Small, single-purpose, open-source tools that sit alongside AI coding agents (Claude Code, Cursor, Codex, Gemini CLI) and catch the specific, recurring failure modes they're prone to. None of them are AI tools themselves — no accounts, no backend, no billing. Each installs in one command and does exactly one thing.
| Tool | Category | What it does |
|---|---|---|
| checkpoint | Safety | Automatic git checkpoints before your AI agent runs a bash command — the gap Claude Code's native /rewind doesn't cover |
| loopkill | Awareness | Detects when your AI agent is stuck in a loop and interrupts it before it burns your quota |
| ctxmeter | Awareness | A live context-window gauge for AI coding CLIs that don't already have one (Claude Code users: use /statusline instead) |
| again | Awareness | Counts how many times you've had to repeat yourself to your AI agent this session |
| permit | Safety | Diagnoses why a Claude Code permission rule silently doesn't apply |
| actually | Trust | Cross-checks a "tests pass" claim against the real last test run in the same session |
| exists | Trust | Checks whether a package your agent just installed (npm/PyPI) actually exists in the real registry |
| spare | Safety | Makes rm/Remove-Item recoverable — the safety net underneath every AI agent's delete command, on macOS, Linux, and Windows |
Not staged output — the actual released binaries, run against a real transcript, making real calls where the tool needs to (exists really asks npm's registry; actually really parses the real failing test output).
actually catches a "tests pass" claim that contradicts the last real test run:
exists catches a hallucinated package name before it's trusted:
spare catches a real rm and makes it recoverable:
Every tool here has a "Known limitations" and "Verified against real data" section in its own README — the point of this section is to put the concrete evidence in one place instead of six, so it's not just a claim.
- checkpoint: hardened against 8 real scenarios beyond the happy path
(no commits yet, detached HEAD, 30 truly concurrent saves, corrupted log
lines, symlinks/binaries) — 2 real bugs found and fixed along the way.
Verified via a real
brew-installed binary:brew upgrade, three full uninstall/reinstall cycles, zero leftover state. - loopkill: verified against a real simulated stuck process — zero false positives on a spinner, correct detection and kill on a genuine loop, not a synthetic pass/fail assertion.
- ctxmeter: verified that piping through it never swallows the wrapped CLI's real output — the gauge renders on stderr, stdout reaches the terminal byte-for-byte unmodified.
- again: the real Claude Code transcript shape was confirmed correct against a live session, not assumed — and its token counts are the actual measured length of a repeated prompt, deliberately never called "tokens saved," because that would require guessing a counterfactual no tool can actually know.
- permit: cut from three commands down to one (
doctor) after discovering Claude Code already had native features doing the same job. Shipping less, on purpose, once redundant work was found, rather than keeping code around because it existed. - actually: swept against ~500 real assistant messages from this project's own development session — 0 false positives after fixing 3 real bugs found along the way, including a silent-data-loss bug where one oversized transcript line could drop detection for the rest of a session with zero error output.
- exists: its test suite makes live calls to the real npm and PyPI registries rather than mocking them, because the entire point is whether the real registry agrees. A pre-release audit found its own pitch had quietly oversold what it protects against (it can't catch a package name that's already been squatted, only one that's still unclaimed) — the README was rewritten to say so plainly rather than leave the stronger, less accurate claim standing.
- spare: built from four real, severe, publicly filed incidents
(a custom OS with 78+ verified build iterations destroyed, 160+ hours of
work lost, an entire
~/Desktopincluding installed applications gone) — none of whichcheckpointwould have caught, sincegit stashstructurally cannot capture untracked files. Its own CI actually executes the built binary on realwindows-latest,macos-latest, andubuntu-latestGitHub Actions runners on every push, catching a real Windows-specific bug (%USERPROFILE%vs$HOME) that cross- compiling alone — the only verification every other tool here has ever had for its Windows/Linux binaries — would never have surfaced.
Each one is a narrow fix for a single, specific, well-evidenced annoyance —
in the spirit of tools like fzf, ripgrep, and atuin, not a bundled
suite you have to buy into all at once. Use one, use all eight, doesn't
matter — none of them depend on each other.
Each tool has its own README with full details. The short version, once
releases exist (see each tool's docs/RELEASING.md):
go install github.com/Soldsoul86/AAA/checkpoint@latest
go install github.com/Soldsoul86/AAA/loopkill@latest
go install github.com/Soldsoul86/AAA/ctxmeter@latest
go install github.com/Soldsoul86/AAA/again@latest
go install github.com/Soldsoul86/AAA/permit@latest
go install github.com/Soldsoul86/AAA/actually@latest
go install github.com/Soldsoul86/AAA/exists@latest
go install github.com/Soldsoul86/AAA/spare@latest
Or build from source today, before any release exists:
git clone https://github.com/Soldsoul86/AAA
cd AAA/checkpoint && go build -o checkpoint .
This is a monorepo — one repo, independent Go modules, one folder each.
Every tool has its own go.mod, its own tests, its own
.goreleaser.yaml, and its own release workflow
(.github/workflows/release-<tool>.yml), triggered only by that tool's own
tag prefix (e.g. checkpoint-v0.1.0) so releasing one tool never touches
the others.
MIT, per tool (see each folder's LICENSE).