ai.
Notes on AI, building software, and things learned the interesting way.
running 4× Claude Code Max still isn't enough — here's what actually helps
A power user's read on caveman, context compression, and every "save tokens" plugin your feed is pushing. Most of them optimize the cheapest slice of the bill.
reverse-engineering Warp's cli-agent notification protocol
Warp's agent sidebar runs on one OSC 777 escape sequence and a six-field JSON envelope. here's the full protocol, reverse-engineered from the three official adapter repos.
using remote SSH with Codex app
Codex app now has alpha SSH remote support, but during rollout you may still need the hidden feature flag, the beta appcast build, or a shell-init fix.
Tip: enable 1M context in Codex with one command
Run one command, let the script create a backup of ~/.codex/config.toml, patch the 1M settings if needed, then restart Codex.
using claude code's new native ssh remote on a mac mini / darwin
Claude Code's SSH remote rejects darwin hosts, but a lightweight OrbStack Linux VM on a Mac Mini turns the whole thing into a workable remote dev setup.
auto-approve Claude Code plan mode via PermissionRequest
Claude Code's plan approval dialog can be auto-approved by hooking PermissionRequest correctly and matching ExitPlanMode instead of Stop.