Releases

What changed, when. Latest first.

v0.16.2latest

  • README output-savings section now matches what lakonai gain actually shows: ~52% leaner prose on a rule-free Claude baseline (best case on chatty turns), with real workloads landing between the deterministic input savings (~46%) and that ceiling.
  • Framed honestly — never inflated into a headline.
npm install -g lakonai@0.16.2

v0.16.1

  • Fixed the output benchmark in lakonai gain — it now actually measures instead of getting stuck on "not measured yet".
  • The rule is isolated via the CLI's own --setting-sources project flag, keeping macOS Keychain auth intact (the old empty-config-dir trick logged Claude out).
npm install -g lakonai@0.16.1

v0.16.0

  • lakonai gain now shows **output** savings too, not just input — measures how much terser the model writes via your local AI CLI (no API key, weekly, TTY-only).
  • compress-memory takes a freeform instruction + validation levels: default (lossless), --prune (drop prose, protect code), --rewrite (free restructure).
  • lakonai upgrade updates via your package manager and refreshes the rule block. oh-my-zsh-style Update now? [Y/n] prompt from gain / version.
  • Smarter shim offer — only when a hook-less agent is detected. README rewritten lean (442 → ~170 lines).
npm install -g lakonai@0.16.0

v0.15.1

  • Promoted the slogan **"Speak less. Ship more."** to a hero heading.
npm install -g lakonai@0.15.1

v0.15.0

  • Platform-agnostic auto-learning — lakonai now learns heavy commands from ~/.lakon/log.jsonl (appended by every lakonai call on **any** agent), not just the Claude transcript. Throttled hourly; disable with LAKON_NO_LEARN=1.
  • With the shim, all four input-side features now run automatically on every agent for shell-mediated work.
npm install -g lakonai@0.15.0

v0.14.0

  • Universal Read-guard on the shell path: lakonai cat/head/tail/less/bat refuse junk reads (lockfiles, node_modules/, build artifacts) using the same deny rules as the Claude Read hook.
  • Junk-read protection is now automatic on **every** agent for shell-mediated reads.
npm install -g lakonai@0.14.0

v0.13.0

  • Universal PATH shim — lakonai shim makes shell-output filtering automatic on **every** agent (Codex, Cursor, Windsurf, Cline, Gemini), not just Claude Code. No hook API, no model cooperation.
  • Drops wrapper binaries for ls/grep/rg/find/cat/tree/head into ~/.lakon/shim/ and prepends it to PATH. Opt-in; lakonai shim --off reverts.
npm install -g lakonai@0.13.0

v0.12.0

  • Automatic MCP catalog compression — on Claude Code install, lakonai wraps each stdio MCP server so its tool/prompt/resource **descriptions** are compressed before they hit context. Offline, reversible, opt-out LAKON_NO_MCP=1.
  • Memory-file compression — lakonai compress-memory <file> shrinks authored memory (CLAUDE.md, notes) via a local AI CLI you already have. **No API key.** Backed up + validated byte-for-byte; sensitive filenames refused.
npm install -g lakonai@0.12.0

v0.11.0

  • lakonai shrink <mcp-server-cmd> — a stdio MCP proxy that compresses tool/prompt/resource description fields offline. Requests and results pass through untouched.
  • Bundled the MCP shrinker into lakonai as a subcommand — one install, one version.
npm install -g lakonai@0.11.0

v0.10.0

  • lakonai doctor — per-platform health check: CLI on PATH, rule installed, hooks registered.
  • Honest capability matrix in the README — exactly what each of the 6 agents gets.
  • Output benchmark harness measuring how much less the model writes under the rule via your own claude CLI.
npm install -g lakonai@0.10.0

v0.9.0

  • Per-turn reinforcement re-injects the terse rule each turn so the model doesn't drift back to verbose prose (opt out LAKON_NO_REINFORCE=1).
  • Terse workflow commands /lakonai:commit and /lakonai:review; benchmark harness fails CI if a filter regresses.
npm install -g lakonai@0.9.0

v0.8.0

  • Declarative filter engine — a 9-stage data-driven pipeline; adding a command filter is one entry in defs.js.
  • Test-runner filtering (jest, vitest, mocha, pytest, ava, go test, cargo test): collapse passes, keep failures + summary.
  • Lint / build / cloud filters: tsc, eslint, ruff, cargo clippy, make, docker, kubectl, aws, package installs. Coverage expanded from 4 to ~30 commands; suite at 270 tests, 100% coverage.
  • Removed the legacy lakon / lak aliases — the command is lakonai.
npm install -g lakonai@0.8.0

v0.7.0

  • Rebrand: package and brand renamed to **lakonai**. The lakonai command is the primary CLI.
  • Bash rewrite, slash commands, and project-rule paths all renamed to lakonai; install upgrades existing blocks cleanly.
npm install -g lakonai@0.7.0

v0.6.0

  • Update notifications via SessionStart hook (checks npm 1×/day, opt-out LAKON_NO_UPDATE_CHECK=1).
  • version command; visual revamp of gain; 100% test coverage.
npm install -g lakonai@0.6.0

v0.5.0

  • Stop hook records per-turn LLM token usage in gain — a separate "session output" block.
  • Multi-profile Claude Code support via CLAUDE_CONFIG_DIR.
npm install -g lakonai@0.5.0

v0.4.0

  • Per-project rules (Cursor / Windsurf / Cline) scoped behind --here; install no longer touches the current dir unless asked.
npm install -g lakonai@0.4.0

v0.3.0

  • Grep guard: auto-cap head_limit=30 with a one-shot hint.
  • Read denylist expanded (lockfiles, build artifacts, .next, .cache).
  • "Think in code" rule — the model writes one-off scripts instead of ingesting bulk data.
npm install -g lakonai@0.3.0