v0.1.2

Fixed

• User message filtering: Changed from message-level to block-level filtering — system-injected content (<system-reminder>, etc.) is now stripped per text block instead of discarding the entire user message, preventing legitimate instructions from being lost

Changed

• Simplified architecture: Merged compact-save.sh + compact_save.py into a single self-contained Python script
• Simplified tool call extraction: Replaced per-tool-type logic and priority system with unified parameter lookup
• Simplified restore: Removed legacy and generic fallback — session-specific snapshot only
• Large file optimization: Only reads last 2MB of transcripts to avoid timeout on long sessions
• Encoding safety: Added encoding='utf-8', errors='replace' to handle non-UTF-8 characters gracefully
• Better error messages: Errors now print to stderr instead of being silently swallowed
• Session isolation: snapshots saved per-session (compact-snapshot-<session_id>.md)
• Restore script parses session_id from stdin JSON to locate the correct snapshot
• Auto-cleanup: snapshot deleted after restore; stale snapshots (>10 min) removed during save
• User message numbering now chronological (1, 2, 3...) instead of reverse
• Tool call cap set to 20
• Key implementation details documented in CLAUDE.md and README.md

Removed

• compact-save.sh bash wrapper (merged into Python)
• Transcript search fallback (hook always provides transcript_path)
• Legacy v0.1.x snapshot fallback in restore script
• Priority-based tool call capping