Technical notes on version-controlling AI context — architecture, patterns, and the occasional opinion from the people building it.
Six situations teams actually hit when they adopt Contexo on a project that already exists — a mature repo with docs, a codebase with none, an overgrown CLAUDE.md, a wiki in another repo — plus what happens after the import, when a generated scaffold meets reality and a teammate corrects an imported page.
Third in the Under the hood series: the two on-ramps that fill an empty Contexo repo. How ctx migrate discovers an existing knowledge base and ctx generate manufactures one from a code graph — sharing a single Detector/Catalog engine, a text directive, and one manifest, with the agent doing every judgement call.
Second in the Under the hood series: a code-level walk through Contexo's data path. What actually crosses the wire when your agent pushes (a list of pages, not your transcript), why the raw session buffer never leaves your laptop, how your tokens are hashed, and the honest answer to 'does Contexo keep my API keys?'
A deep dive into Contexo's engine: how the CLI and server stay LLM-free by handing distillation to the agent already in the loop (the push handshake), and how section-aware diffs make a page's history legible in a way git diff can't.
Six concrete scenarios where Contexo earns its place — walked through step by step, from the commands you run to the MCP handshakes your agent handles. Every step maps to what Contexo actually does today.
The one idea behind Contexo, and why we're starting this blog: the context your AI agents run on is code now — so it deserves diffs, history, and one shared source of truth. Here's the “context tax” it solves, what Contexo is, and how it works.