I think command-line tools should emit a small, machine-readable execution manifest: which steps actually ran locally, which were satisfied by a local or remote cache, which contacted a remote service, and which invoked a plugin or AI provider. This is execution provenance, not surveillance telemetry. A step ID, status, execution location, cache source, service category, timestamp, and input/output digest would be enough to make many failures explainable.
The practical payoff is straightforward: debugging a surprising result, reviewing privacy boundaries, investigating an incident, and comparing builds across machines. SLSA and systems such as Bazel’s Build Event Protocol show useful precedents, though neither is a universal record of every interaction. I’d favor a one-line terminal summary plus an opt-in detailed view, with the manifest exportable as a small local file.
There are real downsides: noisy output, sensitive endpoint names, and the false confidence that a record proves reproducibility. Would you trust or actually use this data? Counterexamples and alternative designs welcome.