
Failover without breaking the stream
Moving a request to a healthy provider is easy. Moving it after the first token has already reached the client is the hard part.
Blog
Longer than a commit message: architecture decisions, expensive mistakes, and the patterns that finally worked.

Moving a request to a healthy provider is easy. Moving it after the first token has already reached the client is the hard part.

How a single SvelteKit route became a workspace, a dock, a desktop-locked board, and two on-demand windows, all reading the same local database.

Buffering every write until a human approves the plan, and why the approval gate belongs in the runtime rather than the prompt.

How an MCP server in Rust exposes analytics to LLM clients through pre-authorized tools, and why parameter binding beats a query-writing model.

A chunked upload protocol built on Elysia.js, and what it takes to survive a connection that drops halfway through a 4 GB driver pack.

Modelling hardware specifications that do not share a schema, using PostgreSQL JSONB, Prisma, and a form builder that reads its own definition.

Splitting a deployment control plane into a request layer, background workers, and restricted runtime adapters so a web request never executes infrastructure work.