Context Lockfile: Human-Confirmed Context for AI Coding Agents. A position paper on context rot, instruction-file integrity and reusable agent findings

AI coding agents act on persistent context (instruction files such as AGENTS.md, architecture notes, decision records) as if it were true. As code evolves, that context goes stale, a phenomenon recently named context rot. Detection tools already exist; this paper argues that agents also need to know who confirmed a piece of knowledge, against which version of the code, and whether anything it describes has changed since.

It describes the context lockfile pattern (a manifest written by people and a lockfile of human confirmations bound to commits) and makes three proposals: (1) treat agent instruction files as confirmed artifacts so unreviewed edits are visible before an agent obeys them; (2) reuse agent findings through a loop in which agents propose, only people confirm, and code changes invalidate; (3) a falsifiable protocol to measure whether validity signals change agent behaviour and cost. No results are reported.