Shadow Execution
For some actions, the fastest way to find out whether a change is safe is to actually try it — somewhere it can't do damage. v0.0.8 adds Shadow Execution: running a proposed change in a fully local, disposable copy of its environment before recommending the real thing.
Two supported domains
v0.0.8 ships two shadow-execution domains, both deliberately chosen because they can be shadow-executed with zero real external systems touched:
- File mutation — a proposed file change is applied to an isolated staged copy of the working tree, never the real one.
- SQLite migration — a proposed schema migration runs against a throwaway, file-backed database, never a real one.
Infrastructure-plan and deployment-configuration domains were explicitly declined for this release, since either would need real credentials or network reachability that this feature's own safety rules forbid by default. That's a documented scope decision, not an oversight.
What it actually caught
Shadow execution found two real fragilities in this release that passive evidence alone would not have reliably caught: a syntactically-broken JSON mutation that only surfaced when parsed inside the isolated copy, and a unique-index migration that only failed against representative seeded duplicate data.
Safety, tested adversarially
Every shadow runner operates only against local, disposable resources — temp directories, throwaway databases. Negative controls are real, adversarial tests, not just happy-path checks: a path-traversal attempt is refused rather than silently clamped; a symlink pointing outside the sandbox is never followed (proven by placing a real secret file outside the sandbox and confirming it stays untouched after an attempted write); network-shaped and credential-shaped targets are refused outright, regardless of how valuable the information would be.
A passing shadow run is explicitly not equivalent to a production proof — every result carries fidelity/coverage metadata naming exactly what it does and does not model.
See also
- Adaptive Verification Budgets — shadow execution is one verification tier a budget can escalate into.