Agently Docs

EN

中文

EN

中文

Appearance

Resize and Strategy Extension ​

Applies to: 4.0.8.1+

resize is the entrypoint for Session window governance:

  • it controls context_window
  • it can also update memo

1. auto_resize decision flow ​

How to read this diagram ​

  • the analyzer decides whether resizing should happen and which strategy name to use
  • the executor actually mutates context_window and memo

2. Default behavior: max_length + simple_cut ​

The default analyzer:

  • returns simple_cut when session.max_length is an integer and the window is over limit

The default executor simple_cut:

  • keeps messages from newest backward until total length is within max_length

Example:

python
agent.set_settings("session.max_length", 12000)

3. Automatic and manual triggers ​

  • automatic: when auto_resize=True (default), resize runs after history writes
  • manual: call session.resize() whenever you want
python
agent.activate_session(session_id="resize_demo")
session = agent.activated_session

session.resize()
# await session.async_resize()

4. Custom strategy: keep last N rounds ​

python
def analysis_handler(full_context, context_window, memo, session_settings):
    if len(context_window) > 10:
        return "keep_last_ten"
    return None


def keep_last_ten(full_context, context_window, memo, session_settings):
    return None, list(context_window[-10:]), memo


session.register_analysis_handler(analysis_handler)
session.register_execution_handlers("keep_last_ten", keep_last_ten)

5. When to disable automatic resize ​

If you want to batch writes and compress later in one pass:

python
from agently.core import Session

session = Session(auto_resize=False, settings=agent.settings)
# ... batch writes
session.resize()

6. Common pitfalls ​

  • setting session.max_length without activating a session: no effect
  • returning a strategy name that was never registered: warning, no update
  • returning the wrong executor shape: always return (new_full, new_window, new_memo)