v1.10.0
v1.10.0 — net savings, safer compress, honest numbers
Changed 2
- /caveman-stats now reports net savings by subtracting skill costs (~1–1.5k input tokens per turn) instead of only counting output saved
- Skill hardening improved for negation safety, tool-call silence, and language-drift
Fixed 3
- Compress no longer loses data on Windows through explicit UTF-8 and atomic writes to prevent file truncation on crash
- Installer moved bin/ to cli/ with hardened Windows quoting and clean uninstall path
- Hooks now share one parser (caveman-parse.js), unwrap the payload envelope, and survive resume/clear/compact re-fires so mid-session level switches are no longer reset by SessionStart re-fire
Install
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex
Features
/caveman-statsreports net savings, not just gross (#145, #677). The skill costs ~1–1.5k input tokens per turn. Stats now subtracts that instead of only counting output saved, so a net-negative workload shows up as net-negative.
Fixes
- Compress no longer loses data on Windows (#652, #655, #686). Explicit UTF-8 plus atomic writes. A crash mid-write can no longer truncate the file it was rewriting.
- Installer moved
bin/→cli/, with hardened Windows quoting and a clean uninstall path. - Hooks share one parser (
caveman-parse.js), unwrap the payload envelope, and surviveresume/clear/compactre-fires — a mid-session/caveman <level>switch is no longer reset by a SessionStart re-fire. - Skill hardening: negation safety, tool-call silence, language-drift.
Docs
- Token savings are now split by workload. 65% is the chat-prose number. On agentic coding runs, JetBrains independently measured 8.5% across 86 SkillsBench tasks, with no detectable quality change. Both numbers are in the README with the mechanism behind the gap.
CI
- Full test suite (Node 18/20/22 + Python 3.11) runs on every push and PR.
Upgrade note
PINNED_REF moves to v1.10.0. This release adds a new hook module, and the detached installer verifies hooks against the pinned tag — so curl | bash installs resolve against v1.10.0 from here on. Existing installs should re-run the install command to pick up the new hook set.
Full changelog: https://github.com/JuliusBrussee/caveman/compare/v1.9.1...v1.10.0