obol-ai 0.3.46 → 0.3.48

package/CHANGELOG.md

@@ -1,3 +1,9 @@
+## 0.3.47
+- fix elapsed scope bug in background task completion
+- fix TTS summary to speak in first person
+- inject proactive follow-up messages into conversation history
+- fix OAuth token expiry in heartbeat scheduled events
+
 ## 0.3.44
 - exclude electron from npm package
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "obol-ai",
-  "version": "0.3.46",
+  "version": "0.3.48",
   "description": "Self-evolving AI assistant that learns, remembers, and acts on its own. Persistent vector memory, self-rewriting personality, proactive heartbeats.",
   "main": "src/index.js",
   "bin": {

package/src/claude/router.js

@@ -14,6 +14,18 @@ function buildRouterMessages(recentHistory, userMessage) {
   return [...trimmed, { role: 'user', content: userMessage }];
 }
 
+function extractRecentUserMessages(recentHistory, userMessage, count = 5) {
+  const userMsgs = recentHistory
+    .filter(m => m.role === 'user')
+    .map(m => typeof m.content === 'string'
+      ? m.content
+      : m.content.filter(b => b.type === 'text').map(b => b.text).join(''))
+    .filter(Boolean)
+    .slice(-count);
+  userMsgs.push(userMessage);
+  return userMsgs;
+}
+
 function tokenize(s) {
   return new Set(s.toLowerCase().split(/\W+/).filter(Boolean));
 }
@@ -38,9 +50,7 @@ async function routeMessage(client, memory, userMessage, { vlog, onRouteDecision
 2. What model complexity does it need?
 
 Reply with ONLY a JSON object:
-{"need_memory": true/false, "search_queries": ["query1", "query2"], "model": "sonnet|opus"}
-
-search_queries: 1-5 optimized search queries based on the full conversation context. Cover distinct topics, people, entities, time periods, or projects referenced. Single-topic messages need just one query. Use more queries when the message references multiple people, projects, or threads.
+{"need_memory": true/false, "model": "sonnet|opus"}
 
 Memory: casual messages (greetings, jokes, simple questions) → false. References to past, people, projects, preferences → true.
 
@@ -57,15 +67,11 @@ If recent context shows an ongoing task (sonnet/opus was just used, multi-step w
       if (jsonStr) decision = JSON.parse(jsonStr);
     } catch {}
 
-    const queries = Array.isArray(decision.search_queries) && decision.search_queries.length > 0
-      ? decision.search_queries.slice(0, 3)
-      : decision.search_query ? [decision.search_query] : [];
-
     if (decision.model !== 'sonnet' && decision.model !== 'opus') {
       decision.model = 'sonnet';
     }
 
-    vlog(`[router] model=${decision.model} memory=${decision.need_memory || false}${queries.length ? ` queries=${JSON.stringify(queries)}` : ''}`);
+    vlog(`[router] model=${decision.model} memory=${decision.need_memory || false}`);
 
     onRouteDecision?.({
       model: decision.model,
@@ -80,23 +86,19 @@ If recent context shows an ongoing task (sonnet/opus was just used, multi-step w
     if (decision.need_memory && memory) {
       const budget = decision.model === 'opus' ? 60 : 40;
       const poolPerQuery = decision.model === 'opus' ? 25 : 20;
-      const searchQueries = queries.length > 0 ? queries : [userMessage];
-
-      const recentMemories = await memory.byDate('7d', { limit: Math.ceil(budget / 3) });
+      const conversationQueries = extractRecentUserMessages(recentHistory, userMessage);
 
       const semanticResults = await Promise.all(
-        searchQueries.map(q => memory.search(q, { limit: poolPerQuery, threshold: 0.4 }))
+        conversationQueries.map(q => memory.search(q, { limit: poolPerQuery, threshold: 0.4 }))
       );
       const semanticMemories = semanticResults.flat();
 
       const seen = new Set();
       const combined = [];
-      for (const m of [...recentMemories, ...semanticMemories]) {
+      for (const m of semanticMemories) {
         if (!seen.has(m.id)) {
           seen.add(m.id);
-          const ageDays = m.created_at ? (Date.now() - new Date(m.created_at).getTime()) / 86400000 : 7;
-          const recencyBonus = Math.max(0, 1 - ageDays / 7) * 0.3;
-          m._score = (m.similarity || 0.5) * 0.5 + (m.importance || 0.5) * 0.2 + recencyBonus;
+          m._score = (m.similarity || 0.5) * 0.7 + (m.importance || 0.5) * 0.3;
           combined.push(m);
         }
       }
@@ -112,7 +114,7 @@ If recent context shows an ongoing task (sonnet/opus was just used, multi-step w
         if (!isDup) topFacts.push(m);
       }
 
-      vlog(`[memory] ${topFacts.length} facts (${recentMemories.length} recent, ${semanticMemories.length} semantic, budget=${budget})`);
+      vlog(`[memory] ${topFacts.length} facts from ${conversationQueries.length} conversation queries (${semanticMemories.length} candidates, budget=${budget})`);
       onRouteUpdate?.({ memoryCount: topFacts.length });
 
       memoryBlock = formatMemoryBlock(topFacts);

package/src/runtime/background.js

@@ -110,6 +110,8 @@ TASK: ${task}`;
       claude.clearHistory(`bg-${taskState.id}`);
       clearStatus();
 
+      const elapsed = Math.floor((Date.now() - taskState.startedAt) / 1000);
+
       if (!result?.trim()) {
         taskState.status = 'error';
         taskState.error = 'No result returned';
@@ -119,7 +121,6 @@ TASK: ${task}`;
       } else {
         taskState.status = 'done';
         taskState.result = result;
-        const elapsed = Math.floor((Date.now() - taskState.startedAt) / 1000);
         if (silent) {
           await sendLong(ctx, result);
         } else {

package/src/runtime/heartbeat.js

@@ -7,6 +7,7 @@ const { runAnalysis } = require('../analysis');
 const { runProactiveNews } = require('../news');
 const { createSelfMemory } = require('../memory/self');
 const { createAnthropicClient, ensureFreshToken } = require('../claude/client');
+const { markdownToTelegramHtml } = require('../telegram/utils');
 
 
 const ANALYSIS_HOURS = new Set([4, 7, 10, 13, 16, 19, 22]);
@@ -145,7 +146,6 @@ async function runNewsForUser(bot, config, userId) {
 
 const MAX_ATTEMPTS = 3;
 const STALE_MS = 2 * 60 * 60 * 1000;
-const EVENT_TIMEOUT_MS = 30_000;
 
 async function processEvent(bot, config, scheduler, event) {
   const tz = event.timezone || 'UTC';
@@ -320,40 +320,30 @@ async function runAgenticEvent(bot, config, event) {
   const timezone = event.timezone || getUserTimezone(config, event.user_id);
 
   const query = event.description || event.instructions;
-  const context = await buildProactiveContext(tenant, timezone, query).catch(() => '');
-
-  const systemParts = [];
-  if (tenant.personality?.soul) systemParts.push(tenant.personality.soul);
-  if (tenant.personality?.user) systemParts.push(`About this user:\n${tenant.personality.user}`);
-  if (context) systemParts.push(`Current context:\n${context}`);
-  systemParts.push(
-    'You are executing a scheduled task. Respond ONLY with a clean, natural-language message for the user. ' +
-    'Do NOT output JSON, code blocks, tool calls, or structured data. Write as a direct Telegram message.'
-  );
-
-  const client = await getFreshClient(config);
-
-  const controller = new AbortController();
-  const timeout = setTimeout(() => controller.abort(), EVENT_TIMEOUT_MS);
+  const proactiveContext = await buildProactiveContext(tenant, timezone, query).catch(() => '');
+
+  const contextPrefix = proactiveContext
+    ? `[Scheduled task context — ${timezone}]\n${proactiveContext}\n\n`
+    : '';
+  const prompt = `${contextPrefix}Scheduled task "${event.title}":\n${event.instructions}`;
+
+  const chatId = `scheduled-${event.id}-${Date.now()}`;
+  const { text } = await tenant.claude.chat(prompt, {
+    chatId,
+    userName: 'ScheduledTask',
+    userId: event.user_id,
+    userDir: tenant.userDir,
+    toolPrefs: tenant.toolPrefs,
+    config,
+    scheduler: tenant.scheduler,
+    messageLog: tenant.messageLog,
+  });
 
-  let response;
-  try {
-    response = await client.messages.create({
-      model: 'claude-sonnet-4-6',
-      max_tokens: 300,
-      system: systemParts.join('\n\n'),
-      messages: [{ role: 'user', content: event.instructions }],
-    }, { signal: controller.signal });
-  } finally {
-    clearTimeout(timeout);
-  }
+  tenant.claude.clearHistory(chatId);
 
-  const text = response.content.filter(b => b.type === 'text').map(b => b.text).join('\n').trim();
-  if (!text) throw new Error('Empty response from model');
+  if (!text?.trim()) throw new Error('Empty response from model');
 
-  await bot.api.sendMessage(event.chat_id, text).catch(() =>
-    bot.api.sendMessage(event.chat_id, text, { parse_mode: undefined })
-  );
+  await sendScheduledMessage(bot, event.chat_id, text);
 
   tenant.claude.injectHistory(event.chat_id, 'assistant', text);
   if (tenant.messageLog) {
@@ -361,4 +351,31 @@ async function runAgenticEvent(bot, config, event) {
   }
 }
 
+async function sendScheduledMessage(bot, chatId, text) {
+  const html = markdownToTelegramHtml(text);
+  if (html.length <= 4096) {
+    await bot.api.sendMessage(chatId, html, { parse_mode: 'HTML' }).catch(() =>
+      bot.api.sendMessage(chatId, text)
+    );
+    return;
+  }
+
+  let remaining = html;
+  while (remaining.length > 0) {
+    if (remaining.length <= 4096) {
+      await bot.api.sendMessage(chatId, remaining, { parse_mode: 'HTML' }).catch(() =>
+        bot.api.sendMessage(chatId, remaining)
+      );
+      break;
+    }
+    let splitAt = remaining.lastIndexOf('\n', 4096);
+    if (splitAt === -1 || splitAt < 2000) splitAt = 4096;
+    const chunk = remaining.substring(0, splitAt);
+    await bot.api.sendMessage(chatId, chunk, { parse_mode: 'HTML' }).catch(() =>
+      bot.api.sendMessage(chatId, chunk)
+    );
+    remaining = remaining.substring(splitAt).trimStart();
+  }
+}
+
 module.exports = { setupHeartbeat };

package/src/telegram/handlers/text.js

@@ -20,7 +20,7 @@ async function sendTtsVoiceSummary(ctx, tenant, responseText) {
     max_tokens: 200,
     messages: [{
       role: 'user',
-      content: `Summarize the following assistant message in 1-2 short spoken sentences. Use plain conversational language — no markdown, no code, no lists. Just what was said or done:\n\n${responseText.substring(0, 3000)}`,
+      content: `Summarize the following message in 1-2 short spoken sentences. Write in first person as if YOU are speaking directly to the user — say "I" not "the assistant". Use plain conversational language — no markdown, no code, no lists:\n\n${responseText.substring(0, 3000)}`,
     }],
   });
 

package/src/ws/server.js

@@ -466,7 +466,7 @@ async function generateTtsAudio(ws, tenant, responseText) {
     max_tokens: 200,
     messages: [{
       role: 'user',
-      content: `Summarize the following assistant message in 1-2 short spoken sentences. Write in first person as the assistant speaking directly. Use plain conversational language \u2014 no markdown, no code, no lists:\n\n${responseText.substring(0, 3000)}`,
+      content: `Summarize the following message in 1-2 short spoken sentences. Write in first person as if YOU are speaking directly to the user \u2014 say "I" not "the assistant". Use plain conversational language \u2014 no markdown, no code, no lists:\n\n${responseText.substring(0, 3000)}`,
     }],
   });