ANDRE LUSTOSA / WRITING
A working library.
Notes on building dependable software, contributing upstream, and helping engineering teams move with confidence.
Archive
All writing
12 entries
AI Agents in CI/CD: Security, Robustness, and Productivity
A graduate-level study of agentic CI/CD: authority and prompt injection, bounded verification, distributed recovery, test oracles, and causal measurement of productivity.
Build Pipelines Should Explain Their Decisions
A pipeline becomes more useful when each stage records the evidence behind release, rejection, and later maintenance decisions.
Bounded Autonomy Requires Separate Control Planes
Agentic automation remains governable when model choice, execution, policy, and evidence are separate interfaces with independent owners and tests.
Hermetic Builds Begin at the Acquisition Boundary
A build becomes explainable when every network-dependent input is acquired explicitly, recorded, and separated from the steps that construct the artifact.
A Linux Updater Is a Deployment Transaction
Safe self-updates preserve a compatible installed set through staging, validation, privileged activation, startup checks, and recovery.
Serving llama.cpp Across My Home Network and Tailscale
Building llama.cpp on Fedora with Vulkan, then sharing a 27B model with friends and other machines across my home network and Tailscale.
When Repeated Release Work Deserves a Shared Stream
A worked release example shows how to separate shared build policy from product decisions, preserve immutable identity, and decide whether the added platform is worth owning.
The Case for Publishing a Useful Package Early
PyPI names are scarce, pending publishers do not reserve them, and active conflicts are hard to unwind. A small working release can be the responsible first public commitment.
How to Trace an Agent That Can Crash or Be Killed
The supervisor must own trace identity, timeouts, and final status because the agent may never get a chance to report how it stopped.
A Release Has Three Trust Decisions
Version identity, permission to publish, and artifact integrity answer different questions. A release pipeline is trustworthy only when it connects all three.
Testing the Package Beyond the Source Tree
Several upstream fixes point to one packaging rule: exercise built artifacts, disabled paths, and dependency boundaries directly.
One Review Inbox for Two Forges
tongs treats code review as one terminal workflow across GitHub and GitLab, then designs each layer around the differences that remain.