Igor Ganapolsky IgorGanapolsky

Hermes Mobile — phone remote for AI agents on YOUR Mac

Not another cloud agent. Control Cursor / Claude Code / Hermes gateway from your phone.

Free chat to your Mac gateway (Tailscale when off Wi‑Fi)
Leash: approve/deny dangerous tools from the couch (10 free approvals/week)
Keys stay on your machine

Install (Android live)

Hermes Mobile Privacy Policy

Effective date: July 4, 2026

Hermes Mobile is published by Max Smith KDP LLC. Hermes Mobile lets you connect your phone to your own computer so you can monitor and control Hermes workflows from the mobile app.

Data We Collect

Hermes Mobile collects limited data needed to operate and improve the app:

AI Agent Reliability Checklist

The 12-point audit for teams running AI coding agents in production.

Why this exists

After 6 months of ThumbGate telemetry across Claude Code, Cursor, and Codex installations, we identified the 12 failure patterns that cause 94% of all agent incidents. This checklist helps you find them in your setup before they cost you money.

Stop Your AI Agent Before It Breaks Production

The Problem

AI coding agents (Claude Code, Cursor, Codex) make the same categories of mistakes:

Editing files outside task scope — agent "helpfully" refactors a file you didn't ask it to touch
Hallucinating function signatures — calls a function with wrong params, breaks the build
Making unauthorized changes — modifies config files, env vars, or deployment scripts without confirmation
Ignoring instructions — skips test files, deletes comments, reformats code you told it not to touch

The 4 Stripe Connect failure modes that quietly cost marketplaces real money

A field teardown for anyone building split payments on a platform — and how to catch each one before it hits a real payout.

Most Stripe Connect integrations demo perfectly and then lose money in production, because the failure modes don't show up until real volume, real retries, and real disputes arrive. Here are the four I see most, ranked by blast radius, with the check that catches each one.

1. Non-idempotent charge creation → double charges on retry

A network blip makes your server retry a PaymentIntent/charge create. Without an Idempotency-Key, Stripe happily creates two. The customer is charged twice; you eat the refund and the dispute. Catch it: every create/capture call carries a stable idempotency key derived from your order id (not a random per-attempt UUID). Test it by firing the same request twice and asserting one charge.

LipoShield — Support

For questions, bug reports, or feedback about LipoShield, contact:

Email: iganapolsky@gmail.com

LipoShield is an educational wellness tool. It does not provide medical advice. For health concerns, please consult a qualified healthcare professional.

Privacy Policy: https://gist.github.com/IgorGanapolsky/7e5c32cd76f1cf8b31c963f4b0efe584 Terms of Service: https://gist.github.com/IgorGanapolsky/aaa78f12fbcef4824b2d4e9bd0dd7d8a

	"""Hyperagent (hyperagent.com + Nov 3, 2026 Product Update) Architectural Steal for Swappa/ShopGoodwill Desk.

	Implements the 7 core systems extracted from https://hyperagent.com/thread/cmql52i8o0dd606adahnyadrt?autostart=onboarding-four-paths,
	/Users/iganapolsky/Downloads/hyper.pdf, and https://www.hyperagent.com/llms.txt:

	1. Onboarding Four Paths + 5-Agent Physical-Goods Arbitrage Team (`ByteHedge`, `ChronoSim`, `ShieldGrid`, `VaultBot`, `PulseRay`)
	& 16-Stage Arbitrage State Machine (`DISCOVERED` -> ... -> `SOLD`).
	2. Rubrics & Blind Judge Evaluator (`rubrics-and-evaluations.md`):
	- 3-7 weighted criteria (auto-normalized to 1.0) with 1-to-5 scoring guides and hard-failure gates.
	- Blind Judge separation: scores ONLY output + rubric (never sees internal reasoning or self-grades).

	#!/usr/bin/env python3
	"""
	yolo-roast — Zero-install Viral "Mac Agent Stack Roaster" & 1-Command Critique CLI.

	Inspired by `npx ccusage` / `neofetch`:
	1. Scans live macOS process tree (`ps`, `vm_stat`, `sysctl`) in <200ms.
	2. Evaluates the 5 real-world macOS coding-agent failure signatures documented in
	public GitHub issues (e.g., anthropics/claude-code#11315 129GB RAM freeze,
	#4953 120GB OOM, #67433 500MB/min idle leak, #82941 kalloc.1024 kernel panic,
	#98864 renderer bloat, and orphaned stdio MCP servers).

	{
	"name": "GCS resumable upload - fail-closed verification workaround",
	"nodes": [
	{
	"parameters": {},
	"type": "n8n-nodes-base.manualTrigger",
	"typeVersion": 1,
	"position": [-720, 0],
	"id": "e5fd6958-6a52-49fa-b4e9-fb94fa74171f",
	"name": "When clicking Test workflow"