whetstone-cli 3.0.0

Installer and CLI for Claude Code token optimization (Headroom + RTK + Memory)
# Whetstone

A single Rust binary that installs and orchestrates three upstream tools for
Claude Code (and friends): **Headroom** (context proxy), **RTK** (Bash-output
hook), and **ICM** (project memory). Idempotent setup, version-pinned manifest,
reversible v2 → v3 migration, release automation.

```
whetstone setup ──┬── Headroom (context compression proxy, upstream)
                  ├── RTK (Bash-output rewrite hook, upstream)
                  └── ICM (embedded SQLite project memory, upstream)
```

Compression is upstream's job. The glue — installing, versioning, configuring,
migrating, and tearing down cleanly — is whetstone's.

## Install

```bash
cd ~/my-project
curl -fsSL https://raw.githubusercontent.com/z19r/whetstone/main/install.sh | bash
```

Or from source: `cargo install whetstone && whetstone setup`

See [docs/install.md](docs/install.md) for prerequisites, setup details, and project configuration.

## Architecture

```
User → AI Coding Tool
         ├── Bash calls → [RTK Hook]            → compressed output
         ├── Context    → [Headroom Proxy :8787]→ LLM API
         └── Memory     → [ICM, embedded SQLite]→ persistent context
```

## What each piece does

**Whetstone (this repo)** — single Rust binary. Installs the three tools below,
ensures `uv` is present (offers to install it), version-pins everything in
`.claude/whetstone.json`, exposes `setup`, `update`, `migrate`, `doctor`, and
`release`. No runtime dependencies of its own.

**Headroom** *(upstream — [headroom-ai](https://pypi.org/project/headroom-ai/))* —
HTTP proxy in front of the LLM provider. Multi-stage pipeline: cache alignment,
content routing, statistical JSON compression, AST-aware code compression,
score-based message dropping. Compression numbers belong to Headroom; run
`curl localhost:8787/stats` to measure your own.

**RTK** *(upstream — [rtk-ai/rtk](https://github.com/rtk-ai/rtk))* — Bash-output
compression via PreToolUse hook. Compresses output before it enters the context
window. Caveat: it only fires on **Bash** tool calls; Claude Code's native
Read/Grep/Glob bypass it. Compression isn't free either — RTK's own tracker has
logged a ~18% net cost-increase case. Run `rtk gain` and `rtk discover` to keep
yourself honest, and consider RTK's audit mode if a particular rewrite feels
suspect.

**ICM** *(upstream — [rtk-ai/icm](https://github.com/rtk-ai/icm))* — embedded
SQLite memory store; skills, hooks, and CLI installed by
`icm init --mode standard`. Whetstone v2's bundled MemStack/skills/rules layer
is gone — ICM owns its own assets now.

## Documentation

| Doc | Contents |
|-----|----------|
| [Installation]docs/install.md | Prerequisites, quick start, new/existing project setup |
| [CLI Reference]docs/cli-reference.md | All v3 commands, flags, RTK quick reference + caveats |
| [Migration Guide]docs/migration.md | v2 → v3 with `migrate`, `--dry-run`, `--rollback` |
| [Editor Setup]docs/editors.md | Claude Code, Cursor, Copilot, Windsurf, Cline, Aider, Codex, Gemini CLI, OpenCode + compatibility matrix |
| [Headroom Service]docs/headroom-service.md | systemd, launchd, and background setup |
| [Configuration]docs/configuration.md | Global/per-project files, environment variables |
| [Troubleshooting]docs/troubleshooting.md | Common issues, uninstall, manual removal |

## Development

```bash
just build          # Debug build
just test           # Run all tests
just lint           # Clippy lints
just fmt            # Format code
just check          # Build + test + lint
```

Source layout in [CLAUDE.md](CLAUDE.md).