Hook Functions

Dayu uses a registry-based hook mechanism so policies and application behavior can change without replacing the stable runtime service loops. Implementations register an alias with ClassFactory; templates and visualization configs select that alias at runtime.

Resolution model

Part	Role
ClassFactory	Registers one implementation under a ClassType and alias.
Context.get_algorithm()	Reads the selected name and parameters and creates the hook instance.
dependency/core/lib/algorithms/__init__.py	Imports each hook-family package so discovery starts during core initialization.
Shared hook loader	Imports public single-file hooks and explicit package entry points from dependency/core/lib/algorithms/loader.py.
Templates and YAML catalogs	Select aliases through <TYPE>_NAME, <TYPE>_PARAMETERS, lists, or hook_name.

Most single hooks use this pattern:

- name: GEN_FILTER_NAME
  value: simple
- name: GEN_COMPRESS_NAME
  value: simple

Processor scenario extractors and Monitor metrics use alias lists, while result and system visualizations select one hook_name per visualization item.

Package discovery

Each hook-family package examines only its direct children. A public .py file is imported as a flat hook entry. A multi-file implementation must provide <package>/hook.py; that entry point imports and exposes its registered implementations through __all__. Internal model, DRL, utility, and ablation modules are not discovered recursively, and package __init__.py files should remain lightweight.

Hook families that declare optional dependencies skip an entry only when an external package is unavailable and emit a warning. Missing core.* imports still fail immediately so repository errors are not mistaken for optional features.

Main hook groups

Area	Examples
Generator	Getter filter, frame filter/process/compress, before-schedule, after-schedule, getter, and before-submit hooks.
Scheduler	Configuration extraction, scenario/policy retrieval, startup, source selection, deployment, redeployment, and agents.
Processor	Processor family, queue discipline, and scenario extraction.
Monitor	Resource and queue-state samplers.
Visualization	Result and system visualizers selected by YAML.

Add an implementation

1. Implement the base signature used by the caller.
2. Use a public .py file for a self-contained hook, or a package with hook.py for a multi-file implementation or related variants.
3. Register each class with @ClassFactory.register(ClassType.<TYPE>, alias="<name>") and expose it through the entry module's __all__.
4. Import package internals explicitly from hook.py; do not rely on recursive discovery or eager package imports.
5. Expose the alias in a template, list, or visualization config.
6. Document constructor parameters and optional dependencies, then add focused tests and update the hook catalog.

The call order is part of the runtime contract, especially for Generator identity reservation and Scheduler deployment validation. Consult the implementation-facing Hook Guide and Hook Catalog before adding an alias.