Distillation

Date: Sep 14, 2026

• 🟡 (23:22) Session compaction implementation defines DEFAULT_BUFFER = 20_000, DEFAULT_KEEP_TOKENS = 8_000, TOOL_OUTPUT_MAX_CHARS = 2_000, and SUMMARY_OUTPUT_TOKENS = 4_096.
• 🟡 (23:22) Session compaction uses SUMMARY_TEMPLATE, which requires an exact Markdown structure with sections in fixed order, including ## Objective, ## Important Details, ## Work State with ### Completed, and ## Next Move; instructions require terse bullets rather than prose paragraphs and moving completed work from Active to Completed.
• 🟡 (23:22) Session compaction serializes and truncates tool-result content as [Tool result]: ...; shell messages are represented as [Shell]: ${message.command}\n${truncate(message.output)}.
• 🟡 (23:22) Session compaction streams summary generation and appends text when LLMEvent.is.textDelta(event) is true; its context calculation includes context - Math.max(output, config.buffer).
• 🟡 (23:22) Working-tree diff for packages/core/schema.json keeps schema version: "7" and dialect: "sqlite" but changes schema snapshot ID from f14a9b18-8207-487e-a3d3-227e629ba9ad to 57591e35-6f8f-4cda-932c-6a08e85caae7.
• 🟡 (23:22) packages/core/schema.json diff expands the serialized SQLite schema metadata for session recovery/execution, including session_execution references, continuation_message_id, columns, foreign keys, primary keys, and indexes.