@mygsoft/context-diet

v1.3.2

Published

3 days ago

Smart AI Context Optimizer & Adaptive MCP Server with Global Language Support

0High
0Medium
0Low

mygsoft

mcp ai context-optimization tree-sitter token-saving

📉 Context-Diet v1.3.2

Predictive AI Context Optimizer, Terminal Log Distiller & Universal MCP Infrastructure (Global OSS Synergy Edition)

Context-Diet is a high-performance engine designed to slash AI token costs by 85% to 99% while boosting reasoning accuracy. By combining AST-based folding, multi-turn context working-set caching, session delta diffs, virtual type bundling, RTK terminal log distillation, Headroom CCR reversible compression, 2-Tier atomic decision memory, and Repomix zero-body skeleton packing, Context-Diet ensures your AI agents (Claude, Cursor, Antigravity, Windsurf, Cline) process only the code that truly matters.

📑 Table of Contents

🚀 Quick Start

Initialize your project and AI IDE environments in one command:

# Interactive setup (prompts: Global, Local workspace only, or Both)
npx @mygsoft/context-diet init

# Or install globally across all AI IDEs on this PC (without creating local files)
npx @mygsoft/context-diet init --global

# Or install locally for the current project only (without modifying global settings)
npx @mygsoft/context-diet init --local

The init command prompts you to choose your desired installation scope:

🌐 Both (Recommended): Automatically deploys persistent zero-waste rules and MCP to Antigravity (~/.gemini/config), Claude Code (~/.claude/CLAUDE.md), Cursor (~/.cursor/rules/), Windsurf (~/.codeium/windsurf/), and Cline (~/.clinerules), AND configures current project workspace rules (GEMINI.md, .cursorrules, etc.).
🖥️ Global Only (--global / -g): Sets up PC-wide global rules and MCP servers across all detected AI IDEs without creating files in your current directory.
📁 Local Only (--local / -l): Injects rules, lifecycle guards (hooks.json & enforce-diet.cjs), and skills only into the current project workspace, leaving your global IDE settings 100% untouched.

🔌 One-Click MCP Installation

Automatically registers Context-Diet into all detected IDEs:

# Auto-detect and register into Claude Desktop, Cursor, Antigravity, Windsurf, Cline
diet install-mcp

# List detected clients and config paths
diet install-mcp --list

# Simulate installation without writing files
diet install-mcp --dry-run

# Force install to all supported clients
diet install-mcp --all

Supported MCP Clients:

Claude Desktop: %APPDATA%\Claude\claude_desktop_config.json (Windows) / ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
Cursor: ~/.cursor/mcp.json (Global) & .cursor/mcp.json (Workspace)
Antigravity IDE: ~/.gemini/config/mcp_config.json & .agents/mcp_config.json
Windsurf: ~/.codeium/windsurf/mcp_config.json
Cline / Roo Code: VS Code global extension settings

📦 Multi-OS Packages & Installation

Context-Diet is distributed as ready-to-run native packages for Gumroad and npm:

For detailed step-by-step setup and troubleshooting, refer to INSTALL.md.

💎 Editions & Feature Matrix (Community vs Pro)

Context-Diet is released under the Functional Source License, Version 1.1 (FSL-1.1-MIT).

Source-Available: The entire source code is 100% transparent and public on GitHub.
Freedom to Build: Free to inspect, use locally, modify, and integrate into internal personal and team workflows.
Fair Competition & Anti-Crack: Creating competing commercial services or bypassing feature gating/paywalls is strictly prohibited. Automatically converts to standard MIT License 2 years from release date.

💡 ROI Payback Guarantee: Once Context-Diet saves you over $14 (Early-Bird) or $29 (Regular) in LLM API token costs, the Pro license has paid for itself completely! Pro Lifetime Edition can be activated via Polar with full offline cryptographic support (diet activate <key> --offline).
🏢 Team & Enterprise Edition (Roadmap): Centralized token telemetry, shared team cache, and organization-wide PR gates are planned for future release.

🛠 CLI Command Reference

1. `diet init [options]`

Initializes Context-Diet zero-waste protocols, lifecycle hooks, agent skills, and MCP servers across your workspace and global AI environments.

# Interactive setup (prompts: Global, Local only, or Both)
diet init

# Install globally across all AI IDEs on your PC (no local project files created)
diet init --global

# Install only for the current project workspace (global settings untouched)
diet init --local

# Non-interactive auto-setup for all supported formats
diet init -y --all

-g, --global: Install persistent zero-waste rules, lifecycle hooks, and MCP servers globally (~/.gemini/config, ~/.claude/CLAUDE.md, ~/.cursor/rules/, ~/.codeium/windsurf/, ~/.clinerules) without touching the local project folder.
-l, --local: Install rules, lifecycle guards, and MCP servers only to the current project workspace without modifying global settings.
-y, --yes: Non-interactive auto-setup without interactive prompts (defaults to both).
--all: Inject guidelines and register into all supported AI IDE formats simultaneously.

2. `diet read <path> [options]`

Reads a file with AST folding, predictive caching, session working-set tracking, and automated secret masking (SecretSanitizer).

diet read src/main.ts -s handleAuth

-s, --symbol <name>: Unfold a specific function, class, or method.
-i, --intent <keywords>: Unfold code blocks matching keywords.
-t, --types-only: Extract only pure type signatures and interfaces, stripping function bodies (95% token savings).
-b, --bundle-types: Inline imported local types at the bottom to eliminate round-trip reads.
-f, --force: Bypass working-set cache and force a complete re-dump.
--no-delta: Disable unified diff output on modified files.
--ref <id>: Expand stored raw log reference (Headroom CCR pattern).

2. `diet exec <command...>`

Runs terminal commands (tests, builds, lints, git, docker) and distills output using 12 RTK profiles, removing 95% of noise while preserving critical errors and stack traces. Stashes raw output to .context-diet/raw_logs/ for reversible restoration.

diet exec npm test
diet exec git status
diet exec docker build .

3. `diet expand <ref_id> [options]`

Reversibly restores original terminal or command raw logs on demand using Headroom CCR (Chunk-wise Continuous Retrieval) pattern.

diet expand log_1729001234_a1b2
diet expand log_1729001234_a1b2 -s 1 -e 50

-s, --start-line <number>: 1-indexed start line.
-e, --end-line <number>: 1-indexed end line.

4. `diet doctor [--fix] [-y, --yes]`

Comprehensive system and dependency health diagnosis.

diet doctor
diet doctor --fix

Checks Node.js version, package manager (npm/pnpm/yarn/bun), native SQLite acceleration, Tree-sitter WASM grammars, and MCP registrations.

5. `diet fix-deps [-y, --yes]`

Safely repairs/rebuilds missing WASM binaries and native dependencies with interactive TTY consent and zero root project pollution.

diet fix-deps

6. `diet stats [path]`

Displays telemetry on saved tokens, cost reductions ($), and cache hit rates.

diet stats

7. `diet audit <path>`

Audits compression ratio, parsing speed, and syntax integrity for any source file.

diet audit src/core/parser.ts

8. `diet search <query> [path]`

Fast hybrid code search powered by native Rust rg (ripgrep) / fd subprocesses with automatic pure Node.js fallback.

diet search "handleAuth" src/

9. `diet tree [path]`

Generates an optimized, token-friendly directory tree.

diet tree src/ -d 2

-d, --depth <number>: Maximum traversal depth (default: 2).

10. `diet query <pattern> [target_path]`

Performs fast AST structural pattern matching (ast-grep pattern) across files or directories without regex fragility.

diet query "function $NAME($$$ARGS) { $$$BODY }" src/

-l, --lang <language>: Language filter (typescript, javascript, python, go, rust, java, csharp, cpp).
-m, --max-results <number>: Maximum matching nodes to return (default: 20).

11. `diet pack [target_path]`

Packs project into a zero-body architecture skeleton for instant AI cold-start onboarding (Repomix & PageRank Repo-Map pattern).

# Output XML skeleton to stdout or file
diet pack src/ -f xml -o architecture-skeleton.xml

-f, --format <format>: Output format (xml or markdown, default: xml).
-o, --output <file>: Output file path (default: stdout).
-m, --max-tokens <number>: Upper limit on skeleton tokens (default: 20000).
--budget <number>: Target token budget for PageRank graph selection.

12. `diet memory [action] [args...]` (2-Tier Atomic Memory)

Manages long-term project architectural decisions, conventions, and constraints with a 2-Tier memory architecture (inspired by Mem0 & Letta).

# Record an architectural decision (atomic fact)
diet memory add "Use SQLite for Journal" "Avoid JSON files for write-heavy telemetry" -c architecture -t "db,perf"

# Search and recall decisions by keyword
diet memory recall "SQLite"

# Generate dense prompt decision capsule for AI prompt prefix (optional query filter)
diet memory prompt
diet memory prompt "SQLite"

# List all recorded decisions
diet memory list

-c, --category <category>: Category (architecture, convention, constraint, decision).
-t, --tags <tags>: Comma-separated tags.
Session Reset Protocol: On task/topic switch, record critical facts with diet memory add, reset the chat session (/clear or new chat), and restore only 1~3 targeted facts with diet memory recall. This limits turn accumulation and reduces how often /compact is needed (it complements /compact, it does not replace it).

13. `diet subagent [action] [sub_agent_id]` 🔒 (Pro)

Manages partitioned working-set contexts for concurrent sub-agents, preventing noise contamination and distilling ultra-compact handoff diff capsules.

# Initialize partitioned sub-agent session
diet subagent create worker-1 -p root-session

# Inspect active sub-agent sessions
diet subagent list

# Distill handoff XML diff summary (< 300 tokens)
diet subagent handoff worker-1

# Complete session and merge file changes to parent cache
diet subagent complete worker-1 --merge

-p, --parent-id <id>: Parent session or agent identifier.
--merge: Merge sub-agent file changes back to parent context working-set cache upon completion.

14. `diet server` (or `diet serve`, `diet mcp`)

Launches the Context-Diet MCP server directly in stdio mode.

15. `diet clean`

Removes injected AI guidelines and resets rule files.

16. `diet activate <key> [options]`

Activates Context-Diet Pro license with Gumroad online validation or offline signature.

# Online activation with automatic IDE MCP registration
diet activate YOUR-LICENSE-KEY --install-all

# Offline cryptographic activation
diet activate YOUR-LICENSE-KEY --offline

-e, --email <email>: Associate license with your purchase email.
--offline: Skip online server validation and use offline cryptographic check.
--install-all: Automatically register MCP server into all detected IDEs upon successful activation.

17. `diet license [options]`

Displays active license tier, machine instance ID, and status.

diet license
diet license --deactivate

--deactivate: Deactivates and unbinds active license from current machine and reverts to Community Edition.

18. `diet check [path] [options]` (alias: `guard`)

Audits project for AI context health (oversized files, high-volume barrel files, giant data files) to prevent context pollution in CI/CD.

# Audit current repository
diet check

# Enforce strict CI threshold
diet check --ci --min-score 85

# Output machine-readable JSON
diet check --json

--max-lines <number>: Maximum recommended lines per source file (default: 500).
--min-score <number>: Minimum passing health score (default: 80).
--ci: Exit with code 1 if health score fails threshold (for GitHub Actions / CI).
--json: Output results in structured JSON format.

19. `diet ui [options]`

Launches the embedded real-time Web Dashboard and token analytics server with instant browser launch.

# Launch dashboard on default port (3333) and open browser
diet ui

# Custom port or host without auto-opening browser
diet ui --port 8080 --host 0.0.0.0 --no-open

-p, --port <number>: Port to listen on (default: 3333, automatically falls back up to 3343 if occupied).
-h, --host <host>: Host address to bind (default: 127.0.0.1).
--no-open: Do not automatically open the browser on launch.

Dashboard Highlights:

Real-time KPI Metrics: Live cumulative tokens saved, dollar cost savings ($), folded lines, and compression efficiency rate.
Active AI Context Working-Set Cache: Inspect active deduplicated file sessions, hit ratios (X / 8), last referenced timestamps, and Myers delta diff statuses.
Telemetry & Operations Journal: Live event stream with instant text filter for file paths, commands, and queries.
Modern Custom Pagination: 25 rows per page default with custom dark glassmorphism popover dropdown (10, 25, 50, 100, All) and page navigation buttons.
Interactive SVG Time-Series Chart: Visualizes token compression trends with interactive hover tooltips.
In-App License Management: Activate and deactivate Pro Lifetime licenses with zero terminal hassle.

🧠 Core Capabilities (Multi-Tier Optimization)

Tier 1: AST Code Diet (Tree-sitter WASM)
- Semantic folding for TypeScript, JavaScript, Python, Go, Rust, Java, C#, and C++.
- Types-Only Mode (-t) & Virtual Type Bundler (-b) preserving multi-line types and signatures.
Tier 2: Predictive Computation Cache (SmartCacheEngine)
- High-throughput AST caching with LRU eviction and background module import warmup.
Tier 3: Universal AI Context Working-Set Cache with Disk Persistence
- Multi-turn deduplication replacing already-read files with 35-token pointers (99% savings).
- Session Delta Diff: sends compact unified diffs (@@ ... @@) using Myers diff algorithm.
- 3-Way Guardrails: 1,000 LRU entries, 1-hour inactivity TTL, and 8-hit safe re-sync.
Tier 4: RTK Terminal Log Distiller (diet_exec)
- Captures verbose terminal logs (tests, builds, lints, git, docker, cargo, dotnet) with 12 specialized profiles and repeated line (×N) folding (95% log reduction).
Tier 5: Architecture Skeleton Packer (diet pack / Repomix & PageRank Repo-Map)
- Zero-body architecture skeleton bundler for instant AI onboarding with up to 98% token reduction.
- PageRank dependency graph centrality ranking prioritizing entry points and core type hubs.
Tier 6: Data Leak Guard & Secret Masking (SecretSanitizer)
- Real-time Gitleaks pattern matching (20+ sensitive credential formats) auto-redacting secrets in file reads and execution outputs.
Tier 7: 2-Tier Atomic Decision Memory & Session Reset (diet memory & diet_memory)
- Atomic facts extraction & persistent storage (.context-diet/decisions.json) inspired by Mem0 & Letta.
- Targeted prompt capsules (50~150 tokens) eliminating full-history turn bloat.
- Topic-switching session reset protocol with working state (diet_memory action state) re-injected by a SessionStart hook after /compact or /clear, so compaction keeps the thread. It complements /compact; it does not replace it.
Tier 8: Native Hybrid Accelerated Search (HybridSearchEngine)
- Native Rust rg / fd subprocess acceleration (<10ms query times) with pure Node.js fallback.
Tier 9: Dynamic MCP Tool Gating (DynamicToolGatingEngine)
- Profile-based tool schema filtering (CONTEXT_DIET_TOOL_PROFILE=minimal|core|code|debug|full, default core) cutting 40%~60% fixed schema tokens (800~1,500 tokens/turn).
Tier 10: Bundle AST Skimmer & ASAR Reader (BundleAstSkimmer, AsarReader)
- Single-line minified bundle & Electron .asar AST skimmer extracting signatures and SourceMaps without exploding context (95%~99% savings).
Tier 11: Call Graph & IPC Tracer (CallGraphEngine, IpcTracer)
- High-precision caller/callee trees and Electron IPC / API routing without source body bloat (70%~85% savings).
Tier 12: Evidence v2 Protocol (EvidenceProtocolAdapter)
- Structured evidence metadata ([EVIDENCE], [LIMITATION], [RECOVERED_GRAPH]) eliminating hallucinated logic and reducing retry loops by 50%.

Additional Guardrails & Ergonomics:

Headroom CCR Reversible Storage (diet_expand): Distilled raw terminal logs are safely stashed in .context-diet/raw_logs/ and accessible via ref:<id>.
Negative Compression Guard (NegativeCompressionGuard): Guarantees zero token bloat by automatically falling back to lossless raw content if compression ratio is unfavorable.
Transparent Execution Rewriter (TransparentExecutionRewriter): Intercepts shell execution commands and transparently filters high-noise outputs.
Sub-Agent Context Isolation & Distillation Adapter (SubAgentAdapter): Partitions sub-agent working sets and distills handoff summaries to prevent main context contamination.
Real-Time Web Dashboard (diet ui): Zero-dependency embedded dashboard with SVG time-series analytics, interactive pagination, and in-app Pro licensing.

📊 Performance Benchmarks (Full Ultra-Lean vs Standard Agent)

Estimates, not measurements. Prompt caching bills cached input at ~10% of the base price, so cost savings are smaller than token savings. Measure on your own tasks with npm run bench (A/B harness in bench/; runs are billed).

🚀 Pro Features Impact (Virtual Type Bundler & Sub-Agent Isolation)

📦 MCP Tools for Agents

Note: Telemetry, diagnostic audits, and search are available directly via the CLI (diet stats, diet audit, diet search) and the real-time Web Dashboard (diet ui) to keep MCP prompt schemas lean.

🩺 System Diagnostics & Resilient Fallback

Graceful Fallback (FallbackJournal): If native better-sqlite3 fails to compile or load due to Node ABI mismatches, Context-Diet automatically falls back to pure JSON storage (.context-diet/journal.json). Operations never crash.
TTY Deadlock Guard: When invoked by background MCP daemons without an interactive terminal (!process.stdin.isTTY), prompts are automatically bypassed to prevent deadlocks.
Isolated Repair: Dependency repairs use the detected package manager (bun, pnpm, yarn, npm) without modifying user workspace dependencies.
ContextGuard & Lifecycle Interceptor: Hardware/hook-level prevention of raw terminal dumps and oversized (>100 lines) file reads.
Execution Environment Sanitization: Subprocess execution in diet_exec strips parent IDE environment variables (VSCODE_*) to ensure clean CLI runner execution.
Lossless Rollback & Negative Compression Guard: Prevents context inflation by falling back to raw output when compression ratio is negative.
Real-time Secret Masking: Prevents confidential tokens and credentials from leaking into agent prompt contexts.

🏗 Development & Testing

# Install dependencies & fetch WASM grammars
npm run setup

# Build TypeScript
npm run build

# Run Vitest test suite (30 test files, 268 tests, 100% pass)
npm test

# Launch Web Dashboard locally
npm run ui

# Bundle & Obfuscate (Phase 1)
npm run build:bundle

# Package standalone native binaries (Phase 2)
npm run build:binary

# Build complete multi-OS distribution bundles (Windows, macOS, Linux, NPM)
npm run build:releases

📄 License

Distributed under the Functional Source License, Version 1.1, MIT Future License (FSL-1.1-MIT). See LICENSE for more information.

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

📉 Context-Diet v1.3.2

📑 Table of Contents

🚀 Quick Start

🔌 One-Click MCP Installation

📦 Multi-OS Packages & Installation

💎 Editions & Feature Matrix (Community vs Pro)

🛠 CLI Command Reference

1. diet init [options]

2. diet read <path> [options]

2. diet exec <command...>

3. diet expand <ref_id> [options]

4. diet doctor [--fix] [-y, --yes]

5. diet fix-deps [-y, --yes]

6. diet stats [path]

7. diet audit <path>

8. diet search <query> [path]

9. diet tree [path]

10. diet query <pattern> [target_path]

11. diet pack [target_path]

12. diet memory [action] [args...] (2-Tier Atomic Memory)

13. diet subagent [action] [sub_agent_id] 🔒 (Pro)

14. diet server (or diet serve, diet mcp)

15. diet clean

16. diet activate <key> [options]

17. diet license [options]

18. diet check [path] [options] (alias: guard)

19. diet ui [options]