@soederpop/luca 0.2.1 → 0.2.3

package/docs/CNAME

@@ -0,0 +1 @@
+luca-js.soederpop.com

package/docs/examples/assistant-hooks-reference.ts

@@ -0,0 +1,171 @@
+/**
+ * Example hooks.ts — showcases the assistant hook system.
+ *
+ * Every named export is a hook. Hooks are awaited via triggerHook(),
+ * so async work completes BEFORE the assistant proceeds.
+ *
+ * Available hooks:
+ *
+ *   Lifecycle:
+ *     created()              — after prompt/tools/hooks load (before start)
+ *     beforeStart()          — before conversation wiring, blocks start()
+ *     started()              — conversation is ready, wire up tools here
+ *     afterStart()           — everything is live, blocks start() until done
+ *     formatSystemPrompt(a, prompt) => string — rewrite the system prompt
+ *
+ *   Ask flow:
+ *     beforeInitialAsk(a, question, options) — first ask() only
+ *     beforeAsk(a, question, options) => string? — every ask(), return rewrites question
+ *     answered(a, result) — after response, before ask() returns
+ *
+ *   Tool execution:
+ *     beforeToolCall(a, ctx) — inspect/rewrite args, set ctx.skip to bypass
+ *     afterToolCall(a, ctx)  — inspect/rewrite result after execution
+ *
+ *   Forwarded from conversation (awaited before the bus event):
+ *     turnStart(a, info)  turnEnd(a, info)  chunk(a, delta)
+ *     preview(a, text)    response(a, text)  toolCall(a, name, args)
+ *     toolResult(a, name, result)  toolError(a, name, error)
+ *
+ * The first argument is always the assistant instance.
+ * `assistant` and `container` are also available as globals (injected by the VM).
+ */
+import type { Assistant, AGIContainer } from '@soederpop/luca/agi'
+
+declare global {
+  var assistant: Assistant
+  var container: AGIContainer
+}
+
+// ─── Lifecycle ──────────────────────────────────────────────────
+
+/**
+ * Runs before the conversation is created. Good for async setup
+ * that needs to finish before the assistant is usable.
+ */
+export async function beforeStart() {
+  console.log(`[hooks] preparing ${assistant.assistantName}...`)
+}
+
+/**
+ * Conversation is wired — register tools, plugins, extensions.
+ * This is the most common hook: equivalent to the old bus-based pattern.
+ */
+export function started() {
+  // Give the assistant shell tools
+  assistant.use(container.feature('codingTools'))
+
+  // Add write operations from fileTools
+  const fileTools = container.feature('fileTools')
+  assistant.use(fileTools.toTools({ only: ['editFile', 'writeFile'] }))
+  fileTools.setupToolsConsumer(assistant)
+}
+
+/**
+ * Everything is live. Good for loading state that depends on
+ * the conversation being fully initialized. Blocks start().
+ */
+export async function afterStart() {
+  // e.g. load a knowledge base into the system prompt
+  const fs = container.feature('fs')
+  const notesPath = assistant.paths.resolve('notes.md')
+
+  if (await fs.exists(notesPath)) {
+    const notes = await fs.readFile(notesPath)
+    assistant.addSystemPromptExtension('notes', `\n## Your Notes\n${notes}`)
+  }
+}
+
+/**
+ * Rewrite the system prompt before the conversation is created.
+ * Return the new prompt string.
+ */
+export async function formatSystemPrompt(_assistant: Assistant, prompt: string) {
+  const today = new Date().toLocaleDateString('en-US', {
+    weekday: 'long', year: 'numeric', month: 'long', day: 'numeric',
+  })
+  return `${prompt}\n\nToday is ${today}.`
+}
+
+// ─── Ask Flow ───────────────────────────────────────────────────
+
+/**
+ * Runs only on the very first ask(). Good for one-time greetings
+ * or loading context that depends on the first user message.
+ */
+export async function beforeInitialAsk(_assistant: Assistant, question: string) {
+  console.log(`[hooks] first message from user: "${question}"`)
+}
+
+/**
+ * Runs before every ask(). Return a string to rewrite the question.
+ * Useful for injecting context the model should see but the user
+ * shouldn't have to type every time.
+ */
+export async function beforeAsk(_assistant: Assistant, question: string) {
+  // Example: inject recent git context into every question
+  const proc = container.feature('proc')
+  const { stdout } = await proc.exec('git log --oneline -5 2>/dev/null || true')
+
+  if (stdout.trim()) {
+    return `${question}\n\n<context>\nRecent commits:\n${stdout.trim()}\n</context>`
+  }
+}
+
+/**
+ * Fires after the model responds, before ask() returns.
+ * Good for logging, analytics, or auto-saving.
+ */
+export async function answered(_assistant: Assistant, result: string) {
+  console.log(`[hooks] response length: ${result.length} chars`)
+}
+
+// ─── Tool Execution ─────────────────────────────────────────────
+
+/**
+ * Fires before every tool call. The ctx object is mutable:
+ *   ctx.name  — tool name
+ *   ctx.args  — arguments (rewrite to modify)
+ *   ctx.skip  — set true to bypass execution
+ *   ctx.result — set when skipping to provide a result
+ */
+export async function beforeToolCall(_assistant: Assistant, ctx: any) {
+  console.log(`[hooks] tool call: ${ctx.name}(${JSON.stringify(ctx.args)})`)
+
+  // Example: block dangerous commands
+  if (ctx.name === 'runCommand' && ctx.args.command?.includes('rm -rf')) {
+    ctx.skip = true
+    ctx.result = JSON.stringify({ error: 'Blocked: destructive command not allowed' })
+  }
+}
+
+/**
+ * Fires after every tool call. The ctx object contains:
+ *   ctx.name   — tool name
+ *   ctx.args   — original arguments
+ *   ctx.result — the result string (rewrite to modify what the model sees)
+ *   ctx.error  — error object if the tool threw
+ */
+export async function afterToolCall(_assistant: Assistant, ctx: any) {
+  if (ctx.error) {
+    console.error(`[hooks] tool ${ctx.name} failed:`, ctx.error.message)
+  }
+
+  // Example: truncate huge tool outputs so they don't blow the context
+  if (ctx.result && ctx.result.length > 10000) {
+    const truncated = ctx.result.slice(0, 10000)
+    ctx.result = `${truncated}\n\n... (truncated from ${ctx.result.length} chars)`
+  }
+}
+
+// ─── Forwarded Conversation Events ──────────────────────────────
+
+/**
+ * Fires at the start of each completion turn. turn > 1 means
+ * the model is continuing after tool calls.
+ */
+export function turnStart(_assistant: Assistant, info: { turn: number; isFollowUp: boolean }) {
+  if (info.isFollowUp) {
+    console.log(`[hooks] follow-up turn ${info.turn}`)
+  }
+}