Open source file-system and knowledge based agent template. Build AI agents that stay up to date with your knowledge base
improve ui and generate agent options
Hugo Richard74be21a31650
diff --git a/apps/chat/app/components/ChatLoading.vue b/apps/chat/app/components/ChatLoading.vueindex 27ad911..be448eb 100644--- a/apps/chat/app/components/ChatLoading.vue+++ b/apps/chat/app/components/ChatLoading.vue@@ -1,11 +1,28 @@<script setup lang="ts">import { motion } from 'motion-v'+import { useIntervalFn, useRafFn } from '@vueuse/core'++interface CommandResult {+ command: string+ stdout: string+ stderr: string+ exitCode: number+ success: boolean+}++interface ToolExecutionResult {+ commands: CommandResult[]+ success: boolean+ durationMs: number+ error?: string+}interface ToolCall {toolCallId: stringtoolName: stringargs: Record<string, unknown>- state: 'loading' | 'done'+ state: 'loading' | 'done' | 'error'+ result?: ToolExecutionResult}const props = defineProps<{@@ -15,116 +32,89 @@ const props = defineProps<{}>()const messages = [- 'Searching the documentation',- 'Reading through the docs',+ 'Exploring the sources',+ 'Searching for content','Looking for relevant files',- 'Analyzing the content',+ 'Analyzing the results','Finding the best answer','Checking related resources','Almost there',]-const finishedMessage = 'Search done'+const finishedMessage = 'Done'+const chars = 'abcdefghijklmnopqrstuvwxyz'const currentIndex = ref(0)+const displayedText = ref('')+const targetText = computed(() => {- if (!props.isLoading) {- return finishedMessage- }+ if (!props.isLoading) return finishedMessagereturn props.text || messages[currentIndex.value]})-const displayedText = ref(targetText.value)-const chars = 'abcdefghijklmnopqrstuvwxyz'+let frame = 0+let totalFrames = 15+let fromText = ''+let toText = ''-function scrambleText(from: string, to: string) {- const maxLength = Math.max(from.length, to.length)- let frame = 0- const totalFrames = 15-- const animate = () => {- frame++- let result = ''-- for (let i = 0; i < maxLength; i++) {- const progress = frame / totalFrames- const charProgress = progress * maxLength-- if (i < charProgress - 2) {- result += to[i] || ''- } else if (i < charProgress) {- result += chars[Math.floor(Math.random() * chars.length)]- } else {- result += from[i] || ''- }- }+const { pause: pauseScramble, resume: resumeScramble } = useRafFn(() => {+ frame+++ const maxLength = Math.max(fromText.length, toText.length)+ let result = ''- displayedText.value = result+ for (let i = 0; i < maxLength; i++) {+ const progress = frame / totalFrames+ const charProgress = progress * maxLength- if (frame < totalFrames) {- requestAnimationFrame(animate)+ if (i < charProgress - 2) {+ result += toText[i] || ''+ } else if (i < charProgress) {+ result += chars[Math.floor(Math.random() * chars.length)]} else {- displayedText.value = to+ result += fromText[i] || ''}}- requestAnimationFrame(animate)-}+ displayedText.value = result++ if (frame >= totalFrames) {+ displayedText.value = toText+ pauseScramble()+ }+}, { immediate: false })-let textInterval: ReturnType<typeof setInterval> | null = null+function scrambleText(from: string, to: string) {+ fromText = from+ toText = to+ frame = 0+ resumeScramble()+}watch(targetText, (newText, oldText) => {if (newText !== oldText && newText && oldText) {scrambleText(oldText, newText)}-})+}, { immediate: true })-// Stop text rotation when loading finishes-watch(() => props.isLoading, (isLoading) => {- if (!isLoading && textInterval) {- clearInterval(textInterval)- textInterval = null- }+onMounted(() => {+ displayedText.value = targetText.value || ''})-function getToolLabel(tool: ToolCall) {- const { toolName, args, state } = tool- const isDone = state === 'done'-- if (toolName === 'search_and_read') {- return isDone- ? `Searched "${args?.query || '...'}"`- : `Searching "${args?.query || '...'}"`- }-- if (toolName === 'read') {- const paths = args?.paths as string[] | undefined- const path = paths?.[0] || '...'- return isDone- ? `Read ${path}`- : `Reading ${path}`- }-- return toolName-}+const { pause, resume } = useIntervalFn(() => {+ currentIndex.value = (currentIndex.value + 1) % messages.length+}, 3500, { immediate: false })-onMounted(() => {- // Text rotation only when loading- if (!props.text && props.isLoading) {- textInterval = setInterval(() => {- currentIndex.value = (currentIndex.value + 1) % messages.length- }, 3500)+watch(() => props.isLoading, (isLoading) => {+ if (isLoading && !props.text) {+ resume()+ } else {+ pause()}-})--onUnmounted(() => {- if (textInterval) clearInterval(textInterval)-})+}, { immediate: true })</script><template><div class="flex flex-col gap-2 mb-4">- <!-- Main loader with matrix and text --><div class="flex items-center text-sm text-muted overflow-hidden"><motion.divv-if="isLoading"@@ -144,25 +134,16 @@ onUnmounted(() => {</motion.span></div>- <!-- Tool calls displayed below --><divv-if="toolCalls?.length"- class="flex flex-col gap-1.5"+ class="flex flex-col gap-0.5":class="isLoading ? 'pl-[31px]' : 'pl-0'">- <motion.div+ <ToolCallItemv-for="tool in toolCalls":key="tool.toolCallId"- :initial="{ opacity: 0, x: -4 }"- :animate="{ opacity: 1, x: 0 }"- :transition="{ duration: 0.15 }"- class="flex items-center gap-2"- >- <span class="size-1.5 rounded-full bg-current opacity-40" />- <span class="text-xs text-dimmed truncate max-w-[300px]">- {{ getToolLabel(tool) }}- </span>- </motion.div>+ :tool+ /></div></div></template>diff --git a/apps/chat/app/components/ChatMatrix.vue b/apps/chat/app/components/ChatMatrix.vueindex 3d25609..443742d 100644--- a/apps/chat/app/components/ChatMatrix.vue+++ b/apps/chat/app/components/ChatMatrix.vue@@ -1,4 +1,6 @@<script setup lang="ts">+import { useIntervalFn } from '@vueuse/core'+interface Props {size?: numberdotSize?: number@@ -14,34 +16,19 @@ const props = withDefaults(defineProps<Props>(), {const totalDots = computed(() => props.size * props.size)const activeDots = ref<Set<number>>(new Set())-// Patterns for 4x4 grid (indices 0-15)-// Grid layout:-// 0 1 2 3-// 4 5 6 7-// 8 9 10 11-// 12 13 14 15const patterns = [- // Spiral inward[[0], [1], [2], [3], [7], [11], [15], [14], [13], [12], [8], [4], [5], [6], [10], [9]],- // Wave horizontal[[0, 4, 8, 12], [1, 5, 9, 13], [2, 6, 10, 14], [3, 7, 11, 15]],- // Diamond pulse[[5, 6, 9, 10], [1, 4, 7, 8, 11, 14], [0, 3, 12, 15], [1, 4, 7, 8, 11, 14], [5, 6, 9, 10]],- // Loading bar[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15]],- // Corners rotate[[0], [3], [15], [12]],- // Cross pulse[[5, 6, 9, 10], [1, 2, 4, 7, 8, 11, 13, 14], [0, 3, 12, 15]],- // Snake[[0], [1], [2], [3], [7], [6], [5], [4], [8], [9], [10], [11], [15], [14], [13], [12]],- // Diagonal wave[[0], [1, 4], [2, 5, 8], [3, 6, 9, 12], [7, 10, 13], [11, 14], [15]],]let patternIndex = 0let stepIndex = 0-let interval: ReturnType<typeof setInterval> | null = nullfunction nextStep() {const pattern = patterns[patternIndex]@@ -69,14 +56,7 @@ const dotStyle = computed(() => ({height: `${props.dotSize}px`,}))-onMounted(() => {- interval = setInterval(nextStep, 120)- nextStep()-})--onUnmounted(() => {- if (interval) clearInterval(interval)-})+useIntervalFn(nextStep, 120, { immediateCallback: true })</script><template>diff --git a/apps/chat/app/components/ModelSelect.vue b/apps/chat/app/components/ModelSelect.vueindex 14fcee2..2d81acc 100644--- a/apps/chat/app/components/ModelSelect.vue+++ b/apps/chat/app/components/ModelSelect.vue@@ -1,7 +1,6 @@<script setup lang="ts">const { model, models, formatModelName } = useModels()-// Map providers to icons (some don't have simple-icons)const providerIcons: Record<string, string> = {openai: 'i-simple-icons-openai',anthropic: 'i-simple-icons-anthropic',diff --git a/apps/chat/app/components/SourceCard.vue b/apps/chat/app/components/SourceCard.vueindex 25f8d45..7b8057c 100644--- a/apps/chat/app/components/SourceCard.vue+++ b/apps/chat/app/components/SourceCard.vue@@ -1,4 +1,6 @@<script setup lang="ts">+import { useTimeoutFn } from '@vueuse/core'+interface SourceData {id: stringtype: 'github' | 'youtube'@@ -25,12 +27,14 @@ const emit = defineEmits<{const isSyncing = ref(false)+const { start: resetSyncState } = useTimeoutFn(() => {+ isSyncing.value = false+}, 2000, { immediate: false })+function handleSync() {isSyncing.value = trueemit('sync')- setTimeout(() => {- isSyncing.value = false- }, 2000)+ resetSyncState()}const sourceUrl = computed(() => {diff --git a/apps/chat/app/components/ToolCallItem.vue b/apps/chat/app/components/ToolCallItem.vuenew file mode 100644index 0000000..7937e0b--- /dev/null+++ b/apps/chat/app/components/ToolCallItem.vue@@ -0,0 +1,252 @@+<script setup lang="ts">+import { motion, AnimatePresence } from 'motion-v'+import { useTimeoutFn } from '@vueuse/core'++interface CommandResult {+ command: string+ stdout: string+ stderr: string+ exitCode: number+ success: boolean+}++interface ToolExecutionResult {+ commands: CommandResult[]+ success: boolean+ durationMs: number+ error?: string+}++interface ToolCall {+ toolCallId: string+ toolName: string+ args: Record<string, unknown>+ state: 'loading' | 'done' | 'error'+ result?: ToolExecutionResult+}++const props = defineProps<{+ tool: ToolCall+}>()++const isExpanded = ref(false)+const displayedLabel = ref('')++const fullLabel = computed(() => {+ const { toolName, args, state } = props.tool+ const commands = args?.commands as string[] || (args?.command ? [args.command as string] : [])++ if (toolName === 'search_and_read') {+ const query = args?.query || '...'+ return state === 'loading' ? `Searching "${query}"` : `Searched "${query}"`+ }++ if (toolName === 'read') {+ const paths = args?.paths as string[] | undefined+ const path = paths?.[0]?.split('/').pop() || '...'+ return state === 'loading' ? `Reading ${path}` : `Read ${path}`+ }++ if (!commands.length) {+ return state === 'loading' ? 'Running...' : 'Done'+ }++ const summaries = commands.slice(0, 2).map(summarizeCommand)+ const suffix = commands.length > 2 ? ` +${commands.length - 2}` : ''+ return `${summaries.join(', ')}${suffix}`+})++function summarizeCommand(cmd: string): string {+ const trimmed = cmd.trim()+ const firstPart = (trimmed.split('|')[0] ?? '').trim()++ if (firstPart.startsWith('grep ') || firstPart.startsWith('rg ')) {+ const match = trimmed.match(/(?:grep|rg)\s+(?:-[a-zA-Z]+\s+)*["']?([^"'\s|]+)["']?/)+ return match ? `grep "${match[1]}"` : 'grep'+ }+ if (firstPart.startsWith('find ')) {+ const nameMatch = trimmed.match(/-name\s+["']?([^"'\s]+)["']?/)+ return nameMatch ? `find ${nameMatch[1]}` : 'find'+ }+ if (firstPart.startsWith('cat ') || firstPart.startsWith('head ') || firstPart.startsWith('tail ')) {+ const pathMatch = firstPart.match(/(?:cat|head|tail)\s+(?:-[a-zA-Z0-9]+\s+)*(.+)/)+ const fileName = pathMatch?.[1]?.trim().split('/').pop() || 'file'+ return `read ${fileName}`+ }+ if (firstPart.startsWith('ls ')) {+ const pathMatch = firstPart.match(/ls\s+(?:-[a-zA-Z]+\s+)*(.*)/)+ const path = pathMatch?.[1]?.trim() || '.'+ const dirName = path === '.' ? 'docs' : path.split('/').pop()+ return `ls ${dirName}`+ }++ const [cmdName] = firstPart.split(/\s+/)+ return cmdName || 'command'+}++let typewriterTimeout: ReturnType<typeof setTimeout> | null = null++function typewrite(text: string, index = 0) {+ if (index <= text.length) {+ displayedLabel.value = text.slice(0, index)+ typewriterTimeout = setTimeout(() => typewrite(text, index + 1), 20)+ }+}++watch(fullLabel, (newLabel) => {+ if (typewriterTimeout) clearTimeout(typewriterTimeout)+ if (newLabel.length > displayedLabel.value.length) {+ typewrite(newLabel, displayedLabel.value.length)+ } else {+ displayedLabel.value = newLabel+ }+}, { immediate: true })++const canExpand = computed(() => {+ const { args, result } = props.tool+ return args?.commands || args?.command || result?.commands?.length+})++const { start: startExpand, stop: stopExpand } = useTimeoutFn(() => {+ isExpanded.value = true+}, 150, { immediate: false })++const { start: startCollapse, stop: stopCollapse } = useTimeoutFn(() => {+ isExpanded.value = false+}, 2000, { immediate: false })++watch(() => props.tool.state, (newState, oldState) => {+ if (!canExpand.value) return++ stopExpand()+ stopCollapse()++ if (newState === 'loading' && oldState !== 'loading') {+ startExpand()+ } else if (newState === 'done' && oldState === 'loading') {+ startCollapse()+ }+}, { immediate: true })++onUnmounted(() => {+ if (typewriterTimeout) clearTimeout(typewriterTimeout)+ stopExpand()+ stopCollapse()+})++const commandsToShow = computed(() => {+ const { args, result } = props.tool++ if (result?.commands?.length) {+ return result.commands+ }++ const cmds = args?.commands as string[] || (args?.command ? [args.command as string] : [])+ return cmds.map(cmd => ({+ command: cmd,+ stdout: '',+ stderr: '',+ exitCode: 0,+ success: true,+ }))+})++function truncateOutput(text: string, maxLines = 8): string {+ const lines = text.split('\n').filter(l => l.trim())+ if (lines.length <= maxLines) return lines.join('\n')+ return `${lines.slice(0, maxLines).join('\n')}\n··· ${lines.length - maxLines} more lines`+}+</script>++<template>+ <motion.div+ :initial="{ opacity: 0, x: -6 }"+ :animate="{ opacity: 1, x: 0 }"+ :transition="{ duration: 0.35, ease: [0.25, 0.1, 0.25, 1] }"+ >+ <button+ class="flex items-center gap-2 w-full text-left py-0.5 group rounded -mx-1 px-1"+ :class="[+ canExpand ? 'cursor-pointer hover:bg-muted/5' : 'cursor-default',+ isExpanded && 'bg-muted/5',+ ]"+ @click="canExpand && (isExpanded = !isExpanded)"+ >+ <span class="relative flex size-1.5 shrink-0">+ <span+ v-if="tool.state === 'loading'"+ class="animate-ping absolute inline-flex h-full w-full rounded-full bg-current opacity-75"+ />+ <span+ class="relative inline-flex rounded-full size-1.5 bg-current"+ :class="tool.state === 'loading' ? 'opacity-100' : 'opacity-40'"+ />+ </span>++ <span class="text-xs text-dimmed font-mono flex-1 truncate">+ {{ displayedLabel }}<span v-if="tool.state === 'loading'" class="animate-pulse">_</span>+ </span>++ <span+ v-if="tool.result?.durationMs && tool.state === 'done'"+ class="text-[10px] text-muted/40 tabular-nums"+ >+ {{ tool.result.durationMs }}ms+ </span>++ <motion.span+ v-if="canExpand"+ :animate="{ rotate: isExpanded ? 180 : 0 }"+ :transition="{ duration: 0.35, ease: [0.25, 0.1, 0.25, 1] }"+ >+ <UIcon name="i-lucide-chevron-down" class="size-3 text-muted/40" />+ </motion.span>+ </button>++ <AnimatePresence>+ <motion.div+ v-if="isExpanded && commandsToShow.length"+ :initial="{ height: 0, opacity: 0 }"+ :animate="{ height: 'auto', opacity: 1 }"+ :exit="{ height: 0, opacity: 0 }"+ :transition="{ duration: 0.4, ease: [0.25, 0.1, 0.25, 1] }"+ class="overflow-hidden"+ >+ <div class="ml-3.5 pl-2 border-l border-muted/20 mt-1 mb-2 space-y-2">+ <div+ v-for="(cmd, idx) in commandsToShow"+ :key="idx"+ class="space-y-1"+ >+ <div class="flex items-start gap-1.5 text-[11px]">+ <span class="text-muted/60 font-mono shrink-0">$</span>+ <code class="text-muted/80 font-mono break-all">{{ cmd.command }}</code>+ </div>++ <div+ v-if="cmd.stdout || cmd.stderr"+ class="bg-muted/5 rounded px-2 py-1.5 font-mono text-[10px] text-dimmed/70 overflow-x-auto max-h-40 overflow-y-auto"+ >+ <pre v-if="cmd.stdout" class="whitespace-pre-wrap">{{ truncateOutput(cmd.stdout) }}</pre>+ <pre v-if="cmd.stderr" class="whitespace-pre-wrap text-muted/60 mt-1">{{ truncateOutput(cmd.stderr, 3) }}</pre>+ </div>++ <div+ v-else-if="tool.state === 'loading'"+ class="text-[10px] text-muted/40 italic"+ >+ running...+ </div>++ <div+ v-else-if="tool.state === 'done' && tool.result"+ class="text-[10px] text-muted/30 italic"+ >+ (no output)+ </div>+ </div>+ </div>+ </motion.div>+ </AnimatePresence>+ </motion.div>+</template>diff --git a/apps/chat/app/pages/chat/[id].vue b/apps/chat/app/pages/chat/[id].vueindex f3d124f..9e03730 100644--- a/apps/chat/app/pages/chat/[id].vue+++ b/apps/chat/app/pages/chat/[id].vue@@ -96,20 +96,27 @@ function copy(e: MouseEvent, message: UIMessage) {}, 2000)}-const lastMessage = computed(() => chat.messages.at(-1))-const showThinking = computed(() => {- if (chat.status !== 'streaming') return false- const msg = lastMessage.value- if (msg?.role !== 'assistant') return false- const parts = msg.parts as Array<{ type: string }> | undefined- return !parts?.some(p => p.type === 'text')-})+interface CommandResult {+ command: string+ stdout: string+ stderr: string+ exitCode: number+ success: boolean+}++interface ToolExecutionResult {+ commands: CommandResult[]+ success: boolean+ durationMs: number+ error?: string+}interface ToolCallInfo {toolCallId: stringtoolName: stringargs: Record<string, unknown>- state: 'loading' | 'done'+ state: 'loading' | 'done' | 'error'+ result?: ToolExecutionResult}function getMessageToolCalls(message: UIMessage): ToolCallInfo[] {diff --git a/apps/chat/server/api/chats/[id].post.ts b/apps/chat/server/api/chats/[id].post.tsindex 132ed2b..8ada4ed 100644--- a/apps/chat/server/api/chats/[id].post.ts+++ b/apps/chat/server/api/chats/[id].post.ts@@ -5,49 +5,52 @@ import { and, eq } from 'drizzle-orm'import { createSavoir } from '@savoir/sdk'import { log, useLogger } from 'evlog'import { generateTitle } from '../../utils/chat/generate-title'+import { routeQuestion, buildSystemPromptWithComplexity } from '../../utils/router/route-question'-const SYSTEM_PROMPT = `You are an AI assistant specialized in the Nuxt/Nitro ecosystem.+const SYSTEM_PROMPT = `You are an AI assistant that answers questions based on the available sources.-## CRITICAL: Documentation First+## CRITICAL: Sources First-Your knowledge may be outdated. ONLY answer based on what you find in the documentation.-- If you can't find information in the docs, say "I couldn't find this in the documentation"-- NEVER make up information or guess - only state what you found in docs-- NEVER mention package versions unless you verified them in the docs+Your knowledge may be outdated. ONLY answer based on what you find in the sources.+- If you can't find information, say "I couldn't find this in the available sources"+- NEVER make up information or guess - only state what you found+- Always cite the source file when quoting content-## Tools+## Search Strategy (IMPORTANT for speed)-You have access to documentation search tools:+1. **Explore first** - Start by discovering the available sources:+ \`\`\`bash+ ls /docs # See what sources are available+ \`\`\`-- **search_and_read**: Search AND read files in one step (PREFERRED - faster and more efficient)-- **read**: Read specific files by path (use when you already know the exact file path)+2. **Target specific directories** - Don't search everything at once:+ \`\`\`bash+ ls /docs/[source]/ # Explore the structure+ grep -r "keyword" /docs/[source] --include="*.md" -l | head -10+ \`\`\`-## How to Search Effectively--1. **Use ONE simple keyword**: "composables" ✓, "vue composables usage" ✗-2. **Always prefer search_and_read** - it's faster and returns content directly-3. **Search by concept**: "middleware", "plugins", "routing", "state", "data-fetching"-4. **Search by package/module name**: "unstorage", "ofetch", "nitro", "h3", "unhead"+3. **Use simple keywords**: one word is better than a phrase## Workflow1. User asks a question-2. Call **search_and_read** with a relevant keyword-3. Read the results and find the answer-4. If not found, try different keywords or broader terms-5. Answer based ONLY on what you found in the documentation+2. Explore available sources with \`ls\`+3. Identify which source(s) are relevant+4. Search within those specific directories+5. Read the relevant files+6. Answer based ONLY on what you found## Response Style- Be concise and helpful-- Include relevant code examples from the docs+- Include relevant code examples when available- Use markdown formatting-- Cite the source file when quoting documentation+- Cite the source file path`defineRouteMeta({openAPI: {- description: 'Chat with AI about Nuxt documentation.',+ description: 'Chat with AI about the available sources.',tags: ['ai'],},})@@ -118,7 +121,10 @@ export default defineEventHandler(async (event) => {apiUrl: getRequestURL(event).origin,apiKey: savoirConfig.apiKey || undefined,onToolCall: (info) => {- log.info('chat', `[${requestId}] Tool call [${info.state}]: ${info.toolName} ${JSON.stringify(info.args)}`)+ const resultSummary = info.result+ ? ` [${info.result.success ? 'OK' : 'FAIL'}] (${info.result.durationMs}ms)`+ : ''+ log.info('chat', `[${requestId}] Tool call [${info.state}]: ${info.toolName}${resultSummary}`)if (streamWriter) {streamWriter.write({@@ -129,13 +135,16 @@ export default defineEventHandler(async (event) => {toolName: info.toolName,args: info.args,state: info.state,+ result: info.result,},})}},})- log.info('chat', `[${requestId}] Starting agent with ${model}`)+ const agentConfig = await routeQuestion(messages, requestId)++ log.info('chat', `[${requestId}] Starting agent with ${model} (routed: ${agentConfig.complexity}, ${agentConfig.maxSteps} steps)`)const stream = createUIMessageStream({execute: async ({ writer }) => {@@ -143,8 +152,9 @@ export default defineEventHandler(async (event) => {const agent = new ToolLoopAgent({model,- instructions: SYSTEM_PROMPT,+ instructions: buildSystemPromptWithComplexity(SYSTEM_PROMPT, agentConfig),tools: savoir.tools,+ stopWhen: stepCountIs(agentConfig.maxSteps),onStepFinish: (stepResult) => {const stepDurationMs = Date.now() - stepStartTimestepDurations.push(stepDurationMs)@@ -177,6 +187,9 @@ export default defineEventHandler(async (event) => {toolCallCount,stepDurations,totalAgentMs: totalDurationMs,+ routerComplexity: agentConfig.complexity,+ routerMaxSteps: agentConfig.maxSteps,+ routerReasoning: agentConfig.reasoning,})log.info('chat', `[${requestId}] Finished: ${result.finishReason} (total: ${totalDurationMs}ms)`)},diff --git a/apps/chat/server/utils/router/route-question.ts b/apps/chat/server/utils/router/route-question.tsnew file mode 100644index 0000000..abb8543--- /dev/null+++ b/apps/chat/server/utils/router/route-question.ts@@ -0,0 +1,96 @@+import { createGateway } from '@ai-sdk/gateway'+import { generateText, Output } from 'ai'+import { log } from 'evlog'+import type { UIMessage } from 'ai'+import { type AgentConfig, agentConfigSchema, getDefaultConfig } from './schema'++const ROUTER_MODEL = 'google/gemini-2.5-flash-lite'++const ROUTER_SYSTEM_PROMPT = `You are a question classifier for an AI assistant.+Analyze the user's question and determine the appropriate configuration for the agent.++## Classification Guidelines++**trivial** (maxSteps: 3, model: gemini-2.5-flash-lite)+- Simple greetings: "Hello", "Thanks", "Hi there"+- Acknowledgments without questions++**simple** (maxSteps: 6, model: gemini-2.5-flash-lite)+- Single concept lookups: "What is X?", "How to use Y?"+- Direct questions with likely clear answers+- Questions about a single topic++**moderate** (maxSteps: 12, model: gemini-3-flash)+- Comparisons: "Difference between X and Y?"+- Integration questions requiring exploration+- Questions requiring multiple file searches++**complex** (maxSteps: 20, model: gemini-3-flash or claude-opus-4.5)+- Debugging scenarios+- Architecture questions+- Deep analysis requiring extensive research+- Cross-topic questions++Use claude-opus-4.5 only for the most complex cases requiring deep reasoning.`++function extractQuestionFromMessages(messages: UIMessage[]): string {+ // Get the last user message+ const lastUserMessage = [...messages].reverse().find(m => m.role === 'user')+ if (!lastUserMessage) return ''++ // Extract text from message parts+ const textParts = lastUserMessage.parts+ ?.filter((p): p is { type: 'text', text: string } => p.type === 'text')+ .map(p => p.text)+ .join('\n')++ return textParts || ''+}++export async function routeQuestion(+ messages: UIMessage[],+ requestId: string,+): Promise<AgentConfig> {+ const config = useRuntimeConfig()+ const gateway = createGateway({ apiKey: config.savoir?.apiKey })++ const question = extractQuestionFromMessages(messages)+ if (!question) {+ log.info('chat', `[${requestId}] Router: no question found, using default config`)+ return getDefaultConfig()+ }++ try {+ const { output } = await generateText({+ model: gateway(ROUTER_MODEL),+ output: Output.object({ schema: agentConfigSchema }),+ messages: [+ { role: 'system', content: ROUTER_SYSTEM_PROMPT },+ { role: 'user', content: `Question: ${question}` },+ ],+ })++ if (!output) {+ log.warn('chat', `[${requestId}] Router returned no output, using default config`)+ return getDefaultConfig()+ }++ log.info('chat', `[${requestId}] Router decision: ${output.complexity} (${output.model}, ${output.maxSteps} steps) - ${output.reasoning}`)+ return output+ } catch (error) {+ const errorMessage = error instanceof Error ? error.message : 'Unknown error'+ log.error('chat', `[${requestId}] Router failed: ${errorMessage}, using default config`)+ return getDefaultConfig()+ }+}++export function buildSystemPromptWithComplexity(basePrompt: string, agentConfig: AgentConfig): string {+ const complexityHints: Record<AgentConfig['complexity'], string> = {+ trivial: 'This is a simple greeting or acknowledgment. Respond briefly without searching.',+ simple: 'This is a straightforward question. Explore with `ls`, then do a targeted search.',+ moderate: 'This requires some research. Explore the sources structure first, then search within relevant directories.',+ complex: 'This is a complex question. Map out the available sources with `ls`, then systematically search relevant areas.',+ }++ return `${basePrompt}\n\n## Task Complexity: ${agentConfig.complexity}\n${complexityHints[agentConfig.complexity]}`+}diff --git a/apps/chat/server/utils/router/schema.ts b/apps/chat/server/utils/router/schema.tsnew file mode 100644index 0000000..185f898--- /dev/null+++ b/apps/chat/server/utils/router/schema.ts@@ -0,0 +1,29 @@+import { z } from 'zod'++export const agentConfigSchema = z.object({+ complexity: z.enum(['trivial', 'simple', 'moderate', 'complex'])+ .describe('trivial=greeting, simple=single lookup, moderate=multi-search, complex=deep analysis'),++ maxSteps: z.number().min(1).max(30)+ .describe('Agent iterations: 3 trivial, 6 simple, 12 moderate, 20 complex'),++ model: z.enum([+ 'google/gemini-3-flash',+ 'google/gemini-2.5-flash-lite',+ 'anthropic/claude-opus-4.5',+ ]).describe('flash-lite for trivial/simple, flash for moderate, opus for complex'),++ reasoning: z.string().max(200)+ .describe('Brief explanation of the classification'),+})++export type AgentConfig = z.infer<typeof agentConfigSchema>++export function getDefaultConfig(): AgentConfig {+ return {+ complexity: 'moderate',+ maxSteps: 12,+ model: 'google/gemini-3-flash',+ reasoning: 'Default fallback configuration',+ }+}diff --git a/apps/github-bot/server/utils/ai.ts b/apps/github-bot/server/utils/ai.tsindex f02c31f..6b8226c 100644--- a/apps/github-bot/server/utils/ai.ts+++ b/apps/github-bot/server/utils/ai.ts@@ -1,7 +1,83 @@-import { ToolLoopAgent } from 'ai'+import { createGateway } from '@ai-sdk/gateway'+import { generateText, Output, stepCountIs, ToolLoopAgent } from 'ai'import { log } from 'evlog'import { createSavoir } from '@savoir/sdk'import type { IssueContext } from './types'+import { type AgentConfig, agentConfigSchema, getDefaultConfig } from './router-schema'++const ROUTER_MODEL = 'google/gemini-2.5-flash-lite'++const ROUTER_SYSTEM_PROMPT = `You are a question classifier for a documentation assistant bot.+Analyze the user's question and determine the appropriate configuration for the agent.++## Classification Guidelines++**trivial** (maxSteps: 3, model: gemini-2.5-flash-lite)+- Simple greetings: "Hello", "Thanks", "Hi there"+- Acknowledgments without questions++**simple** (maxSteps: 6, model: gemini-2.5-flash-lite)+- Single concept lookups: "What is useAsyncData?", "How to use useFetch?"+- Direct API questions with clear answers++**moderate** (maxSteps: 12, model: gemini-3-flash)+- Comparisons: "Difference between useFetch and useAsyncData?"+- Integration questions: "How to use Nuxt with TypeScript?"+- Questions requiring multiple file searches++**complex** (maxSteps: 20, model: gemini-3-flash or claude-opus-4.5)+- Debugging scenarios: "My middleware isn't working with auth"+- Architecture questions: "How to structure a multi-package monorepo?"+- Deep analysis requiring extensive documentation review++Use claude-opus-4.5 only for the most complex cases requiring deep reasoning.`++function buildRouterInput(question: string, context?: IssueContext): string {+ const parts: string[] = []++ if (context) {+ parts.push(`Repository: ${context.owner}/${context.repo}`)+ parts.push(`Issue #${context.number}: ${context.title}`)+ if (context.body) {+ parts.push(`Issue body: ${context.body.slice(0, 500)}`)+ }+ if (context.labels.length) {+ parts.push(`Labels: ${context.labels.join(', ')}`)+ }+ }++ parts.push(`Question: ${question}`)++ return parts.join('\n')+}++async function routeQuestion(question: string, context?: IssueContext): Promise<AgentConfig> {+ const config = useRuntimeConfig()+ const gateway = createGateway({ apiKey: config.savoir.apiKey })++ try {+ const { output } = await generateText({+ model: gateway(ROUTER_MODEL),+ output: Output.object({ schema: agentConfigSchema }),+ messages: [+ { role: 'system', content: ROUTER_SYSTEM_PROMPT },+ { role: 'user', content: buildRouterInput(question, context) },+ ],+ })++ if (!output) {+ log.warn('github-bot', 'Router returned no output, using default config')+ return getDefaultConfig()+ }++ log.info('github-bot', `Router decision: ${output.complexity} (${output.model}, ${output.maxSteps} steps) - ${output.reasoning}`)+ return output+ } catch (error) {+ const errorMessage = error instanceof Error ? error.message : 'Unknown error'+ log.error('github-bot', `Router failed: ${errorMessage}, using default config`)+ return getDefaultConfig()+ }+}export async function generateAIResponse(question: string,@@ -26,7 +102,9 @@ Once configured, I'll be able to search the documentation and help answer your qlet toolCallCount = 0try {- log.info('github-bot', `Starting agent for issue #${context?.number || 'unknown'}`)+ const agentConfig = await routeQuestion(question, context)++ log.info('github-bot', `Starting agent for issue #${context?.number || 'unknown'} with ${agentConfig.complexity} complexity`)const savoir = createSavoir({apiUrl: config.savoir.apiUrl,@@ -39,9 +117,10 @@ Once configured, I'll be able to search the documentation and help answer your q})const agent = new ToolLoopAgent({- model: 'google/gemini-3-flash',- instructions: buildSystemPrompt(context),+ model: agentConfig.model,+ instructions: buildSystemPrompt(context, agentConfig),tools: savoir.tools,+ stopWhen: stepCountIs(agentConfig.maxSteps),onStepFinish: ({ usage, toolCalls }) => {stepCount++totalInputTokens += usage.inputTokens || 0@@ -91,7 +170,7 @@ Please try again later or open a discussion if this persists.`}}-function buildSystemPrompt(context?: IssueContext): string {+function buildSystemPrompt(context?: IssueContext, agentConfig?: AgentConfig): string {const basePrompt = `You are a documentation assistant with bash access to a sandbox containing docs (markdown, JSON, YAML).## Speed is Important@@ -121,13 +200,23 @@ ls docs/ && grep -rl "keyword" docs/ --include="*.md" | head -10- Include code examples- Use markdown`- if (context) {- return `${basePrompt}+ let prompt = basePrompt-Issue #${context.number}: "${context.title}" in ${context.owner}/${context.repo}`+ if (agentConfig) {+ const complexityHints: Record<AgentConfig['complexity'], string> = {+ trivial: 'This is a simple greeting or acknowledgment. Respond briefly without searching.',+ simple: 'This is a straightforward question. One or two searches should suffice.',+ moderate: 'This requires some research. Take time to find relevant documentation.',+ complex: 'This is a complex question. Thoroughly search the documentation and provide a detailed answer.',+ }+ prompt += `\n\n## Task Complexity: ${agentConfig.complexity}\n${complexityHints[agentConfig.complexity]}`+ }++ if (context) {+ prompt += `\n\nIssue #${context.number}: "${context.title}" in ${context.owner}/${context.repo}`}- return basePrompt+ return prompt}function buildUserMessage(question: string, context?: IssueContext): string {diff --git a/apps/github-bot/server/utils/router-schema.ts b/apps/github-bot/server/utils/router-schema.tsnew file mode 100644index 0000000..185f898--- /dev/null+++ b/apps/github-bot/server/utils/router-schema.ts@@ -0,0 +1,29 @@+import { z } from 'zod'++export const agentConfigSchema = z.object({+ complexity: z.enum(['trivial', 'simple', 'moderate', 'complex'])+ .describe('trivial=greeting, simple=single lookup, moderate=multi-search, complex=deep analysis'),++ maxSteps: z.number().min(1).max(30)+ .describe('Agent iterations: 3 trivial, 6 simple, 12 moderate, 20 complex'),++ model: z.enum([+ 'google/gemini-3-flash',+ 'google/gemini-2.5-flash-lite',+ 'anthropic/claude-opus-4.5',+ ]).describe('flash-lite for trivial/simple, flash for moderate, opus for complex'),++ reasoning: z.string().max(200)+ .describe('Brief explanation of the classification'),+})++export type AgentConfig = z.infer<typeof agentConfigSchema>++export function getDefaultConfig(): AgentConfig {+ return {+ complexity: 'moderate',+ maxSteps: 12,+ model: 'google/gemini-3-flash',+ reasoning: 'Default fallback configuration',+ }+}diff --git a/packages/sdk/src/index.ts b/packages/sdk/src/index.tsindex 67b5c57..6e6c909 100644--- a/packages/sdk/src/index.ts+++ b/packages/sdk/src/index.ts@@ -2,7 +2,7 @@ import { SavoirClient } from './client'import { createBashBatchTool, createBashTool } from './tools'import type { SavoirConfig } from './types'-export type { SavoirConfig, ShellResponse, ShellBatchResponse, ShellCommandResult, SyncOptions, SyncResponse, SnapshotResponse, GitHubSource, YouTubeSource, SourcesResponse, SyncSourceResponse, ToolCallInfo, ToolCallCallback, ToolCallState } from './types'+export type { SavoirConfig, ShellResponse, ShellBatchResponse, ShellCommandResult, SyncOptions, SyncResponse, SnapshotResponse, GitHubSource, YouTubeSource, SourcesResponse, SyncSourceResponse, ToolCallInfo, ToolCallCallback, ToolCallState, ToolExecutionResult, CommandResult } from './types'export { SavoirError, NetworkError } from './errors'export { SavoirClient } from './client'diff --git a/packages/sdk/src/tools/shell.ts b/packages/sdk/src/tools/shell.tsindex de7ab05..1671c1f 100644--- a/packages/sdk/src/tools/shell.ts+++ b/packages/sdk/src/tools/shell.ts@@ -1,7 +1,7 @@import { tool } from 'ai'import { z } from 'zod'import type { SavoirClient } from '../client'-import type { ToolCallCallback } from '../types'+import type { ToolCallCallback, ToolExecutionResult, CommandResult } from '../types'/*** Create the bash tool for AI SDK - runs a single command in the sandbox@@ -22,19 +22,57 @@ Use standard Unix commands to explore and read files.`,})},execute: async ({ command }, { toolCallId }) => {- const result = await client.bash(command)+ const startTime = Date.now()- onToolCall?.({- toolCallId,- toolName: 'bash',- args: { command },- state: 'done',- })+ try {+ const apiResult = await client.bash(command)+ const durationMs = Date.now() - startTime++ const commandResult: CommandResult = {+ command,+ stdout: apiResult.stdout,+ stderr: apiResult.stderr,+ exitCode: apiResult.exitCode,+ success: apiResult.exitCode === 0,+ }++ const executionResult: ToolExecutionResult = {+ commands: [commandResult],+ success: commandResult.success,+ durationMs,+ }++ onToolCall?.({+ toolCallId,+ toolName: 'bash',+ args: { command },+ state: 'done',+ result: executionResult,+ })++ return {+ stdout: apiResult.stdout,+ stderr: apiResult.stderr,+ exitCode: apiResult.exitCode,+ }+ } catch (error) {+ const durationMs = Date.now() - startTime+ const errorMessage = error instanceof Error ? error.message : 'Unknown error'++ onToolCall?.({+ toolCallId,+ toolName: 'bash',+ args: { command },+ state: 'error',+ result: {+ commands: [{ command, stdout: '', stderr: errorMessage, exitCode: 1, success: false }],+ success: false,+ durationMs,+ error: errorMessage,+ },+ })- return {- stdout: result.stdout,- stderr: result.stderr,- exitCode: result.exitCode,+ throw error}},})@@ -61,22 +99,60 @@ Maximum 10 commands per batch.`,})},execute: async ({ commands }, { toolCallId }) => {- const result = await client.bashBatch(commands)+ const startTime = Date.now()- onToolCall?.({- toolCallId,- toolName: 'bash_batch',- args: { commands },- state: 'done',- })+ try {+ const apiResult = await client.bashBatch(commands)+ const durationMs = Date.now() - startTime- return {- results: result.results.map(r => ({+ const commandResults: CommandResult[] = apiResult.results.map(r => ({command: r.command,stdout: r.stdout,stderr: r.stderr,exitCode: r.exitCode,- })),+ success: r.exitCode === 0,+ }))++ const executionResult: ToolExecutionResult = {+ commands: commandResults,+ success: commandResults.every(r => r.success),+ durationMs,+ }++ onToolCall?.({+ toolCallId,+ toolName: 'bash_batch',+ args: { commands },+ state: 'done',+ result: executionResult,+ })++ return {+ results: apiResult.results.map(r => ({+ command: r.command,+ stdout: r.stdout,+ stderr: r.stderr,+ exitCode: r.exitCode,+ })),+ }+ } catch (error) {+ const durationMs = Date.now() - startTime+ const errorMessage = error instanceof Error ? error.message : 'Unknown error'++ onToolCall?.({+ toolCallId,+ toolName: 'bash_batch',+ args: { commands },+ state: 'error',+ result: {+ commands: commands.map(cmd => ({ command: cmd, stdout: '', stderr: '', exitCode: 1, success: false })),+ success: false,+ durationMs,+ error: errorMessage,+ },+ })++ throw error}},})diff --git a/packages/sdk/src/types.ts b/packages/sdk/src/types.tsindex 74e78ca..ae60d22 100644--- a/packages/sdk/src/types.ts+++ b/packages/sdk/src/types.ts@@ -29,7 +29,32 @@ export interface SavoirConfig {/*** Tool call state*/-export type ToolCallState = 'loading' | 'done'+export type ToolCallState = 'loading' | 'done' | 'error'++/**+ * Result of a single command execution+ */+export interface CommandResult {+ command: string+ stdout: string+ stderr: string+ exitCode: number+ success: boolean+}++/**+ * Tool execution result (available when state is 'done' or 'error')+ */+export interface ToolExecutionResult {+ /** Results for each command (single item for bash, multiple for bash_batch) */+ commands: CommandResult[]+ /** Whether all commands succeeded */+ success: boolean+ /** Execution duration in milliseconds */+ durationMs: number+ /** Error message if state is 'error' */+ error?: string+}/*** Tool call information passed to onToolCall callback@@ -39,6 +64,8 @@ export interface ToolCallInfo {toolName: stringargs: Record<string, unknown>state: ToolCallState+ /** Execution result (only available when state is 'done' or 'error') */+ result?: ToolExecutionResult}/**