Hooks

Hooks are modules that run “around” the agent loop to add cross-cutting behavior: logging, approvals, redaction, UI/streaming presentation, and context injection. They’re the closest thing Amplifier has to middleware.

Why hooks matter

Observability: capture a canonical event stream for debugging and tooling.
Safety and governance: approvals and redaction can be enforced consistently.
UX: streaming UI behavior can be swapped without touching the kernel.

Existing hooks (scanned)

This list is rendered from data/catalog.json.

How to think about a hook chain

You can picture hooks as a pipeline around “one turn”:

flowchart LR req["Request (prompt + context)"] --> inj["Context injection hooks"] inj --> approve["Approval/redaction hooks"] approve --> provider["Provider call (model)"] provider --> stream["Streaming UI / output formatting"] provider --> log["Logging/event stream"] stream --> end["User sees output"] log --> end

Important

Hooks affect what gets logged, what the model sees, and what actions are allowed. If you’re auditing behavior, start by checking which hooks your profile enables.