⚡ Prism — Intelligent Token Optimization for Kiro CLI

A transparent PTY proxy that compresses prompts before they reach the model, reducing token usage without changing how you work.

User Input ──► Prism (detects + compresses) ──► Kiro CLI Classic ──► Model

Optimized for Kiro Classic — no character limits, full pipeline on every prompt.

---

How It Works

Prism intercepts your prompts via a PTY wrapper around kiro-cli --classic, runs them through a multi-stage compression pipeline, and forwards the optimized version. You type normally — compression is invisible.

Pipeline

┌─────────────────────────────────────────────────┐
│  1. Noise Stripping (greetings, filler, typos)  │
│  2. Content Detection (code/logs/diff/HTML/text) │
│  3. ML Compression (Kompress ModernBERT model)   │
│  4. Metrics Logging (SQLite + Dashboard)         │
└─────────────────────────────────────────────────┘

---

Quick Start

# Clone and setup
git clone <repo-url> prism && cd prism
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Run (starts model warm-up + dashboard + Kiro proxy)
./start

Prerequisite: kiro-cli must be installed and kiro-cli --classic must work.

By default, Kiro opens at your home directory. To start in a specific project:

./start ~/projects/my-app

To launch the dashboard separately:

streamlit run dashboard.py

---

Features

Feature	Description
Noise Stripping	Removes greetings, filler words, punctuation spam — with fuzzy matching for typos like "hellloooo"
Content-Aware Routing	Detects content type and routes to specialized compressors
Tree-Sitter AST Compression	Removes comments, collapses whitespace while preserving code logic (Python, JS, TS, Rust, Go)
Log Compression	Pattern deduplication, timestamp normalization
Diff Compression	Hunks-only extraction, metadata stripping
HTML Extraction	Content extraction, tag removal
ML Token Compression	ModernBERT-based per-token keep/discard inference via Kompress
Adaptive Prediction	6-feature ML model predicts optimal compression ratio per prompt
Safety Fallback	Reverts to original if compression exceeds 65% or expands the prompt
Live Dashboard	Streamlit dashboard with real-time compression metrics

---

Content-Specific Compressors

Content Type	Strategy
Source Code	Tree-sitter AST-aware — removes comments, collapses whitespace, preserves logic
Logs	Pattern deduplication, timestamp normalization
Git Diffs	Hunks-only extraction, metadata stripping
HTML	Content extraction, tag removal
JSON Arrays	Structural deduplication
Plain Text	Greeting/filler removal + ML token compression

---

Project Structure

prism/
├── kiro_optimizer.py          # Main entry point — PTY proxy + optimization pipeline
├── dashboard.py               # Streamlit metrics dashboard
├── prism_warm.py              # Model warm-up script
├── start                      # Quick-start launcher
├── requirements.txt
├── README.md
└── prism/                     # Core library
    ├── __init__.py
    ├── model_pool.py          # Shared ONNX/SentenceTransformer model pool
    ├── context_pruner.py      # Context pruning (standalone module)
    ├── prediction.py          # Feature extraction + compression ratio prediction
    ├── eval.py                # Quality evaluation (semantic similarity scoring)
    ├── cache/
    │   └── semantic.py        # FAISS-based semantic cache (standalone module)
    ├── transforms/
    │   ├── base.py            # Base transform interface
    │   ├── content_router.py  # Content detection + routing
    │   ├── kompress_ml.py     # ModernBERT ML compression
    │   ├── code_compressor.py # Regex-based code compression
    │   ├── tree_sitter_compressor.py  # AST-aware code compression
    │   ├── diff_compressor.py # Git diff compression
    │   ├── html_extractor.py  # HTML content extraction
    │   ├── log_compressor.py  # Log deduplication
    │   └── smart_crusher.py   # Aggressive last-resort compression
    ├── compression/
    │   ├── masks.py           # Token masking strategies
    │   └── universal.py       # Universal compression wrapper
    ├── observability/
    │   ├── metrics.py         # SQLite metrics collector
    │   └── tracing.py         # Request tracing

---

Dashboard

Launch with streamlit run dashboard.py to see:

• Total tokens saved
• Compression ratio per request
• Content type distribution
• Per-request history with before/after

---

What's Next

Prism works fully as a standalone prompt optimizer today. Additional features are built and ready to activate with minimal integration:

Feature	Hook Needed	Impact
Context Pruning	beforeSubmit with conversation history access	Auto-drop irrelevant turns, keep context window lean
IDE Integration	beforeSubmit prompt-transform in Kiro VS Code extension	Same compression inside the IDE
Semantic Cache	beforeSubmit + ability to short-circuit with a cached response	Skip API call entirely for 92%+ similar prompts

---

Notes

• First launch downloads models from HuggingFace (~2s) — ./start handles warm-up automatically
• Tree-sitter compression requires language packages (included in requirements.txt)

---

Tech Stack

• Python 3.10+
• ONNX Runtime — fast inference for embedding + compression models
• FAISS — vector similarity search
• Tree-sitter — AST parsing for code compression
• Kompress — ModernBERT token compression model
• Streamlit + Plotly — live dashboard
• tiktoken — token counting (cl100k_base)