distill

distill is a Rust-first orchestration layer for Claude Code cost optimization.

It does not reimplement GrepAI, RTK, Rolling Context, Token Optimizer, or claude-pulse. Instead, it detects, validates, and composes them into a coherent stack with sane defaults, install guidance, and conflict checks.

What problem this solves

Claude Code cost and quota pressure usually come from several independent failure modes:

1. Codebase exploration waste — too many broad file reads and weak search.
2. Shell output waste — huge command output gets pushed into context.
3. Long-session decay — compaction destroys useful history.
4. Invisible context overhead — ghost tokens from settings, skills, MCP, memory, and plugin load.
5. Poor observability — users do not see session, weekly, or context pressure early enough.

distill turns that into a layered system:

• GrepAI for semantic code retrieval.
• RTK for shell-output compression.
• Rolling Context for long-session compression.
• Token Optimizer for context audits and compaction survival.
• claude-pulse for usage analytics and historical visibility.
• distill statusline as the built-in local fallback renderer.
• Claude Code plugin + project settings generated by this repo to make the stack easier to install and maintain.

Repository layout

• src/ — Rust CLI
• docs/ — product requirements, architecture, publishing plan
• .claude-plugin/marketplace.json — marketplace metadata for Claude Code
• plugins/distill-stack/ — bundled Claude Code plugin
• npm/ — Node wrapper for NPM distribution
• .github/workflows/ — release scaffolding

CLI

distill init
distill doctor
distill run -- -p "..."
distill statusline
distill print-plugin
distill print-marketplace

Design principles

• Do not vendor upstream tools by default. Install or detect them as external adapters.
• One rewriter per layer. RTK owns shell PreToolUse rewriting. Avoid stacking multiple rewriters.
• Prefer project scope when shareable, local scope when machine-specific.
• Fail soft. Missing external tools should not break Claude Code.
• Keep legal boundaries clean. Some upstream projects are permissive, some are copyleft, and some are source-available only.

Measured findings

Local benchmarking on 2026-04-03 showed that the cheapest tested profile for this repository was:

• distill lean mode
• RTK for Bash rewriting
• Rolling Context for long-session support

The full OpenWolf + RTK + Rolling Context stack was not the cheapest option in this repo-level benchmark.

See docs/STACK_BENCHMARKS.md for the benchmark method, totals, and setup notes for GrepAI, Token Optimizer, and claude-pulse.

Recommended product direction

The most robust public release path is:

1. Publish the Rust binary on crates.io.
2. Publish a thin NPM wrapper that exposes the same CLI using the bin field and lifecycle scripts.
3. Host the repo on GitHub with a Claude Code marketplace and plugin.
4. Let distill init configure project/user/local scope with the least manual work possible.

Suggested repo name

• GitHub: SamoraDC/distill
• Crates: distill
• NPM: @quantumgraph/distill

Current state of this scaffold

The repository now gives you:

• a working Rust CLI with doctor, init, run, and statusline,
• adapter registry coverage for GrepAI, RTK, Rolling Context, Token Optimizer, and claude-pulse,
• safe local/project Claude settings generation with merge + backup behavior,
• a tested NPM wrapper around the Rust binary,
• Claude Code marketplace/plugin files,
• and the publication strategy for crates, NPM, and GitHub.

It still does not auto-install every upstream dependency for you. The current model is an orchestrator with official install guidance and conflict-aware diagnostics, not a redistributor of third-party binaries.