What problem does it solve?
Context compaction in long-running coding sessions strips the reasoning behind past decisions, causing agents to re-litigate settled design choices and waste token budget. This Skill preserves the 'why' by maintaining a machine-readable decision log alongside the human-readable progress file.
Core Features & Use Cases
- Dual-file memory model: Pairs free-form
claude-progress.txt (narrative context) with claude-decisions.json (structured {ts, decision} entries) so both intent and hard choices survive compaction.
- Automatic decision capture: A
pre-compact hook scans transcripts for decision phrases (decided, chose, rejected, switched from X to Y), dedupes, timestamps, and appends up to 20 entries per compaction without ever blocking the compaction itself.
- Append-only contract: Entries are never edited or deleted; contradictions are resolved by newer entries with the reason noted in the progress file.
- Use Case: At the start of a new session after
/clear, read claude-decisions.json before touching code to confirm a prior session already rejected a polling approach, avoiding re-deriving that decision.
Quick Start
Read claude-decisions.json alongside claude-progress.txt at session start and check it before proposing any design change.