Skip to main content
Version: 0.7.0

mellea.stdlib.context.chat

Chat-style context with pluggable compaction.

Classes​

CLASS ChatContext ​

Chat context that accumulates turns and optionally compacts on each add.

By default the context performs no compaction — the full history is retained. Compaction is opt-in: pass compactor= for a custom strategy, or window_size= as sugar for WindowCompactor(size=...).

Independently of compaction, a token budget can cap what view_for_generation forwards to the model: pass token_context_length_limit= for an explicit cap, or model_id= to derive the cap from the model's known context length. These operate at view time (newest-first, dropping oldest) and compose on top of any compaction already applied at add time.

Args:

  • compactor: The compactor invoked on every add. None (the default) means no compaction; full history is kept.
  • window_size: Sugar that constructs a WindowCompactor(size=window_size), whose default pin_predicate is pin_system. Mutually exclusive with compactor. None (the default) means no windowing.

Behavior change (deliberate): before the compaction refactor, window_size=N was a raw last-N view (as_list(N)) with no pinning — a system message could age out and exactly N items were returned. It now pins the system message (which therefore survives) and the size limit counts only the non-pinned body, so the returned view can exceed N items. To recover the old drop-in semantics, pass compactor=WindowCompactor(size=N, pin_predicate=pin_nothing).

  • token_context_length_limit: Explicit token budget cap for view_for_generation. Overrides the model-derived limit when set. None (the default) means no token cap.
  • model_id: Optional model identifier used for automatic context-window sizing. When set and token_context_length_limit is not provided, view_for_generation looks up the model's known context length and uses it as the token budget.

Methods:

FUNC model_id ​

model_id(self) -> str | ModelIdentifier | None

The model identifier bound to this context, or None if unbound.

FUNC new_instance ​

new_instance(self) -> ChatContext

Return a new empty root ChatContext, preserving compactor, token budget, and model_id.

Use this instead of reset_to_new() when you need to preserve the configuration of an existing instance. reset_to_new() is a classmethod that returns a bare ChatContext() with no configuration.

Returns:

  • A fresh root context with the same compactor,
  • token_context_length_limit, and model_id as this instance, but
  • no history.

FUNC add ​

add(self, c: Component | CBlock | ModelOutputThunk) -> ChatContext

Append c and run the compactor; return the resulting context.

Args:

  • c: The component, content block, or model output to append.

Returns:

  • A new ChatContext carrying the same configuration.

FUNC view_for_generation ​

view_for_generation(self) -> list[Component | CBlock | ModelOutputThunk] | None

Return the components to forward to the model.

Compaction is applied at add time (Pattern 1), so the stored history is already post-compaction. A token budget, if configured, is applied here on top of that:

  1. Explicit token_context_length_limit — token cap, overrides model table.
  2. Model-derived context length looked up via model_id.
  3. No token limit — return the full (post-compaction) history.

None is returned when the underlying history is non-linear.

Returns:

  • list[Component | CBlock | ModelOutputThunk] | None: Ordered list of
  • context entries, or None if the history is non-linear.