Skip to content

Repository navigation

CHANGELOG.md

Latest commit

 

History

History
212 lines (154 loc) · 10.4 KB

CHANGELOG.md

File metadata and controls

212 lines (154 loc) · 10.4 KB
Raw

Changelog

All notable changes to the smart-compact plugin documented here.

Format: Common Changelog + Keep a Changelog Versioning: Semantic Versioning

Unreleased

2.2.3.0 - 2026-03-01

Fixed

  • Bug fix: Fix division error in block_idle_session.py hook
    • Fixed inverted logic: changed if not bytes_saved == 0: to if bytes_saved == 0: to correctly exit when there are no savings
    • Added safety check for estimated_tokens being None before division operation
    • Prevents "unsupported operand type(s) for /: 'NoneType' and 'int'" error

2.2.2.0 - 2026-02-25

Changed

  • Refactor: Moved configuration and duplicate logic into const_models_and_config.py
    • block_idle_session.py and notify_about_compaction.py contained the same logic to get the duplicate bytes and estimated tokens
    • Both files also contained multiple methods to get environment variables

2.2.1.0 - 2026-02-23

Changed

  • Refactor: Reorganize monolithic cleanup_conversation.py into modular Python package
    • const_models_and_config.py - Data models, enums, configuration functions, and markers
    • content.py - Content type detection and line/character-level diffing
    • detection.py - Pattern detection for grep, bash, and edit operations
    • extract.py - Transcript I/O and operation extraction functions
    • __init__.py - Clean public API with __all__ exports
  • Main script reduced from 1299 to 368 lines (-72% complexity)
  • Improved maintainability: single responsibility per module
  • Enhanced code organization: easier to test, extend, and understand
  • Maintains full backward compatibility with existing functionality
  • All tests pass: edge cases, smart overlap detection, pattern matching

2.2.0.0 - 2026-02-22

Added

  • Detect and deduplicate identical bash script execution outputs (python, npm, node, dotnet, ruby, java, go)
  • Script outputs treated like file reads: mark earlier identical runs, keep latest
  • Script invocation pattern detection: auto-identify python script.py, npm test, dotnet build, etc.
  • Support for bash-wrapped scripts: bash -c "python script.py"

Changed

  • Breaking: Deduplication now keeps last (latest) occurrence instead of first occurrence
    • Aligns with LLM token reading preference (latest tokens prioritized)
    • More intuitive for forward reading (latest output at bottom)
  • Updated dedup markers to reflect "latest version below" structure
  • Add operation type tracking (op_type field) for context-aware marker messages
  • Separate marker text for bash script output vs file reads for clarity

2.1.1.0 - 2026-02-22

Added

  • Smart edit overlap detection for grep operations: only skip dedup if edits touch grep-matched lines (not conservative "any edit = skip")
  • Extended bash pattern detection: support head [-n N], tail [-n N] / -f, wc [-lcw] commands
  • Extract line numbers from grep output to enable safe overlap checking

Changed

  • Grep deduplication logic: from conservative "any edit = skip" to smart line-level overlap detection
  • Bash pattern detection: from simple cat/head/tail to comprehensive file-reading command coverage
  • Improved filepath extraction with robust flag parsing

2.1.0.0 - 2026-02-22

Added

  • Extend forward-chaining algorithm to deduplicate bash commands that read files (cat, head, tail)
  • Deduplicate grep search operations when later reads exist for same file with no intervening edits
  • Extract bash file paths and outputs for comparison using pattern matching
  • Extract grep patterns and matched line ranges for safe dedup checking
  • Track Edit operations for dedup safety validation between grep and reads
  • Unified operation stream: process all operations (Read, Write, Bash, Grep, Edit) chronologically per file

Changed

  • Bash operations treated like file reads: identical outputs → full dedup, different → partial dedup
  • Grep operations use conservative "later read without edits" check for safety
  • Both bash and grep inherit existing configurable thresholds (min bytes, context margins)
  • Algorithm maintains O(n) complexity with single pass per filepath
  • Full backward compatibility with existing Read/Write/Edit dedup behavior

2.0.0.0 - 2026-02-21

Complete rewrite with forward-chaining algorithm, dual-mode parsing, and configurable thresholds.

Changed

  • Breaking: Replace backward-iterating algorithm with forward-chaining for cleaner, more intuitive logic
  • Breaking: Last read in file always keeps full content (not first); better represents current state
  • Implement content type auto-detection (newline-based): single-line uses char-level diff, multiline uses line-level diff
  • Add configurable dedup thresholds and context margins via environment variables:
    • SMART_COMPACT_DEDUP_MIN_BYTES (default: 1)
    • SMART_COMPACT_MULTILINE_CONTEXT_LINES (default: 1)
    • SMART_COMPACT_SINGLELINE_CONTEXT_CHARS (default: 10)
  • Add raw_content extraction for Write → Read edge cases (immediate reads after writes)
  • Update output formats: short (Savings: XXX bytes (~YYY tokens)) and normal (Found N reads, XXX bytes)
  • Improve logging with --debug flag for troubleshooting
  • Track dedup state via markers (no external cache file needed); enables multi-pass convergence

1.4.1.0 - 2026-02-17

Fixed

  • Fix inflated token estimates (use standard ~4 bytes ≈ 1 token conversion)
  • Fix duplicate byte count accuracy (count actual message content, not original)
  • Fix cross-platform notification path resolution for cached plugin directories
  • Rename config variable SMART_COMPACT_CONTEXT_WINDOW_BYTES → SMART_COMPACT_CONTEXT_WINDOW_TOKENS

1.4.0.0 - 2026-02-17

Added

  • Add Stop hook for duplicate token notifications when idle and awaiting input
  • Add self-documenting deduplication markers with temporal language ("latest", "earlier")
  • Add configurable environment variables:
    • SMART_COMPACT_CACHE_DURATION_MINUTES
    • SMART_COMPACT_CONTEXT_WINDOW_BYTES (now CONTEXT_WINDOW_TOKENS)
    • SMART_COMPACT_NOTIFICATION_THRESHOLD_PERCENT
    • SMART_COMPACT_CACHE_VALIDATOR_THRESHOLD_PERCENT
  • Add cross-platform system notifications via notify_about_compaction.py

Changed

  • Format duplicate bytes as human-readable (KB, MB) for sizes ≥ 1000 bytes
  • Format tokens with K suffix for sizes ≥ 1000 tokens
  • Update marker format from DEDUPLICATION_*_MARKER|OMITTED_CHARS_COUNT: to natural language
  • Update hook registration from Notification event to Stop hook

1.3.0.0 - 2026-02-14

Backward-iterating chain-following algorithm for improved duplicate detection.

Changed

  • Replace two-phase approach with backward-iterating chain-following
  • Iterate from newest to oldest read for context-aware duplicate detection
  • Prioritize previous read over write for comparison (read-to-read before read-to-write)
  • Enable transitive deduplication in single pass (Read A = B = C all caught efficiently)
  • Apply partial dedup with line-level comparison when content differs

1.2.0.0 - 2026-02-14

Line-level partial deduplication for smarter duplicate detection.

Added

  • Phase 1 (primary): Line-by-line comparison for all reads (detects fully-identical and partially-identical)
  • Phase 2 (secondary): Hash-based read-after-write detection (only for remaining reads)
  • Partial dedup markers with per-block placeholders to maintain file structure
  • Smart filtering: skip files < 3 lines, skip partial omits < 3 lines
  • Improved token calculation based on file percentage omitted
  • Support for broader token ratio validation range (0.25-5 tokens/character)

Changed

  • Strategy: Two-phase approach (line-level primary, hash-based secondary)
  • Context margin: ±3 lines around changes (matches Claude Code edit tool)
  • Only create placeholders for omitted blocks ≥ 3 lines
  • Apply file-percentage token estimation for partial dedup

1.1.0.0 - 2026-02-14

Cache validation and token estimation features.

Added

  • Pre-prompt cache validator hook: detect stale transcripts (> 5 minutes idle) and block if dedup would save bytes
  • Token estimation from cache write ratios per file (0.5 < ratio < 5 tokens/character)
  • Add --dry-run-short flag for compact output format
  • Show token estimates in all modes: dry-run, CLI output, and hook messages
  • Format: Savings: XXX bytes (~YYY tokens)

1.0.0.0 - 2026-02-08

First release of transcript deduplication plugin.

Added

  • Hash-based duplicate detection (SHA256) for file content comparison
  • Two-level deduplication: priority to Writes, then keep latest Reads
  • Deterministic dedup markers: DEDUPLICATION_READ_AFTER_WRITE_MARKER and DEDUPLICATION_READ_MARKER
  • Integration modes: hook (automatic SessionEnd), CLI (manual with --dry-run)
  • Transcript format support: minified JSONL, pretty-printed JSON, unicode paths
  • Detailed dry-run reports and minimal hook mode output
  • O(n) single-pass analysis with no false positives