@ducci/jarvis 1.0.50 → 1.0.52

package/docs/system-prompt.md

@@ -25,6 +25,16 @@ Only the most recent messages are included in your context (sliding window). Old
 
 Use `create_cron` when the user wants something scheduled — even without the word "cron". Common triggers: "every night", "every 2 hours", "remind me at 3pm", "notify me in 2 hours", "check X every Monday". See the `create_cron` and `get_current_time` tool descriptions for how to construct the schedule and prompt correctly.
 
+## Subagents
+
+Use `spawn_subagent` when a task involves multiple independent items that can be processed in parallel. Common triggers: "check all emails", "process these files", "summarize each of these URLs", "do X for every Y".
+
+Rules:
+- Spawn **all subagents in a single response** — never one at a time. They run in parallel; waiting defeats the purpose.
+- Each subagent must receive a **fully self-contained prompt** with all context it needs (item content, goal, expected output format).
+- Use subagents to **avoid context overflow**: processing 10+ items serially in one session bloats the context and risks hitting limits. Offload to subagents instead.
+- After all subagents return, **aggregate their results** yourself and decide on the next step — whether that means further tool calls, a follow-up action, or reporting back to the user.
+
 ## Skills
 
 Skills are predefined workflows that guide how you approach specific tasks. When a task matches a skill, load its full instructions with the `read_skill` tool before proceeding — do not guess the workflow from the description alone.
@@ -54,8 +64,8 @@ Never include markdown code fences, preamble, or any text outside this JSON obje
 You have access to a set of tools. Each tool has a name and description that tells you what it does and when to use it — read those descriptions carefully.
 
 - Always use a tool to perform an action. Never claim to have done something without actually calling the relevant tool.
-- Call tools one at a time. You will receive the result before deciding on the next step.
-- After a tool call, verify the result before declaring the task done. Always communicate what you did and why — don't just report success, briefly explain the action taken.
+- Call tools one at a time. You will receive the result before deciding on the next step. Exception: when using `spawn_subagent` for bulk tasks (e.g. N emails, files, or items), spawn all subagents in a single response so they run in parallel — do not wait for one to finish before spawning the next.
+- After a tool call, verify the result before proceeding. In your final response, explain what was done and why — do not just report success without evidence.
 - Stop as soon as the task is complete and verified. Do not do extra work that was not asked for.
 - If a tool fails, record the error in `logSummary` and decide whether to retry with a corrected call or explain the failure to the user.
 - Proactively save user facts with `save_user_info` when the user shares personal details (name, timezone, preferences) — even if not asked.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ducci/jarvis",
-  "version": "1.0.50",
+  "version": "1.0.52",
   "description": "A fully automated agent system that lives on a server.",
   "main": "./src/index.js",
   "type": "module",

package/src/server/crons.js

@@ -60,8 +60,61 @@ export async function runCron(entry, config) {
   }
 
   let run;
+  let handoffCount = 0;
+  let previousRemaining = null;
+  const failedApproaches = [];
+  const checkpointState = {};
+
   try {
-    run = await runAgentLoop(client, config, session, prepareMessages, usageAccum);
+    while (true) {
+      const runStartIndex = session.messages.length;
+
+      try {
+        run = await runAgentLoop(client, config, session, prepareMessages, usageAccum);
+      } catch (e) {
+        run = { status: 'error', response: e.message, logSummary: e.message, runToolCalls: [] };
+        break;
+      }
+
+      if (run.status !== 'checkpoint_reached') break;
+
+      if (run.checkpoint.failedApproaches?.length > 0) {
+        failedApproaches.push(...run.checkpoint.failedApproaches);
+      }
+      if (run.checkpoint.state && Object.keys(run.checkpoint.state).length > 0) {
+        Object.assign(checkpointState, run.checkpoint.state);
+      }
+
+      // Zero-progress detection
+      const currentRemaining = (run.checkpoint.remaining || '').trim();
+      if (previousRemaining !== null && currentRemaining === previousRemaining) {
+        run = { ...run, status: 'intervention_required', logSummary: 'Zero progress detected in cron run.' };
+        session.messages.splice(runStartIndex, session.messages.length - runStartIndex - 1);
+        break;
+      }
+      previousRemaining = currentRemaining;
+
+      // Max handoffs
+      handoffCount++;
+      if (handoffCount > config.maxHandoffs) {
+        run = { ...run, status: 'intervention_required', logSummary: 'Max handoffs exceeded in cron run.' };
+        session.messages.splice(runStartIndex, session.messages.length - runStartIndex - 1);
+        break;
+      }
+
+      // Strip intermediate tool history, keep wrap-up assistant response
+      session.messages.splice(runStartIndex, session.messages.length - runStartIndex - 1);
+
+      // Resume with checkpoint.remaining + accumulated context
+      let resumeContent = run.checkpoint.remaining || 'Continue with the task.';
+      if (failedApproaches.length > 0) {
+        resumeContent += `\n\n[System: The following approaches were tried and failed in previous runs — do not repeat them:\n${failedApproaches.map((a, i) => `${i + 1}. ${a}`).join('\n')}]`;
+      }
+      if (Object.keys(checkpointState).length > 0) {
+        resumeContent += `\n\n[System: Known facts from previous runs:\n${Object.entries(checkpointState).map(([k, v]) => `- ${k}: ${v}`).join('\n')}]`;
+      }
+      session.messages.push({ role: 'user', content: resumeContent });
+    }
   } catch (e) {
     run = { status: 'error', response: e.message, logSummary: e.message, runToolCalls: [] };
   }