- Why it matters if true
- Agent skill/plugin ecosystems let a capability surface change out from under a pinned agent version — new tools, removed tools, changed permissions — with no signal to the agent or the person relying on it. If that drift is large enough and frequent enough, a lockfile (a recorded, diffable snapshot of what capabilities were available and trusted at build time) is worth building. If it isn’t, this is solving a problem nobody has yet.
- Instrument
capdiff — snapshots a plugin/skill surface, diffs it against a later snapshot, and classifies the delta. agentlock is the lockfile product this experiment would justify; it is not itself the entry being tested here — the claim is.- Kill threshold
- TODO: fill in — state the threshold as pre-registered, or say plainly none was set
- Results
- Instrument built. Number not yet trusted — closing the remote-source coverage gap first. TODO: fill in real numbers once trusted
- What I don’t trust yet
- Remote-source coverage is incomplete, so any drift rate measured right now would undercount. Nothing quantitative should be published from this until that gap closes.