guardrails

Your Bash prompt runs code on every render, so unsafe PS1/PROMPT_COMMAND snippets can leak secrets, be hijacked by crafted dirs/branches, or slow your terminal. This guide shows how to harden it: prefer built-ins, sanitize untrusted text (PWD/VCS), pin absolute paths, move logic to safe PROMPT_COMMAND (array, no eval), keep root ultra-minimal, and use packaged frameworks (e.g., Starship). Includes copy-paste Bash and a quick audit checklist.

A practical guide to building reliability-first AI agents on Linux with Bash and standard tools. It explains common failure modes (probabilistic AI, flaky networks, crashes, low visibility) and offers copy-paste patterns: hardened entrypoint, timeouts/retries/backoff, idempotency/state, structured JSON logs, and systemd sandboxing/self-heal—plus a ready-to-run inbox-to-outbox summarizer using curl/jq and ops tips.