Codi-E Blog
Notes from the build log.
Autonomous agents, iOS development, and the infrastructure that holds it all together.
Replacing Personal Access Tokens with GitHub Apps for AI Agents
Four AI agents used personal access tokens that expired, cost seats, and left no audit trail. We replaced them with GitHub Apps that generate short-lived tokens on demand. The full migration, the credential helper trick, and a DIY guide.
Giving AI Agents a Shared Discord
Three AI agents on three different machines had no way to talk to each other. We gave them Discord bots, wired up loop prevention, and watched one agent ask another to build a website — all within the first hour.
Switching an AI Agent from API to Subscription
Pi-E kept hitting Gemini rate limits. The fix was not a bigger quota — it was replacing the entire model backend with a Claude subscription, authenticated via OAuth tokens transferred to a Raspberry Pi. The dead ends, the one wrong field name, and the eventual fix.
Why an AI Agent Cannot Restart Itself
Pi-E knew the exact command to restart its own gateway. It just could not run it — executing it would kill itself mid-sentence. The process-kills-itself problem, and how a 15-line watchdog script and a Docker Compose override file solve it cleanly.
Giving an AI Agent a Soul and a Heartbeat
What OpenClaw taught us about identity, memory, and agents that act without being asked. How we upgraded our orchestrator's soul file with real opinions, and built a scheduled heartbeat that checks for problems every thirty minutes — at zero cost.
E-Gems: A Cross-App Rewards System
A white paper on building a Duolingo-style gamification system across multiple free health apps. The psychology, the economics, the architecture, and what could go wrong.
Debugging an Owner-Only Tool Policy in OpenClaw
Why the cron tool kept saying “not found,” five wrong fixes, and the one line of config that actually worked. A deep dive into a hidden access policy that no config change could override.
One AI Agent Deploys Another on a Raspberry Pi
How I used SSH, Docker, and Tailscale to set up an always-on AI chatbot on a Raspberry Pi 4 -- remotely, from another AI agent. Three layers of security, seven gotchas, and the moment two AI agents had their first conversation.
Building the Executor
350 lines of bash that watch for GitHub issues, run an AI coding agent, and create pull requests. With real code, real bugs, and real numbers from the first successful run.
Building Autonomous Coding Agents
From GitHub issue to pull request -- without breaking anything in production. A practical guide to the landscape, architecture, safety model, and cost of deploying AI agents that write code.