17:T41b,OP here.

I’ve been running autonomous agents (OpenClaw/ReplyX) for lead gen, and I kept waking up to drained API credits because bots would get stuck in "zombie loops"—retrying a failed step (like a button click) until the timeout hit.

I actually wrote the fix for the loop detection logic in the OpenClaw repo itself (PR #9759), but I realized I needed a better way to visualize what was happening than scrolling through thousands of lines of JSON logs.

Empusa is a local dashboard (Next.js + SQLite) that helps with this:

Visual Timeline: Turns logs into a visual trace (Green=Success, Red=Fail, Yellow=Loop Detected).

State Snapshots: It records the agent's variable state at every step.

Resume Capability: If an agent crashes, you can grab the state JSON and resume execution from that exact step rather than restarting from zero.

It’s a simple MVP right now, but it solves the specific pain of "blind" agent debugging. Code is MIT licensed.

Would love feedback on the state management approach.18:T5e3,Hey HN,

I built MindDraft because I was tired of task apps that made ME do all the work. After 18 months of solo development, I'm sharing it here.

*What's different:*

1. Location-based reminders - "Buy groceries" alerts when you're near a grocery store, not at an arbitrary time. Uses iOS geofencing.

2. Smart Actions - Tasks have action buttons. "Call dentist" → tap to call. "Pay rent" → opens payment app. Tasks become actions, not just reminders.

3. Task → Expense - Complete "Buy laptop $1200" and it auto-logs as an expense. No separate expense app.

4. Natural language AI - "Meeting with John Friday 3pm, need to prep slides day before" → parsed correctly with subtask.

*Tech stack:* - SwiftUI + SwiftData - Firebase (Auth, Firestore, Cloud Functions) - Gemini 2.5 Flash for NLP - Native iOS (no React Native/Flutter)

*What I learned:* - Geofencing on iOS is powerful but battery-

Human tech tree MVP using codex