Context Compaction Lab [Core]

Experiment: experiments/exp_14_context_compaction/main.py

Objective

Contrast microcompact (truncate old tool results), autocompact (LLM summary of the middle), force compact (user /compact), and session memory extraction—aligned with src/services/context/compact.ts and per-turn trimming ideas in the query path.

Source mapping (Claude Code)

Piece	TypeScript
Compaction strategies	src/services/context/compact.ts
Ongoing context trimming (related)	src/query.ts

Architecture

Key code walkthrough

Microcompact preserves structure, shrinks stale tool payloads:

def microcompact(
    messages: list[dict[str, Any]],
    config: CompactConfig,
    current_turn: int,
) -> list[dict[str, Any]]:
    """
    Replace old tool_result content with truncated summaries.
    This reduces token count without losing structural information.
    Mirrors the per-iteration microcompact in query.ts.
    """
    result = []
    for i, msg in enumerate(messages):
        if (
            msg.get("role") == "tool_result"
            and i < len(messages) - config.preserve_recent_messages
        ):
            content = msg.get("content", "")
            if isinstance(content, str) and len(content) > config.max_tool_result_chars:
                truncated = content[:config.max_tool_result_chars]
                new_msg = {
                    **msg,
                    "content": f"[microcompacted] {truncated}...",
                    "_original_size": len(content),
                }
                result.append(new_msg)
                continue
        result.append(msg)
    return result

Autocompact calls the LLM for a middle summary:

async def autocompact(
    messages: list[dict[str, Any]],
    config: CompactConfig,
    client: UnifiedLLMClient,
) -> tuple[list[dict[str, Any]], str]:
    """
    When total tokens exceed threshold, summarize middle messages.
    Preserves: system prompt context + first user msg + last N messages.
    Mirrors compactConversation() in compact.ts.
    """
    total = count_messages_tokens(messages)
    if total <= config.autocompact_threshold:
        return messages, ""
    # ... build middle_text, await client.chat for summary, splice messages ...

Force compact in the demo sets autocompact_threshold=0 to always summarize (see main() section “Force Compact”).

How to run

cd experiments
python -m exp_14_context_compaction.main --mock
python -m exp_14_context_compaction.main --provider anthropic
python -m exp_14_context_compaction.main --provider openai

Exercises

1. Track compact_boundary in CompactState and log what the model can still “see.”
2. Replace heuristic extract_session_memory with a small structured LLM call.
3. Integrate microcompact into exp_03 agent_loop after each tool round.

Next experiment

Command System Lab (Comprehensive) wires /compact and other slash commands into the REPL.