Module compact 

History compaction.

Manages conversation history size by summarizing older messages when the context window limit approaches. Implements three compaction strategies:

• Auto-compact: triggered when estimated tokens exceed threshold
• Reactive compact: triggered by API prompt_too_long errors
• Microcompact: clears stale tool results to free tokens

Thresholds

|<--- context window (e.g., 200K) -------------------------------->|
|<--- effective window (context - 20K reserved) ------------------>|
|<--- auto-compact threshold (effective - 13K buffer) ------------>|
|                                                    ↑ compact fires here

Structs

CompactTracking

Tracking state for auto-compact across turns.

TokenWarningState

Token warning state for the UI.

Constants

MAX_OUTPUT_TOKENS_RECOVERY_LIMIT

Maximum recovery attempts for max-output-tokens errors.

Functions

auto_compact_threshold

Calculate the auto-compact threshold.

build_compact_summary_prompt

Build a compact summary request: asks the LLM to summarize the conversation up to a certain point.

compact_boundary_message

Create a compact boundary marker message.

compact_with_llm

Perform full LLM-based compaction of the conversation history.

effective_context_window

Calculate the effective context window (total minus output reservation).

max_output_recovery_message

Build the recovery message injected when max-output-tokens is hit.

microcompact

Perform microcompact: clear stale tool results to free tokens.

parse_prompt_too_long_gap

Parse a “prompt too long” error to extract the token gap.

should_auto_compact

Check whether auto-compact should fire for this conversation.

token_warning_state

Calculate token warning state for the current conversation.