionAction({ text } = {}) { - // The general H5 router intentionally remains rule/skill-only. Session - // action classification is separately opt-in because it runs before every - // WeChat Agent request that contains a context hint. + // WeChat session action classification is separately opt-in because it runs + // before every WeChat Agent request that contains a context hint. if (!policy.enabled || typeof llmProviderService?.createChatCompletion !== 'function') return null; try { const completion = await withTimeout( diff --git a/chat-intent-router.test.mjs b/chat-intent-router.test.mjs index 58bbb6f..b16315a 100644 --- a/chat-intent-router.test.mjs +++ b/chat-intent-router.test.mjs @@ -26,9 +26,14 @@ import { IMAGE_GENERATION_MODE, resolveImageGenerationDecision, coerceRealtimeWebSkill, + isChatLlmRouterEligible, + logChatLlmRouterShadow, resolveChatIntentRouterPolicy, } from './chat-intent-router.mjs'; +/** Ambiguous user text that should miss FAQ/rules and exercise LLM router paths in tests. */ +const AMBIGUOUS_LLM_ROUTER_QUERY = '周末想放松一下，有什么活动建议？'; + test('image generation intent requires a real image for explicit background requests', () => { const decision = resolveImageGenerationDecision({ text: '帮我生成一首唐诗页面，背景是大唐盛景图片，页面要精美', @@ -568,11 +573,11 @@ test('createChatIntentRouter keeps continued sessions on agent without llm route assert.match(result.reason, /延续已有 Agent 会话/); }); -test('createChatIntentRouter uses agent fallback for general conversational analysis (skill-only)', async () => { +test('createChatIntentRouter uses agent fallback for general conversational analysis when llm router disabled', async () => { const llmCalls = []; const resolveCalls = []; const router = createChatIntentRouter({ - enabled: true, + enabled: false, modelProviderKeyId: 'key-router', model: 'deepseek-chat', llmProviderService: { @@ -694,7 +699,7 @@ test('createChatIntentRouter routes memory recall through rules when memory reso assert.equal(result.source, 'rule'); }); -test('createManagedChatIntentRouter hot-loads admin config and stays skill-only', async () => { +test('createManagedChatIntentRouter hot-loads admin config and activates llm routing', async () => { const states = [ { fingerprint: 'off', @@ -743,16 +748,16 @@ test('createManagedChatIntentRouter hot-loads admin config and stays skill-only' const result = await router.classify({ userMessage: { role: 'user', - content: [{ type: 'text', text: '帮我概括一下这段材料的要点' }], - metadata: { displayText: '帮我概括一下这段材料的要点' }, + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, }, }); - assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); - assert.equal(result.source, 'fallback'); - assert.equal(llmCalls.length, 0); + assert.equal(result.route, CHAT_INTENT_ROUTE.DIRECT_CHAT); + assert.equal(result.source, 'llm'); + assert.equal(llmCalls.length, 1); }); -test('createChatIntentRouter uses agent fallback when rules miss (skill-only)', async () => { +test('createChatIntentRouter rejects low-confidence llm direct_chat and falls back to agent', async () => { const llmCalls = []; const router = createChatIntentRouter({ enabled: true, @@ -777,17 +782,17 @@ test('createChatIntentRouter uses agent fallback when rules miss (skill-only)', const result = await router.classify({ userMessage: { role: 'user', - content: [{ type: 'text', text: '帮我整理一下这段材料' }], - metadata: { displayText: '帮我整理一下这段材料' }, + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, }, }); assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); assert.equal(result.source, 'fallback'); - assert.equal(llmCalls.length, 0); + assert.equal(llmCalls.length, 1); }); -test('createChatIntentRouter falls back to agent when rules miss (skill-only)', async () => { +test('createChatIntentRouter falls back to agent when llm router fails', async () => { const llmCalls = []; const router = createChatIntentRouter({ enabled: true, @@ -802,17 +807,53 @@ test('createChatIntentRouter falls back to agent when rules miss (skill-only)', const result = await router.classify({ userMessage: { role: 'user', - content: [{ type: 'text', text: '帮我概括一下这段材料的要点' }], - metadata: { displayText: '帮我概括一下这段材料的要点' }, + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, }, grantedSkills: ['web'], }); + assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); + assert.equal(result.source, 'fallback'); + assert.equal(llmCalls.length, 1); +}); + +test('createChatIntentRouter uses agent fallback when llm router disabled', async () => { + const llmCalls = []; + const router = createChatIntentRouter({ + enabled: false, + llmProviderService: { + async createChatCompletion() { + llmCalls.push(1); + return { ok: false, message: 'router model unavailable' }; + }, + }, + }); + + const result = await router.classify({ + userMessage: { + role: 'user', + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, + }, + }); + assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); assert.equal(result.source, 'fallback'); assert.equal(llmCalls.length, 0); }); +test('classifyWithRules fast-paths FAQ concept questions to direct chat without LLM', () => { + const result = classifyWithRules({ + text: '什么是 SSE？', + sessionId: null, + sessionMessageCount: 0, + }); + assert.equal(result.route, CHAT_INTENT_ROUTE.DIRECT_CHAT); + assert.equal(result.source, 'rule'); + assert.match(result.reason, /FAQ/); +}); + test('classifyWithRules routes memory recall to direct chat on fresh session', () => { const result = classifyWithRules({ text: '你记得我说想去哪儿吗', @@ -932,7 +973,7 @@ test('buildRouterNormalizedDecision maps legacy routes and session hints', () => sessionMessageCount: 0, }, ); - assert.equal(freshAgentSession.session_hint, ROUTER_DECISION_SESSION_HINT.NEW); + assert.equal(freshAgentSession.session_hint, ROUTER_DECISION_SESSION_HINT.REUSE); const memoryRecall = buildRouterNormalizedDecision( { route: CHAT_INTENT_ROUTE.DIRECT_CHAT }, @@ -940,6 +981,12 @@ test('buildRouterNormalizedDecision maps legacy routes and session hints', () => ); assert.equal(memoryRecall.route, ROUTER_DECISION_ROUTE.CHAT); assert.ok(memoryRecall.flags.includes('memory_recall')); + + const goalRun = buildRouterNormalizedDecision( + { route: CHAT_INTENT_ROUTE.AGENT }, + { text: '帮我分阶段完成下季度产品规划' }, + ); + assert.ok(goalRun.flags.includes('goal_run')); }); test('resolveLegacyRouteFromClassification prefers decision when flag enabled', () => { @@ -985,7 +1032,7 @@ test('resolveGatewayAgentSessionId honors session_hint without breaking hard rul decision: { session_hint: ROUTER_DECISION_SESSION_HINT.NEW }, }, }), - null, + 'sess-1', ); assert.equal( resolveGatewayAgentSessionId({ @@ -1370,31 +1417,6 @@ test('logRouterDecisionShadow emits payload only in shadow mode', () => { } }); -test('createChatIntentRouter uses agent fallback for unmatched general questions (skill-only)', async () => { - const llmCalls = []; - const router = createChatIntentRouter({ - enabled: true, - llmProviderService: { - async createChatCompletion() { - llmCalls.push(1); - return { ok: false, message: 'router disabled' }; - }, - }, - }); - - const result = await router.classify({ - userMessage: { - role: 'user', - content: [{ type: 'text', text: '什么是 SSE？' }], - metadata: { displayText: '什么是 SSE？' }, - }, - }); - - assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); - assert.equal(result.source, 'fallback'); - assert.equal(llmCalls.length, 0); -}); - test('classifyWithRules routes page data dev repair before new page-data collect', () => { const result = classifyWithRules({ text: '修复问卷 insert 403，columns_not_allowed', @@ -1405,3 +1427,135 @@ test('classifyWithRules routes page data dev repair before new page-data collect assert.match(result.reason, /Page Data 开发修复/); assert.equal(result.suggested_skill, undefined); }); + +test('isChatLlmRouterEligible respects enabled flag and canary user ids', () => { + assert.equal(isChatLlmRouterEligible({ enabled: false, canaryUserIds: ['john'], userId: 'john' }), false); + assert.equal(isChatLlmRouterEligible({ enabled: true, canaryUserIds: [], userId: 'john' }), true); + assert.equal(isChatLlmRouterEligible({ enabled: true, canaryUserIds: ['john'], userId: 'john' }), true); + assert.equal(isChatLlmRouterEligible({ enabled: true, canaryUserIds: ['john'], userId: 'other' }), false); + assert.equal(isChatLlmRouterEligible({ enabled: true, canaryUserIds: ['john'], userId: null }), false); +}); + +test('createChatIntentRouter limits llm routing to canary users', async () => { + const llmCalls = []; + const router = createChatIntentRouter({ + enabled: true, + canaryUserIds: ['user-canary'], + llmProviderService: { + async createChatCompletion() { + llmCalls.push(1); + return { + ok: true, + reply: JSON.stringify({ + route: 'direct_chat', + confidence: 0.91, + reason: '纯文字问答', + suggested_skill: null, + agent_brief: '', + }), + }; + }, + }, + }); + + const control = await router.classify({ + userId: 'user-control', + userMessage: { + role: 'user', + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, + }, + }); + const canary = await router.classify({ + userId: 'user-canary', + userMessage: { + role: 'user', + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, + }, + }); + + assert.equal(control.route, CHAT_INTENT_ROUTE.AGENT); + assert.equal(control.source, 'fallback'); + assert.equal(canary.route, CHAT_INTENT_ROUTE.DIRECT_CHAT); + assert.equal(canary.source, 'llm'); + assert.equal(llmCalls.length, 1); +}); + +test('createChatIntentRouter shadow mode logs llm suggestion without changing route', async () => { + const lines = []; + const logger = { warn(line) { lines.push(line); }, info(line) { lines.push(line); } }; + const router = createChatIntentRouter({ + enabled: true, + shadowMode: true, + llmProviderService: { + async createChatCompletion() { + return { + ok: true, + reply: JSON.stringify({ + route: 'direct_chat', + confidence: 0.95, + reason: '纯文字问答', + suggested_skill: null, + agent_brief: '', + }), + }; + }, + }, + logger, + }); + + const result = await router.classify({ + userId: 'user-john', + userMessage: { + role: 'user', + content: [{ type: 'text', text: AMBIGUOUS_LLM_ROUTER_QUERY }], + metadata: { displayText: AMBIGUOUS_LLM_ROUTER_QUERY }, + }, + }); + + assert.equal(result.route, CHAT_INTENT_ROUTE.AGENT); + assert.equal(result.source, 'fallback'); + assert.equal(result.llmShadow?.route, CHAT_INTENT_ROUTE.DIRECT_CHAT); + assert.equal(result.llmShadow?.wouldChangeRoute, true); + assert.equal(lines.length, 1); + assert.match(lines[0], /\[chat-llm-router-shadow\]/); + assert.match(lines[0], /wouldChangeRoute":true/); +}); + +test('logChatLlmRouterShadow emits structured payload', () => { + const lines = []; + logChatLlmRouterShadow({ + baseline: { route: CHAT_INTENT_ROUTE.AGENT, source: 'fallback' }, + llmResult: { route: CHAT_INTENT_ROUTE.DIRECT_CHAT, confidence: 0.9, reason: '闲聊' }, + userId: 'user-1', + sessionId: 'sess-1', + }, { logger: { info(line) { lines.push(line); } } }); + assert.equal(lines.length, 1); + assert.match(lines[0], /baselineRoute":"agent_orchestration"/); + assert.match(lines[0], /llmRoute":"direct_chat"/); +}); + +test('resolveChatIntentRouterPolicy parses llm router shadow and canary env', () => { + const policy = resolveChatIntentRouterPolicy({ + env: { + MEMIND_CHAT_LLM_ROUTER_ENABLED: '1', + MEMIND_CHAT_LLM_ROUTER_SHADOW: '1', + MEMIND_CHAT_ROUTER_CANARY_USER_IDS: 'john, testuser1', + }, + }); + assert.equal(policy.enabled, true); + assert.equal(policy.shadowMode, true); + assert.deepEqual(policy.canaryUserIds, ['john', 'testuser1']); +}); + +test('createChatIntentRouter getStatus reflects llm router modes', () => { + const active = createChatIntentRouter({ enabled: true }); + const shadow = createChatIntentRouter({ enabled: true, shadowMode: true }); + const off = createChatIntentRouter({ enabled: false }); + assert.equal(active.getStatus().llmRoutingEnabled, true); + assert.equal(active.getStatus().llmRoutingShadow, false); + assert.equal(shadow.getStatus().llmRoutingEnabled, false); + assert.equal(shadow.getStatus().llmRoutingShadow, true); + assert.equal(off.getStatus().configuredLlmEnabled, false); +}); diff --git a/goal-run-intent.mjs b/goal-run-intent.mjs new file mode 100644 index 0000000..44652f6 --- /dev/null +++ b/goal-run-intent.mjs @@ -0,0 +1,76 @@ +export const GOAL_RUN_INTENT_PATTERNS = [ + /(?:长期目标|当前目标|目标是|计划要|正在建设|准备实现|希望长期)/i, + /(?:分阶段|多步骤|长期任务|持续跟进|帮我规划并执行)/i, +]; + +export function isGoalRunIntent(text) { + const normalized = String(text ?? '').trim(); + if (!normalized) return false; + return GOAL_RUN_INTENT_PATTERNS.some((pattern) => pattern.test(normalized)); +} + +function envFlag(value, fallback = false) { + const raw = String(value ?? '').trim().toLowerCase(); + if (!raw) return fallback; + return ['1', 'true', 'yes', 'on'].includes(raw); +} + +function parseUserIdSet(raw) { + return new Set( + String(raw ?? '') + .split(/[,;\s]+/) + .map((item) => item.trim()) + .filter(Boolean), + ); +} + +export function isGoalRunEnabledForUser(userId, env = process.env) { + if (!envFlag(env.GOAL_RUN_ENABLED, false)) return false; + const canaryUserIds = parseUserIdSet(env.GOAL_RUN_CANARY_USER_IDS); + if (!canaryUserIds.size) return true; + return canaryUserIds.has(String(userId ?? '').trim()); +} + +export function extractUserMessageDisplayText(userMessage) { + if (!userMessage) return ''; + if (typeof userMessage === 'string') return userMessage.trim(); + const metadataText = String(userMessage?.metadata?.displayText ?? '').trim(); + if (metadataText) return metadataText; + const content = userMessage?.content; + if (typeof content === 'string') return content.trim(); + if (Array.isArray(content)) { + for (const item of content) { + if (typeof item === 'string' && item.trim()) return item.trim(); + if (item?.type === 'text' && String(item.text ?? '').trim()) { + return String(item.text).trim(); + } + } + } + return ''; +} + +export function deriveGoalTitleFromText(text, maxLength = 80) { + const normalized = String(text ?? '').replace(/\s+/g, ' ').trim(); + if (!normalized) return '长期任务'; + return normalized.length <= maxLength + ? normalized + : `${normalized.slice(0, Math.max(0, maxLength - 1))}…`; +} + +export function buildGoalContinueUserMessage(feedback = null, { id = null } = {}) { + const displayText = String(feedback ?? '').trim() + ? `继续下一阶段。补充说明：${String(feedback).trim()}` + : '继续执行下一阶段'; + const messageId = id ?? (typeof crypto !== 'undefined' && crypto.randomUUID + ? crypto.randomUUID() + : `goal-continue-${Date.now()}`); + return { + id: messageId, + role: 'user', + content: [{ type: 'text', text: displayText }], + metadata: { + userVisible: true, + displayText, + }, + }; +} diff --git a/goal-run-intent.test.mjs b/goal-run-intent.test.mjs new file mode 100644 index 0000000..6351e5d --- /dev/null +++ b/goal-run-intent.test.mjs @@ -0,0 +1,60 @@ +import assert from 'node:assert/strict'; +import test from 'node:test'; +import { buildGoalContextEnvelope } from './goal-run-context.mjs'; +import { + buildGoalContinueUserMessage, + deriveGoalTitleFromText, + isGoalRunEnabledForUser, + isGoalRunIntent, +} from './goal-run-intent.mjs'; + +test('isGoalRunIntent detects long-running task phrases', () => { + assert.equal(isGoalRunIntent('帮我分阶段完成下季度产品规划'), true); + assert.equal(isGoalRunIntent('你好'), false); +}); + +test('isGoalRunEnabledForUser respects feature flag and canary list', () => { + assert.equal(isGoalRunEnabledForUser('user-a', { GOAL_RUN_ENABLED: '0' }), false); + assert.equal(isGoalRunEnabledForUser('user-a', { GOAL_RUN_ENABLED: '1' }), true); + assert.equal( + isGoalRunEnabledForUser('user-a', { + GOAL_RUN_ENABLED: '1', + GOAL_RUN_CANARY_USER_IDS: 'user-b', + }), + false, + ); + assert.equal( + isGoalRunEnabledForUser('user-a', { + GOAL_RUN_ENABLED: '1', + GOAL_RUN_CANARY_USER_IDS: 'user-a,user-b', + }), + true, + ); +}); + +test('buildGoalContextEnvelope renders friendly goal summary', () => { + const envelope = buildGoalContextEnvelope({ + title: '准备下季度产品规划', + currentCheckpointId: 'cp-1', + checkpoints: [ + { id: 'cp-1', title: '启动', status: 'running', outputSummary: null }, + { id: 'cp-0', title: '调研', status: 'approved', outputSummary: '完成竞品列表' }, + ], + }); + assert.match(envelope, /【Goal Context】/); + assert.match(envelope, /准备下季度产品规划/); + assert.match(envelope, /完成竞品列表/); +}); + +test('deriveGoalTitleFromText truncates long titles', () => { + const title = deriveGoalTitleFromText('a'.repeat(120), 20); + assert.equal(title.length, 20); + assert.match(title, /…$/); +}); + +test('buildGoalContinueUserMessage renders continue text', () => { + const message = buildGoalContinueUserMessage('先聚焦竞品', { id: 'msg-1' }); + assert.equal(message.id, 'msg-1'); + assert.match(message.metadata.displayText, /继续下一阶段/); + assert.match(message.metadata.displayText, /竞品/); +}); diff --git a/goosed-proxy-boundary.mjs b/goosed-proxy-boundary.mjs index 99215b6..154dce0 100644 --- a/goosed-proxy-boundary.mjs +++ b/goosed-proxy-boundary.mjs @@ -70,6 +70,7 @@ export const GOOSED_BOUNDARY_SCAN_IGNORE = [ /^scripts\//, /^mindspace-service\//, /^\.runtime\//, + /^\.release-gate\//, /^node_modules\//, /\.test\.mjs$/, /\.test\.ts$/, diff --git a/memory-v2-admin-config.mjs b/memory-v2-admin-config.mjs index 226bfd0..721b1a7 100644 --- a/memory-v2-admin-config.mjs +++ b/memory-v2-admin-config.mjs @@ -13,6 +13,8 @@ const FIELD_SPECS = [ { env: 'MEMORY_FAIL_OPEN', group: 'global', field: 'failOpen', type: 'boolean' }, { env: 'MEMIND_CHAT_LLM_ROUTER_ENABLED', group: 'chatIntentRouter', field: 'enabled', type: 'boolean' }, + { env: 'MEMIND_CHAT_LLM_ROUTER_SHADOW', group: 'chatIntentRouter', field: 'shadowMode', type: 'boolean' }, + { env: 'MEMIND_CHAT_ROUTER_CANARY_USER_IDS', group: 'chatIntentRouter', field: 'canaryUserIds', type: 'string' }, { env: 'MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID', group: 'chatIntentRouter', field: 'modelProviderKeyId', type: 'string' }, { env: 'MEMIND_CHAT_ROUTER_MODEL', group: 'chatIntentRouter', field: 'model', type: 'string' }, { env: 'MEMIND_CHAT_ROUTER_MODEL_API', group: 'chatIntentRouter', field: 'modelApiType', type: 'string' }, @@ -28,6 +30,7 @@ const FIELD_SPECS = [ { env: 'MEMORY_CANDIDATE_MIN_CONFIDENCE', group: 'candidateMemory', field: 'minConfidence', type: 'number' }, { env: 'MEMORY_CANDIDATE_MAX_PENDING', group: 'candidateMemory', field: 'maxPending', type: 'number' }, { env: 'MEMORY_CANDIDATE_PERSISTENCE_ENABLED', group: 'candidateMemory', field: 'persistenceEnabled', type: 'boolean' }, + { env: 'MEMORY_CANDIDATE_AUTO_ACCEPT_ALL', group: 'candidateMemory', field: 'autoAcceptAll', type: 'boolean' }, { env: 'MEMORY_AGENT_RESOLVE_ENABLED', group: 'runtimeControl', field: 'agentResolveEnabled', type: 'boolean' }, { env: 'MEMORY_AGENT_INJECTION_MODE', group: 'runtimeControl', field: 'agentInjectionMode', type: 'string' }, @@ -283,7 +286,7 @@ function applyEnv(config, secrets, env = process.env) { config.chatIntentRouter.memoryResolveLimit = '8'; } if (!normalizeString(config.chatIntentRouter.timeoutMs)) { - config.chatIntentRouter.timeoutMs = '2500'; + config.chatIntentRouter.timeoutMs = '1200'; } if (config.chatIntentRouter.enabled === undefined) { config.chatIntentRouter.enabled = false; diff --git a/ops/src/App.tsx b/ops/src/App.tsx index becefaf..ef3257e 100644 --- a/ops/src/App.tsx +++ b/ops/src/App.tsx @@ -15,6 +15,7 @@ import { AgentCodeRunPage } from './pages/admin/AgentCodeRunPage'; import { WechatPage } from './pages/admin/WechatPage'; import { SystemPolicyPage } from './pages/admin/SystemPolicyPage'; import { OrchestratorPage } from './pages/admin/OrchestratorPage'; +import { LlmProvidersPage } from './pages/admin/LlmProvidersPage'; export function App() { return ( @@ -50,6 +51,7 @@ export function App() { } /> } /> } /> + } /> } /> diff --git a/ops/src/api/admin.ts b/ops/src/api/admin.ts index 8109abb..3fc3b87 100644 --- a/ops/src/api/admin.ts +++ b/ops/src/api/admin.ts @@ -815,3 +815,81 @@ export async function fetchAgentCodeRunHistory(limit = 50) { `/admin-api/agent-code-run/runs?limit=${limit}`, ); } + +// ─── Unified LLM Providers ──────────────────────────────────────────────────── + +export type LlmProviderKeyRow = { + id: string; + providerId: string; + providerKind: 'builtin' | 'custom'; + providerLabel: string; + name: string; + defaultModel: string; + models: string[]; + apiUrl: string | null; + status: 'active' | 'disabled'; + isSelected: boolean; + apiKeyMasked: string; +}; + +export type LlmGlobalSettings = { + keyId: string | null; + keyName: string | null; + providerLabel: string | null; + globalModel: string | null; + availableModels: string[]; +}; + +export type ChatIntentRouterAdminConfig = { + enabled?: boolean; + shadowMode?: boolean; + canaryUserIds?: string; + modelProviderKeyId?: string; + model?: string; + modelApiType?: string; + minConfidence?: string; + memoryResolveEnabled?: boolean; + memoryResolveLimit?: string; + timeoutMs?: string; + fallbackRoute?: string; +}; + +export type MemoryV2AdminConfigState = { + config: { + chatIntentRouter?: ChatIntentRouterAdminConfig; + [key: string]: unknown; + }; + updatedAt: number | null; + updatedBy: string | null; +}; + +export type MemoryV2RuntimeState = { + source: string; + updatedAt: number | null; + updatedBy: string | null; + fingerprint?: string; + overrides: Record; +}; + +export async function fetchLlmProviderKeys() { + return adminFetch<{ keys: LlmProviderKeyRow[] }>('/admin-api/llm-providers/keys'); +} + +export async function fetchLlmGlobalSettings() { + return adminFetch<{ global: LlmGlobalSettings }>('/admin-api/llm-providers/global'); +} + +export async function fetchMemoryV2Config() { + return adminFetch('/admin-api/memory-v2/config'); +} + +export async function patchMemoryV2Config(patch: Record) { + return adminFetch('/admin-api/memory-v2/config', { + method: 'PATCH', + body: JSON.stringify(patch), + }); +} + +export async function fetchMemoryV2Runtime() { + return adminFetch('/admin-api/memory-v2/runtime'); +} diff --git a/ops/src/components/AdminLayout.tsx b/ops/src/components/AdminLayout.tsx index 715d963..1ff01d7 100644 --- a/ops/src/components/AdminLayout.tsx +++ b/ops/src/components/AdminLayout.tsx @@ -8,6 +8,7 @@ const links = [ { to: '/admin/agent-code-run', label: 'Code Run' }, { to: '/admin/policy', label: '策略中心' }, { to: '/admin/orchestrator', label: '任务编排' }, + { to: '/admin/llm', label: '统一大模型' }, ]; export function AdminLayout() { diff --git a/ops/src/pages/admin/LlmProvidersPage.tsx b/ops/src/pages/admin/LlmProvidersPage.tsx new file mode 100644 index 0000000..4a12f65 --- /dev/null +++ b/ops/src/pages/admin/LlmProvidersPage.tsx @@ -0,0 +1,367 @@ +import { useCallback, useEffect, useMemo, useState } from 'react'; +import { + fetchLlmGlobalSettings, + fetchLlmProviderKeys, + fetchMemoryV2Config, + fetchMemoryV2Runtime, + patchMemoryV2Config, + type ChatIntentRouterAdminConfig, + type LlmGlobalSettings, + type LlmProviderKeyRow, + type MemoryV2RuntimeState, +} from '../../api/admin'; + +type RouterForm = { + enabled: boolean; + shadowMode: boolean; + modelProviderKeyId: string; + model: string; + canaryUserIds: string; +}; + +function formatTime(value: number | null | undefined) { + if (!value) return '—'; + return new Date(value).toLocaleString('zh-CN', { hour12: false }); +} + +function routerConfigToForm( + config: ChatIntentRouterAdminConfig | undefined, + keys: LlmProviderKeyRow[], +): RouterForm { + const keyId = String(config?.modelProviderKeyId ?? '').trim(); + const selectedKey = keys.find((item) => item.id === keyId) ?? null; + const model = String(config?.model ?? '').trim() + || selectedKey?.defaultModel + || selectedKey?.models?.[0] + || ''; + return { + enabled: Boolean(config?.enabled), + shadowMode: Boolean(config?.shadowMode), + modelProviderKeyId: keyId, + model, + canaryUserIds: String(config?.canaryUserIds ?? '').trim(), + }; +} + +function formToPatch(form: RouterForm) { + return { + chatIntentRouter: { + enabled: form.enabled, + shadowMode: form.shadowMode, + modelProviderKeyId: form.modelProviderKeyId || '', + model: form.model || '', + canaryUserIds: form.canaryUserIds, + }, + }; +} + +function ToggleRow({ + label, + hint, + checked, + onChange, + disabled, +}: { + label: string; + hint?: string; + checked: boolean; + onChange: (next: boolean) => void; + disabled?: boolean; +}) { + return ( + + ); +} + +function runtimeMode(overrides: Record) { + const enabled = ['1', 'true', 'yes', 'on'].includes(String(overrides.MEMIND_CHAT_LLM_ROUTER_ENABLED ?? '').toLowerCase()); + const shadow = ['1', 'true', 'yes', 'on'].includes(String(overrides.MEMIND_CHAT_LLM_ROUTER_SHADOW ?? '').toLowerCase()); + if (!enabled) return '关闭（仅规则路由）'; + if (shadow) return 'Shadow 观测（行为不变）'; + return '已激活（规则未命中时走 LLM 路由）'; +} + +export function LlmProvidersPage() { + const [keys, setKeys] = useState([]); + const [globalSettings, setGlobalSettings] = useState(null); + const [form, setForm] = useState(null); + const [savedForm, setSavedForm] = useState(null); + const [runtime, setRuntime] = useState(null); + const [updatedAt, setUpdatedAt] = useState(null); + const [loading, setLoading] = useState(true); + const [saving, setSaving] = useState(false); + const [error, setError] = useState(null); + const [notice, setNotice] = useState(null); + + const activeKeys = useMemo( + () => keys.filter((item) => item.status === 'active'), + [keys], + ); + + const selectedRouterKey = useMemo( + () => activeKeys.find((item) => item.id === form?.modelProviderKeyId) ?? null, + [activeKeys, form?.modelProviderKeyId], + ); + + const modelOptions = useMemo( + () => selectedRouterKey?.models ?? [], + [selectedRouterKey], + ); + + const dirty = useMemo(() => { + if (!form || !savedForm) return false; + return JSON.stringify(form) !== JSON.stringify(savedForm); + }, [form, savedForm]); + + const load = useCallback(async () => { + setLoading(true); + setError(null); + try { + const [keysResult, globalResult, configResult, runtimeResult] = await Promise.all([ + fetchLlmProviderKeys(), + fetchLlmGlobalSettings(), + fetchMemoryV2Config(), + fetchMemoryV2Runtime(), + ]); + const nextKeys = keysResult.keys ?? []; + const nextForm = routerConfigToForm(configResult.config?.chatIntentRouter, nextKeys); + setKeys(nextKeys); + setGlobalSettings(globalResult.global ?? null); + setForm(nextForm); + setSavedForm(nextForm); + setRuntime(runtimeResult); + setUpdatedAt(configResult.updatedAt ?? null); + } catch (err) { + setError(err instanceof Error ? err.message : '加载失败'); + } finally { + setLoading(false); + } + }, []); + + useEffect(() => { + void load(); + }, [load]); + + const updateForm = (patch: Partial) => { + setForm((current) => { + if (!current) return current; + const next = { ...current, ...patch }; + if ('modelProviderKeyId' in patch) { + const key = activeKeys.find((item) => item.id === next.modelProviderKeyId); + const models = key?.models ?? []; + if (!models.includes(next.model)) { + next.model = key?.defaultModel ?? models[0] ?? ''; + } + } + return next; + }); + }; + + const save = async () => { + if (!form) return; + setSaving(true); + setError(null); + setNotice(null); + try { + const result = await patchMemoryV2Config(formToPatch(form)); + const nextForm = routerConfigToForm(result.config?.chatIntentRouter, keys); + setForm(nextForm); + setSavedForm(nextForm); + setUpdatedAt(result.updatedAt ?? null); + const runtimeResult = await fetchMemoryV2Runtime(); + setRuntime(runtimeResult); + setNotice('已保存。Portal 将在下一次 H5 路由请求时热加载新配置，无需重启。'); + } catch (err) { + setError(err instanceof Error ? err.message : '保存失败'); + } finally { + setSaving(false); + } + }; + + if (loading) return

加载中…

; + if (error && !form) return

{error}

; + if (!form) return

配置不可用

; + + return ( +
+
+

统一大模型

+

+ 管理平台 LLM 密钥与全局默认模型。H5 意图路由可单独绑定密钥与模型，不再依赖本地默认配置。 +

+ {globalSettings ? ( +

+ 当前全局默认： + {' '} + {globalSettings.keyName ?? '（未选择密钥）'} + {' / '} + {globalSettings.globalModel ?? '（未设置模型）'} +

+ ) : null} +
+ +
+

已配置密钥

+ {activeKeys.length === 0 ? ( +

暂无可用密钥，请先在统一模型中心录入 Provider Key。

+ ) : ( + + + + + + + + + + + {activeKeys.map((item) => ( + + + + + + + ))} + +
名称Provider默认模型状态
+ {item.name} + {item.isSelected ? 全局选中 : null} + {item.providerLabel}{item.defaultModel}{item.status}
+ )} +
+ +
+

H5 意图路由（LLM Router）

+

+ 规则未命中时的轻量 LLM 分流。Shadow 模式只记录 `[chat-llm-router-shadow]` 与 `intent_routed.llmShadow`，不改变用户侧路由。 +

+ +
+ updateForm({ enabled })} + /> + updateForm({ shadowMode })} + /> + + + + + + +
+ +
+ + + + 最近保存： + {formatTime(updatedAt)} + +
+ {notice ?

{notice}

: null} + {error ?

{error}

: null} +
+ + {runtime ? ( +
+

运行时生效状态

+

+ 来源： + {runtime.source} + {' · '} + 模式： + {runtimeMode(runtime.overrides ?? {})} +

+
+
+ 绑定密钥： + {runtime.overrides.MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID || '（未设置）'} +
+
+ 绑定模型： + {runtime.overrides.MEMIND_CHAT_ROUTER_MODEL || '（未设置）'} +
+
+ 灰度用户： + {runtime.overrides.MEMIND_CHAT_ROUTER_CANARY_USER_IDS || '（全部用户）'} +
+
+ fingerprint： + {runtime.fingerprint ?? '—'} +
+
+
+ ) : null} +
+ ); +} diff --git a/ops/src/pages/admin/SummaryPage.tsx b/ops/src/pages/admin/SummaryPage.tsx index cd66d12..13c97e1 100644 --- a/ops/src/pages/admin/SummaryPage.tsx +++ b/ops/src/pages/admin/SummaryPage.tsx @@ -68,11 +68,10 @@ export function SummaryPage() { Code Run 配置 - + + 统一大模型 / H5 路由 + + 进入用户管理 diff --git a/scripts/verify-h5-llm-router-shadow.mjs b/scripts/verify-h5-llm-router-shadow.mjs new file mode 100644 index 0000000..1263cd3 --- /dev/null +++ b/scripts/verify-h5-llm-router-shadow.mjs @@ -0,0 +1,201 @@ +#!/usr/bin/env node +/** + * Local smoke: verify ambiguous H5 chat triggers LLM router shadow logging. + * Usage: node scripts/verify-h5-llm-router-shadow.mjs + */ +import crypto from 'node:crypto'; +import { loadH5Environment } from './load-env.mjs'; +import { createDbPool } from '../db.mjs'; +import { createUserAuth } from '../user-auth.mjs'; +import { resolveChatIntentRouterPolicy } from '../chat-intent-router.mjs'; + +loadH5Environment(import.meta.dirname); + +const PORTAL = `http://127.0.0.1:${process.env.H5_PORT ?? 8081}`; +const USERNAME = process.env.VERIFY_LLM_ROUTER_USER ?? 'john2'; +const PASSWORD = process.env.JOHN_PASSWORD ?? process.env.H5_ACCESS_PASSWORD ?? '888888'; +const QUERY = '周末想放松一下，有什么活动建议？'; +const MAX_WAIT_MS = 120_000; +const LOG_PATH = process.env.MEMIND_LLM_ROUTER_SHADOW_LOG + ?? '/Users/john/Library/Logs/memind-h5-local.log'; + +function sleep(ms) { + return new Promise((resolve) => setTimeout(resolve, ms)); +} + +function pass(label, detail = '') { + console.log(`PASS ${label}${detail ? `: ${detail}` : ''}`); +} + +function fail(label, detail = '') { + console.error(`FAIL ${label}${detail ? `: ${detail}` : ''}`); + process.exitCode = 1; +} + +async function login() { + const pool = await createDbPool(); + const auth = createUserAuth(pool); + const result = await auth.login({ username: USERNAME, password: PASSWORD, ip: '127.0.0.1' }); + if (!result.ok) { + throw new Error(`john 登录失败: ${result.message ?? 'unknown'}`); + } + await pool.end(); + return result.token; +} + +async function createRun(token) { + const requestId = crypto.randomUUID(); + const response = await fetch(`${PORTAL}/api/agent/runs`, { + method: 'POST', + headers: { + 'Content-Type': 'application/json', + Cookie: `tkmind_user_session=${token}`, + }, + body: JSON.stringify({ + request_id: requestId, + user_message: { + id: crypto.randomUUID(), + role: 'user', + content: [{ type: 'text', text: QUERY }], + metadata: { userVisible: true, displayText: QUERY }, + }, + }), + }); + const body = await response.json().catch(() => ({})); + if (!response.ok) { + throw new Error(`POST /agent/runs ${response.status}: ${JSON.stringify(body)}`); + } + const run = body.run ?? body; + return { runId: run.id, requestId, status: run.status }; +} + +async function waitForRun(token, runId) { + const started = Date.now(); + while (Date.now() - started < MAX_WAIT_MS) { + const response = await fetch(`${PORTAL}/api/agent/runs/${runId}`, { + headers: { Cookie: `tkmind_user_session=${token}` }, + }); + const payload = await response.json().catch(() => ({})); + const body = payload.run ?? payload; + if (!response.ok) { + throw new Error(`GET /agent/runs/${runId} ${response.status}: ${JSON.stringify(payload)}`); + } + if (['succeeded', 'failed'].includes(body.status)) { + return body; + } + await sleep(1000); + } + throw new Error(`run ${runId} 未在 ${MAX_WAIT_MS}ms 内完成`); +} + +async function loadRunEvents(runId) { + const pool = await createDbPool(); + const [rows] = await pool.query( + `SELECT event_type, data_json FROM h5_agent_run_events WHERE run_id = ? ORDER BY id ASC`, + [runId], + ); + await pool.end(); + return rows.map((row) => ({ + event_type: row.event_type, + data: typeof row.data_json === 'string' ? JSON.parse(row.data_json) : row.data_json, + })); +} + +async function waitForPortal() { + const started = Date.now(); + while (Date.now() - started < 60_000) { + try { + const response = await fetch(`${PORTAL}/api/status`); + if (response.ok) return; + } catch { + // retry + } + await sleep(1000); + } + throw new Error(`Portal ${PORTAL} 未在 60s 内就绪`); +} + +async function readShadowLogLines() { + try { + const fs = await import('node:fs/promises'); + const text = await fs.readFile(LOG_PATH, 'utf8'); + return text.split('\n').filter((line) => line.includes('[chat-llm-router-shadow]')); + } catch { + return []; + } +} + +async function main() { + const policy = resolveChatIntentRouterPolicy(); + console.log('--- router policy ---'); + console.log(JSON.stringify({ + enabled: policy.enabled, + shadowMode: policy.shadowMode, + canaryUserIds: policy.canaryUserIds, + timeoutMs: policy.timeoutMs, + }, null, 2)); + + if (!policy.enabled || !policy.shadowMode) { + fail('policy', '需要 MEMIND_CHAT_LLM_ROUTER_ENABLED=1 且 MEMIND_CHAT_LLM_ROUTER_SHADOW=1'); + return; + } + pass('policy', 'shadow 模式已开启'); + + await waitForPortal(); + pass('portal', PORTAL); + + const token = await login(); + pass('login', USERNAME); + + const beforeShadowLines = await readShadowLogLines(); + const { runId } = await createRun(token); + pass('agent run', runId); + + await waitForRun(token, runId); + pass('run finished'); + + const events = await loadRunEvents(runId); + const routed = events.find((row) => row.event_type === 'intent_routed'); + if (!routed?.data) { + fail('intent_routed', '未找到路由事件'); + } else { + const { route, source, reason, llmShadow } = routed.data; + console.log('\n--- intent_routed ---'); + console.log(JSON.stringify(routed.data, null, 2)); + if (source === 'fallback' && route === 'agent_orchestration') { + pass('shadow routing', '规则未命中，Shadow 仍 fallback Agent'); + } else if (source === 'rule') { + fail('shadow routing', `意外命中规则: ${reason ?? route}`); + } else { + fail('shadow routing', `source=${source}, route=${route}`); + } + if (llmShadow) { + pass('llm shadow payload', JSON.stringify(llmShadow)); + } else { + fail('llm shadow payload', 'intent_routed 缺少 llmShadow（LLM 路由可能未启用）'); + } + } + + await sleep(500); + const afterShadowLines = await readShadowLogLines(); + const newShadowLines = afterShadowLines.slice(beforeShadowLines.length); + if (newShadowLines.length > 0) { + pass('shadow log', `${newShadowLines.length} 条 [chat-llm-router-shadow]`); + for (const line of newShadowLines.slice(-3)) { + console.log(line.trim()); + } + } else if (!process.exitCode) { + console.log('NOTE shadow log 未写入 LaunchAgent 日志，但 intent_routed.llmShadow 已记录观测结果'); + } + + if (process.exitCode && process.exitCode !== 0) { + console.error('\n验证未完全通过。'); + } else { + console.log('\n验证通过：Shadow 模式已观测 LLM 路由建议，用户侧仍走 Agent fallback。'); + } +} + +main().catch((err) => { + console.error(err instanceof Error ? err.stack ?? err.message : err); + process.exit(1); +});