Context management

What the model remembers. Module: regista/context/.

The trigger is provider-reported usage, never a local token estimate: when a turn's observed input tokens (input + cache read + cache write) cross max_input_tokens, the loop compacts — it asks the session's own provider to summarize the older history and replaces those messages with one summary message, keeping the most recent messages verbatim.

from regista import Agent, ContextConfig

agent = Agent(
    provider=...,
    instructions=...,
    context=ContextConfig(
        max_input_tokens=150_000,   # None (default) disables compaction
        keep_recent_messages=4,     # preserved verbatim at the end of history
    ),
)

Rules the implementation keeps:

• A split never separates an assistant tool_use from its tool_result reply.
• Thinking blocks are excluded from the transcript sent to the summarizer.
• The summarization call is a regular, hash-verified llm.request/llm.response pair followed by a context.compaction event, and its usage/cost roll into the session totals.

Why compaction goes through the provider seam

Because the summary flows through the same provider interface as everything else, a replayed session re-runs the same compaction logic, is served the recorded summary, and every post-compaction request hash still matches. ContextConfig is recorded in session.start precisely so replay can reconstruct it.