You Don't Need a Pre-Session Hook - Human Judgment Catches What Hooks Miss
You Don't Need a Pre-Session Hook
The temptation is understandable. You want to add a pre-session hook that validates the agent's context, checks for stale state, and ensures everything is aligned before the session begins. It sounds like good engineering.
But the human who notices something is wrong is doing a job that no hook can do - understanding whether the output is correct in context.
What Hooks Can Check
Automated hooks are good at mechanical validation. They can verify that files exist, dependencies are installed, environment variables are set, and the git state is clean. These are binary checks with deterministic answers.
Pre-session hooks can also run linters, type checkers, and test suites. They can confirm that the codebase compiles and that nothing is obviously broken. All of this is useful but insufficient.
What Hooks Cannot Check
Hooks cannot determine whether the agent's approach makes sense for the problem. They cannot tell you that the feature you built yesterday actually solves the wrong version of the user's need. They cannot notice that the architecture decision from three sessions ago is causing subtle problems now.
This is judgment, not validation. It requires understanding the goal, the constraints, and the tradeoffs - things that change with context and cannot be reduced to a script.
The Human Review Layer
Instead of investing in complex pre-session automation, invest in a simple habit: spend two minutes reviewing what the agent did last session before starting the next one. Read the git log. Glance at the diff. Ask yourself whether the direction still makes sense.
This two-minute review catches problems that would take hundreds of lines of hook code to approximate - and it catches them more reliably because you understand the intent behind the work.
When Hooks Do Help
Use hooks for the mechanical stuff. Let them catch the easy errors. But do not mistake mechanical validation for judgment. The human in the loop is not a bottleneck - they are the quality gate that makes everything else work.
Fazm is an open source macOS AI agent. Open source on GitHub.