Latest note
Codex auto review can get expensive
Auto review is useful safety infrastructure, but it can quietly become a large part of Codex usage when tasks keep crossing protected boundaries.
Read noteNotes, when they happen.
Problems, decisions, and discoveries from ongoing systems work, written down when they seem worth remembering.
Selected notes
A curated path through investigations, measured debugging, and recovery work.
A note on performance in the registry control plane
Reducing repeated registry work with cached catalog data, paginated tag loading, and event-driven cache warming.
Why Llama 4 took minutes to load
What looked like a context-size or memory problem on my Strix Halo system turned out to be a model-loading problem tied to mmap.
The runner was the recovery path
Using a workflow runner to recover access to a host, and why runners need to be treated as serious infrastructure.
Explore by subject
Registry Control Plane
Registry operations, delivery, and the project behind them.
Local AI
Running and evaluating AI systems on local hardware.
OpenWrt and networking
Build infrastructure and practical network systems.
Self-hosting
Operating services and the systems around them.
AI security
Practical safety boundaries for AI-assisted systems.
Archive
14 notes2026
The cheapest model call is the one you avoid
How deterministic retrieval, reusable semantic context, and smaller models can reserve frontier-model attention for the work that needs it.
A code graph is not repository memory
What code graphs can tell an agent, what real repository questions demand, and why useful memory needs a semantic layer.
Giving Coding Agents a Map
An introduction to RepoMemory, a local context layer that helps coding agents find the right part of a repository.
Using Ollama models with Codex providers
Why the goal is mixed-provider Codex delegation, and where today’s provider path breaks down.
Prompt injection has a lot in common with SQL injection
Why prompt injection is dangerous, what adversarial testing can catch, and why passing those tests is not a guarantee of safety.
Why Llama 4 took minutes to load
What looked like a context-size or memory problem on my Strix Halo system turned out to be a model-loading problem tied to mmap.
Building a remote MCP time server
Building a remote MCP server so AI agents can ask for real time, timezone, and duration data instead of guessing.
The runner was the recovery path
Using a workflow runner to recover access to a host, and why runners need to be treated as serious infrastructure.
A note on performance in the registry control plane
Reducing repeated registry work with cached catalog data, paginated tag loading, and event-driven cache warming.
Registry Control Plane is public on GitHub
Registry Control Plane is now public on GitHub, along with its container packages.
Making the registry control plane more responsive
A short note on making the registry control plane easier to use on phones and smaller tablets.
Adding a local cache to a Forgejo runner
Using a temporary local BuildKit cache in a Forgejo runner to speed up repeated registry control plane builds on limited hardware.
Working on a registry control plane
A May 2026 snapshot of building a registry management layer around Docker Distribution.
A quick note on my ASU builder changes
Operational changes that made my OpenWrt ASU builds more reliable when the official builders were overloaded.