alohomora

🐉 Alohomora

"Your guard locked a door it was never asked to watch."

The asymmetry is the whole point: a noisy advisory costs tokens, but a false-positive blocking guard costs the job. And it does not merely stop the work — it teaches agents to route around it, in ways nobody reviews and no test covers. Every over-broad guard trains the exact behaviour the guard exists to prevent, then hides the evidence, because a workaround leaves no failure to find.

What to scan

• Every guard that refuses rather than warns — pre-tool hooks, pre-commit gates, CI blockers
• Their path/scope resolution: what root they resolve against, and what they do with anything outside it
• Their block messages, read as instructions — agents obey block text verbatim
• Their escape hatches: existence, TTL, attribution, ceiling
• Their test suites, specifically the negative cases: the legitimate inputs they must not block

This spell's subject is guards you wrote. On a repo whose only guards are husky, lint-staged and branch protection, the prescriptions below do not apply — you cannot cap git commit --no-verify, and branch protection's escape is an admin override the platform already attributes in its audit log. Report N/A — no repo-owned blocking guards, and say that is what you found rather than reporting a clean sweep of an empty set.

Anti-patterns to find

1. Out-of-scope paths resolved as in-scope. Relative, home-prefixed (~/…) or symlinked paths resolved against the wrong root. Measured example: a repo-isolation guard blocked a write to a directory outside the repository entirely, because it resolved a ~-prefixed absolute path as repo-relative — the guard's own subject was never involved.
2. A shared-resource guard applied to work that does not touch the resource. A stage-everything guard firing inside a different repository; a lease check on an operation that takes no lease. The rule is right; its domain is wrong.
3. Matching on the symptom instead of emitter AND condition. The guard keys on a surface string rather than on who emitted it and under what condition — so the eventual narrowing suppresses real cases along with false ones, and nobody can tell which.
4. No escape hatch, or an unbounded one. Either the guard has no override at all — so the only route past it is a workaround nobody reviews — or the override is untimed, unattributed and uncapped, which is not an escape but a permanent hole with better manners.
5. A block message naming a dead route. Block text is an instruction agents follow literally. A remedy that no longer exists is worse than no remedy: it sends every blocked agent down a path that fails a second time, and the second failure looks like the agent's fault.
6. Severity mismatched to consequence. Blocking where an advisory would do (refusing legitimate work over a stylistic rule), and advising where only a block will hold (a destructive or irreversible action behind a warning agents can and do proceed past).