Context coverage is the share of a codebase that has documented, code-anchored team knowledge attached to it — a measure of how much of the system a new engineer or AI agent can be briefed on before they touch it. High coverage means most areas carry their decisions, gotchas, and invariants; low coverage means large stretches of code that anyone must read cold and reason about from scratch.
The metric matters because it makes an otherwise invisible gap legible. A team rarely knows which parts of its system are well understood and which depend on a single person’s memory until something breaks there. Expressing coverage against the files and modules that actually exist turns "we should document more" into a concrete map of where context is present, where it is stale, and where it is missing entirely.
Coverage is a means, not an end, and it can be gamed: a single over-broad note claiming to cover everything inflates the number while explaining nothing. Useful coverage comes from narrow, well-anchored notes — one concept pointed at the handful of files it genuinely concerns — so the areas counted as covered are areas a reader is actually briefed on, not areas a catch-all happens to match.