squeez 0.3.1 → 0.4.0

package/CHANGELOG.md

@@ -0,0 +1,236 @@
+# Changelog
+
+All notable changes to squeez are documented here.
+
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+This project uses [Semantic Versioning](https://semver.org/).
+
+## [Unreleased]
+
+## [0.4.0] - 2026-04-16
+
+### Added
+- feat(opus47): dual-model cost table in benchmark — Sonnet 4.6 ($3/MTok) + Opus 4.7 ($5/MTok)
+- feat(opus47): protocol now educates LLM on xhigh thinking-block overhead (~2× budget burn)
+
+### Changed
+- feat(opus47): compact_threshold_tokens 120k → 90k (earlier compaction trigger for new tokenizer)
+- feat(opus47): agent_spawn_cost 270k → 350k (xhigh subagents burn more tokens)
+- feat(opus47): burn_rate_warn_calls 20 → 30 (wider pre-exhaustion warning window)
+- feat(opus47): state_warn_calls 5 → 10 (more runway to save session state)
+- feat(opus47): ultra_trigger_pct 0.72 → 0.65 (go Ultra compression earlier)
+- feat(opus47): MCP context_pressure thresholds — use_state_first ≥90%→≥75%, compact_soon ≥70%→≥55%
+- fix: DEFAULT_AGENT_SPAWN_COST constant synced to config default (200k → 270k)
+
+## [0.3.2] - 2026-04-16
+
+### Added
+- feat(setup): auto-detect lang from CLAUDE.md/locale + stderr update notification (#29)
+- v0.3.0 — token economy, pt-BR caveman, aggressive defaults, full hook install
+- structured session memory, cross-session search, error snippets, configurable tunables
+- feat(setup): create config.ini on first install
+- feat(i18n): locale-aware compress-md with pt-BR support (#25)
+- fuzzy dedup, adaptive intensity, benign summarize, MCP server, temporal memory, protocol payload
+- Windows setup via cargo/npm + squeez update fix
+- feat(benchmark): add squeez benchmark command with 19-scenario test suite
+- feat(ci): re-add promote-develop workflow with correct registration
+- feat(init): inject persona into ~/.claude/CLAUDE.md on every session start
+- feat(npm+crates): add npm package and publish jobs to release workflow
+- feat(statusline): add statusline.sh to repo; install via install.sh; show all-time stats
+- compress-md, caveman persona, squeez update, track-result hook
+- feat(context): adaptive intensity always picks Ultra
+- feat(context): aggressive token-engine — adaptive intensity, redundancy cache, summarize fallback
+- add Windows support to install.sh and release.yml
+- add home_dir() helper with USERPROFILE fallback for Windows
+- cross-platform wrap — shell_command() helper, cfg(unix) signal gates, libc unix-only
+- add GitHub Copilot CLI support
+- add OpenCode plugin for automatic Bash compression
+- add universal installer support for macOS and Linux
+- expand benchmarks, edge-case tests, and README polish
+- add SessionStart + PostToolUse hooks, update install.sh and README
+- wrap.rs — artifact capture (files/errors/git/tests) + compact warning header
+- squeez init — session start, prior memory banner, session finalization (TDD)
+- squeez track — token accumulation + session event log (TDD)
+- memory.rs — Summary, read_last_n, write_summary, prune_old (TDD)
+- session.rs — CurrentSession, append_event, date math (TDD)
+- json_util helpers + compact_threshold_tokens, memory_retention_days config
+- benchmark system — all fixtures passing
+- build.sh compiles + installs squeez, registers hook
+- PreToolUse hook script
+- fs and runtime handlers — all handler tests passing
+- cloud, database, network handlers (TDD)
+- typescript and build handlers (TDD)
+- package_mgr and test_runner handlers (TDD)
+- git and docker handlers (TDD)
+- wrap subcommand — sh -c, 120s timeout, signal forwarding, bypass check
+- Handler trait with cmd+config, generic handler, command router, filter_stdin
+- truncation strategy (TDD) — all strategy tests passing
+- grouping strategy (TDD)
+- dedup strategy (TDD)
+- smart_filter strategy with node_modules frame removal (TDD)
+- implement config.ini parser (TDD)
+
+### Changed
+- chore: bump version to 0.3.2
+- docs: add shields.io badge for changelog link
+- chore: bump version to 0.3.1
+- chore: bump version to 0.2.3
+- docs: update README with 11 MCP tools, structured memory, and advanced tunables
+- chore: bump version to 0.2.2
+- docs: rewrite README with full install instructions (curl/npm/cargo) and all commands
+- chore: refresh bench/report.md from latest run [skip release]
+- chore(ci): remove promote-develop workflow (re-adding to fix broken registration)
+- chore: bump version to 0.2.1 + auto-release on merge to main [skip release]
+- chore: bump version to 0.2.0; add Cargo.toml metadata
+- ci: PR template, linked-issue enforcement, no-claude-trailer check
+- docs: update README with PR1+PR2 features, fresh benchmarks, changelog
+- ci: auto-delete branches after PR merge
+- docs: update Windows installation instructions and prerequisites in README
+- docs: add Windows support and Rust install instructions to README
+- refactor: use session::home_dir() in config and init (Windows compat)
+- ci: trigger promote workflow (test)
+- ci(workflow): avoid failing promote job when PR creation is not permitted; add PR_CREATION_TOKEN guidance
+- docs(readme): add changelog, contribution workflow, and benchmark notes
+- ci(workflow): create/update PR from develop to main on push
+- chore: add .worktrees/ to gitignore
+- chore: add docs/ to gitignore (local design specs)
+- docs: update README for GitHub Copilot CLI support
+- security: fix remaining LOW findings in wrap.rs
+- security: fix 2 follow-up findings from second audit
+- security: fix all 8 findings from security audit
+- ci: build and install binary before running benchmarks
+- chore: add GitHub sponsor button
+- chore: ignore .omc/ directory, remove from tracking
+- chore: update benchmark report (7/7 passing)
+- docs: update README — real benchmarks, Phase 2 config, how it works
+- style: cargo fmt on session.rs and test_session.rs
+- chore: open source files — README, LICENSE, CI/release workflows, install.sh
+- chore: scaffold squeez Rust project
+
+### Fixed
+- fix(ci): run update-docs even if publish jobs fail (already published)
+- fix(ci): continue-on-error for publish jobs so update-docs always runs
+- fix(ci): add pull-requests write permission to update-docs job
+- fix(update): cargo-managed self-update + Windows deferred install + session update check (#27)
+- fix(ci): update-docs creates PR instead of pushing directly to protected main
+- fix(npm): copy README.md into npm package before publish; add readme field
+- fix(ci): add toolchain: stable to all dtolnay/rust-toolchain steps
+- fix(ci): handle no previous tags in generate-changelog.sh (pipefail + grep -v)
+- fix(ci): dispatch release.yml from auto-release to bypass GITHUB_TOKEN chain limit [skip release]
+- fix(ci): fix flaky test tmp dir collision + simplify auto-release
+- fix(bench): exclude synthetic fixtures from filter-mode bench/run.sh
+- fix(ci): remove secrets in step if-conditions — check token inside script instead
+- fix(ci): rename promote-develop → promote-to-main to fix broken workflow registration
+- fix(ci): bump github-script to v7 in promote-develop to force workflow re-registration
+- fix(compress_md): use byte slice comparison to avoid UTF-8 char boundary panic
+- fix(statusline): remove Ctx% — show only squeez compression metrics
+- fix(statusline): green ✓, proper capitalization, real-time context %, all metrics
+- resolve gap analysis — new handlers, wider redundancy window, token budget
+- install.sh — clarify PostToolUse no-matcher intent, guard version echo
+- wrap.rs — safe git hex check, O(n) dedup, full cmd in event log
+- use ends_with for files_committed check in init finalize
+- json_util — guard extract_u64 empty digit, add json_util test coverage
+- add pre-flight binary check to bench/run.sh
+- atomic settings.json write + uninstall safety guards
+- harden pretooluse hook (shlex.quote, error handling, single path source)
+- wrap — process_group(0) for correct signal forwarding, tighten is_streaming
+
+### Fixed
+- `squeez update` on Windows now detects cargo-managed installs and delegates to `cargo install squeez` automatically, avoiding exe-lock issues
+- `install_target_path()` uses `current_exe()` instead of hardcoded `~/.claude/squeez/bin`, so the correct binary is always updated regardless of install method
+- Windows self-update fallback spawns a detached `cmd.exe` to move the staged binary after process exit, instead of requiring manual `move /Y`
+- `install_atomic` now returns `Ok(bool)` distinguishing immediate vs deferred installs; success message no longer appears when install was only staged
+
+### Added
+- `session-start.sh`: rate-limited daily update check — prints `[squeez] Update available` to stderr (visible in terminal) when a new version is detected
+- `squeez setup`: auto-detects user language from `~/.claude/CLAUDE.md` and system locale (`LANG`/`LC_ALL`) and writes correct `lang =` into `config.ini` on first install
+
+## [0.3.0] - 2026-04-14
+
+### Added
+- Token economy layer: sub-agent cost tracking via PostToolUse hook (~200K tokens/spawn estimate)
+- Context pressure prediction: sliding-window burn rate with `[budget: ~N calls left]` header
+- Session efficiency scoring: compression ratio, tool choice, context reuse, budget conservation (basis points)
+- Tool result size budgets: per-tool output caps via PreToolUse hook (`read_max_lines`, `grep_max_results`)
+- Auto-calibration on install: `squeez calibrate` runs benchmarks and generates optimized `config.ini`
+- 2 new MCP tools: `squeez_agent_costs`, `squeez_session_efficiency` (11 → 13 total)
+- 2 new CLI commands: `squeez calibrate`, `squeez budget-params <tool>`
+- 5 new config keys: `agent_warn_threshold_pct`, `burn_rate_warn_calls`, `agent_spawn_cost`, `read_max_lines`, `grep_max_results`
+- 29 new economy module unit tests
+- `src/economy/` module: `agent_tracker.rs`, `burn_rate.rs`, `efficiency.rs`, `calibrate.rs`, `budget.rs`
+
+### Changed
+- PreToolUse hook now matches all tools (was Bash-only) for budget enforcement
+- `build.sh` runs `squeez calibrate` automatically after install
+- Wrap header now shows burn rate and agent cost warnings
+- Session finalization computes and stores efficiency scores in `memory/summaries.jsonl`
+- `SessionContext` extended with agent tracking and burn window fields
+
+## [0.2.3] - 2026-04-10
+
+### Added
+- Structured session memory: `investigated`, `learned`, `completed`, `next_steps` fields in summaries
+- Cross-session search: `squeez_search_history`, `squeez_file_history` MCP tools
+- Error snippets: first 128 chars stored alongside FNV fingerprints
+- Configurable tunables: `max_call_log`, `recent_window`, `similarity_threshold`, `ultra_trigger_pct`
+- `squeez_seen_error_details`, `squeez_session_detail`, `squeez_session_stats` MCP tools (6 → 11 total)
+- Locale-aware `compress-md` with pt-BR support (`lang = pt` in config or `--lang pt`)
+- `config.ini` auto-created on first install via `squeez setup`
+
+### Changed
+- MCP tool count: 6 → 11
+- README rewritten with full install instructions and all commands
+
+### Fixed
+- Windows setup via cargo/npm + `squeez update` fix
+- CI: dispatch `release.yml` from `auto-release` to bypass GITHUB_TOKEN chain limit
+
+## [0.2.2] - 2026-03-28
+
+### Added
+- `squeez benchmark` command: 19-scenario reproducible benchmark suite with quality scoring
+- JSON output mode (`--json`) and scenario filtering (`--scenario`)
+
+### Changed
+- README rewritten with full install instructions (curl/npm/cargo)
+
+### Fixed
+- Flaky test tmp dir collision in CI
+- `bench/run.sh` excludes synthetic fixtures (filter-mode only)
+- CI workflow registration fixes for `promote-develop`
+
+## [0.2.1] - 2026-03-20
+
+### Added
+- Persona injection into `~/.claude/CLAUDE.md` on every session start
+- npm package and crates.io publish jobs in release workflow
+- Statusline with real-time compression metrics
+
+### Fixed
+- `compress_md`: byte slice comparison to avoid UTF-8 char boundary panic
+- Statusline: proper capitalization, green checkmark, real-time context %
+
+## [0.2.0] - 2026-03-15
+
+### Added
+- Context engine: adaptive intensity, redundancy cache, summarize fallback
+- `compress-md` command: pure-Rust markdown compressor
+- Caveman persona prompt injection
+- `squeez update` self-update command
+- `track-result` PostToolUse hook for file/error tracking
+- Cloud CLI handlers (kubectl, gh, aws, gcloud, az)
+- Database handlers (psql, prisma, mysql)
+- Data tool handlers (jq, yq, terraform)
+- Windows support in `install.sh` and `release.yml`
+
+### Changed
+- Wider redundancy window, token budget improvements
+- CI: PR template, linked-issue enforcement
+
+[Unreleased]: https://github.com/claudioemmanuel/squeez/compare/v0.3.2...HEAD
+[0.3.0]: https://github.com/claudioemmanuel/squeez/compare/v0.2.3...v0.3.0
+[0.2.3]: https://github.com/claudioemmanuel/squeez/compare/v0.2.2...v0.2.3
+[0.2.2]: https://github.com/claudioemmanuel/squeez/compare/v0.2.1...v0.2.2
+[0.2.1]: https://github.com/claudioemmanuel/squeez/compare/v0.2.0...v0.2.1
+[0.2.0]: https://github.com/claudioemmanuel/squeez/releases/tag/v0.2.0
+[0.3.2]: https://github.com/claudioemmanuel/squeez/releases/tag/v0.3.2

package/README.md

@@ -4,7 +4,7 @@
 [![npm](https://img.shields.io/npm/v/squeez.svg)](https://www.npmjs.com/package/squeez)
 [![Crates.io](https://img.shields.io/crates/v/squeez.svg)](https://crates.io/crates/squeez)
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
-[Changelog](CHANGELOG.md)
+[![Changelog](https://img.shields.io/badge/changelog-📋-blue.svg)](CHANGELOG.md)
 
 End-to-end token optimizer for Claude Code, OpenCode, and GitHub Copilot CLI. Compresses bash output up to **95%**, collapses redundant calls, and injects a terse prompt persona — automatically, with zero new runtime dependencies.
 
@@ -98,25 +98,25 @@ Measured on macOS (Apple Silicon). Token count = `chars / 4` (matches Claude's ~
 
 | Scenario | Before | After | Reduction | Latency |
 |----------|--------|-------|-----------|---------|
-| `summarize_huge` | 82,257 tk | 420 tk | **-99%** | 62.9 ms |
-| `repetitive_output` | 4,692 tk | 37 tk | **-99%** | 102 µs |
-| `ps_aux` | 40,373 tk | 2,352 tk | **-94%** | 702 µs |
-| `git_log_200` | 2,692 tk | 289 tk | **-89%** | 75 µs |
-| `tsc_errors` | 731 tk | 101 tk | **-86%** | 20 µs |
-| `cargo_build_noisy` | 2,106 tk | 452 tk | **-79%** | 75 µs |
-| `docker_logs` | 665 tk | 186 tk | **-72%** | 16 µs |
-| `find_deep` | 424 tk | 134 tk | **-68%** | 29 µs |
-| `git_status` | 50 tk | 16 tk | **-68%** | 5 µs |
-| `verbose_app_log` | 4,957 tk | 1,991 tk | **-60%** | 93 µs |
-| `npm_install` | 524 tk | 232 tk | **-56%** | 16 µs |
-| `crosscall_redundancy_3x` | 486 tk | 241 tk | **-50%** | 57.7 ms |
-| `ls_la` | 1,782 tk | 886 tk | **-50%** | 52 µs |
-| `env_dump` | 441 tk | 287 tk | **-35%** | 9 µs |
-| `git_copilot` | 640 tk | 421 tk | **-34%** | 28 µs |
-| `md_prose` | 187 tk | 138 tk | **-26%** | 251 µs |
-| `md_claude_md` | 316 tk | 247 tk | **-22%** | 299 µs |
-| `git_diff` | 502 tk | 497 tk | **-1%** | 17 µs |
-| `kubectl_pods` | 1,513 tk | 1,513 tk | **-0%** | 20 µs |
+| `summarize_huge` | 82,257 tk | 420 tk | **-99%** | 55.9 ms |
+| `repetitive_output` | 4,692 tk | 37 tk | **-99%** | 205 µs |
+| `ps_aux` | 40,373 tk | 2,352 tk | **-94%** | 2.3 ms |
+| `git_log_200` | 2,692 tk | 289 tk | **-89%** | 153 µs |
+| `tsc_errors` | 731 tk | 101 tk | **-86%** | 25 µs |
+| `cargo_build_noisy` | 2,106 tk | 452 tk | **-79%** | 218 µs |
+| `docker_logs` | 665 tk | 186 tk | **-72%** | 34 µs |
+| `find_deep` | 424 tk | 134 tk | **-68%** | 71 µs |
+| `git_status` | 50 tk | 16 tk | **-68%** | 13 µs |
+| `verbose_app_log` | 4,957 tk | 1,991 tk | **-60%** | 232 µs |
+| `npm_install` | 524 tk | 232 tk | **-56%** | 41 µs |
+| `crosscall_redundancy_3x` | 486 tk | 241 tk | **-50%** | 51.5 ms |
+| `ls_la` | 1,782 tk | 886 tk | **-50%** | 155 µs |
+| `env_dump` | 441 tk | 287 tk | **-35%** | 19 µs |
+| `git_copilot` | 640 tk | 421 tk | **-34%** | 99 µs |
+| `md_prose` | 187 tk | 138 tk | **-26%** | 679 µs |
+| `md_claude_md` | 316 tk | 247 tk | **-22%** | 932 µs |
+| `git_diff` | 502 tk | 497 tk | **-1%** | 36 µs |
+| `kubectl_pods` | 1,513 tk | 1,513 tk | **-0%** | 28 µs |
 
 ### Aggregate
 
@@ -127,8 +127,8 @@ Measured on macOS (Apple Silicon). Token count = `chars / 4` (matches Claude's ~
 | Markdown / context files | **-23.5%** |
 | Wrap / cross-call engine | **-99.2%** |
 | Quality (signal terms preserved) | **19 / 19 pass** |
-| Latency p50 (filter mode) | **6.4 ms** |
-| Latency p95 (incl. wrap/summarize) | **63 ms** |
+| Latency p50 (filter mode) | **5.9 ms** |
+| Latency p95 (incl. wrap/summarize) | **56 ms** |
 
 ### Estimated cost savings — Claude Sonnet 4.6 · $3.00 / MTok input
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "squeez",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "Hook-based token compressor for Claude Code, Copilot CLI, and OpenCode. Compresses bash output up to 95%, collapses redundant calls, injects caveman persona.",
   "keywords": [
     "claude-code",