prunetool-app/
  prunetool.exe    ← gateway server + dashboard
  prune.exe        ← AI chat CLI
  _internal/       ← bundled runtime (Python, libs, grammars)

PRUNE_CODEBASE_ROOT=C:\Users\kunda\Intellij Workspace\Projects\SpringBootRestApplication

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GROQ_API_KEY=gsk_...

PS C:\Users\kunda\Intellij Workspace\Projects\SpringBootRestApplication> .\prune.exe chat

you> describe_project
 create .prunetool folder and inside creates necessary .js files 
[prune] Project index found — 449 files, 10,523 symbols (scanned 2026-04-29)
[prune] Loading project context... done (~5,200 tokens)
— project context is now active for this session.

prunetool.exe starts
        ↓
Starts gateway on http://localhost:8000  (scanner + /prune API)
Starts proxy   on http://localhost:8080  (OpenAI-compatible IDE endpoint)
        ↓
First run?  → opens http://localhost:8000/#/setup in browser
            → paste API key + set project folder
        ↓
No index yet?  → auto-scans your project once (~15-60s), then never again
Already indexed?  → loads instantly

POST /re-scan
        ↓
   → builds skeleton.json          (every function, class, enum with line numbers)
   → builds folder_map.json        (which folders import from which)
   → writes terminal_context.md    (combined snapshot for /describe)
   → writes last_scan.json         (timestamp, file count, symbol count)
   → [background] auto_annotations.json  (one-sentence AI summary per file via Groq)
        ↓
File watcher runs in background
   → detects code file changes
   → rebuilds skeleton + folder_map + terminal_context automatically
        ↓
Watchdog monitors prune library/ folder
   → detects when you save session notes
   → dashboard shows rescan badge so you can click Project Scan

you type a question
        ↓
Scout model (Groq Llama 8B — fast, ~$0.001/query)
   1. Pre-filters: scores all symbols by keyword overlap → top 1,500
      (max 5 symbols per file — prevents large files crowding out others)
   2. Each symbol shown with: file path, line number, purpose hint, enum values
   3. Scout picks the ~5-10 most relevant files for your question
   4. Extracts only the relevant sections from those files
   5. Assembles compact context: ~3-8K tokens instead of 100K+
        ↓
Complexity classifier (folder spread — no extra API call)
   - 1 folder selected        → simple  → fast cheap model (Haiku, Gemini Flash)
   - 2-3 folders selected     → medium  → balanced model (Sonnet, GPT-4o)
   - 4+ folders selected      → heavy   → powerful model (Opus, o1)
   - checks daily token budget → warns at 90%, switches model at 95%
        ↓
Your chosen LLM gets: pruned context + your question
        ↓
Answer streamed back to your terminal

// provider: anthropic,groq

// provider: anthropic,groq
// Edit this first line to choose providers.
//
// Access methods:
//   anthropic:
//     CLI → claude CLI ✓ detected
//     API → ANTHROPIC_API_KEY not set
//   groq:
//     CLI → groq CLI not detected
//     API → GROQ_API_KEY ✓ detected
//
// PruneTool uses CLI first, API key as fallback.
// Complexity legend:
//   simple = typo fix, rename, add one line, small bug, one-file tweak
//   medium = new function, small feature, explain one file
//   heavy  = architecture, refactor multiple files, explain the whole system
module.exports = {
  models: [
    { id: "claude-haiku-4-5-20251001", label: "Claude Haiku",      model: "claude-haiku-4-5-20251001", complexity: "simple",  dailyTokenGoal: 50000 },
    { id: "claude-sonnet-4-6",         label: "Claude Sonnet",      model: "claude-sonnet-4-6",         complexity: "medium",  dailyTokenGoal: 50000 },
    { id: "claude-opus-4-6",           label: "Claude Opus",        model: "claude-opus-4-6",           complexity: "complex", dailyTokenGoal: 50000 },
    { id: "llama-3-1-8b-instant",      label: "Llama 3.1 8B",      model: "llama-3.1-8b-instant",      complexity: "simple",  dailyTokenGoal: 50000 },
    { id: "llama-3-3-70b-versatile",   label: "Llama 3.3 70B",     model: "llama-3.3-70b-versatile",   complexity: "medium",  dailyTokenGoal: 50000 },
  ]
};

prune.exe chat              Start chat (gateway auto-opens, model picker appears)

describe_project    Load project context into this session
/model <llm model>      choose and Switch model in session. llm model wont change until end of session.
/model auto          Switch to auto-routing by llama instant from .env
/model auto exit     stops auto mode of choosing llm by llama instant from .env
/models              Show active provider models as per auth list and usage
/status              Show gateway status
/clear               Clear conversation history
/quit                Exit

you> describe_project
    ↓
No index?            → runs auto scan first (first time only)
Index < 1 hour old   → loads immediately
Index > 1 hour old   → "Last scan was 3h ago. Rescan? (y/n)"

[prune] Loading project context... done (~5,200 tokens)
— project context is now active for this session.

~/.prunetool/
  .env                    your API keys and project path
  daily_stats.json        token usage per model (resets daily)
  model_contexts.json     cached context window sizes (24h TTL)
  active_model.txt        last selected model
  llms_prunetoolfinder.js your model configuration

<your-project>/
  .prunetool/
    last_scan.json        scan timestamp, file count, symbol count
    skeleton.json         symbol index (every function, class, enum)
    folder_map.json       folder dependency graph
    auto_annotations.json one-line AI summary per file
    annotations.json      user-written folder notes
    project_metadata.json file counts, directory tree
    terminal_context.md   combined snapshot loaded by describe_project
  prune library/
    library.md            session knowledge (written by /save docs)
    PROGRESS.md           current status and next steps

{
  "mcpServers": {
    "prunetool": {
      "url": "http://localhost:8765/mcp"
    }
  }
}

codex mcp add prunetool -- /path/to/prunetool.exe mcp

prunetool/
  server/gateway.py         FastAPI gateway — all HTTP endpoints
  mcp_server.py             HTTP MCP server (port 8765)
  mcp_stdio.py              stdio MCP entry point
  proxy_server.py           OpenAI-compatible local proxy (port 8080)
  prune_cli.py              AI chat CLI (prune.exe)
  prunetool_main.py         Binary entry point (prunetool.exe)
  start_mcp.py              Dev startup script
  llms_prunetoolfinder.js   Shipped default model config
  indexer/
    skeletal_indexer.py     Tree-sitter + regex code parser
    folder_mapper.py        Import graph builder
  pruner/
    pruning_engine.py       Scout ranking + file extraction
    scout.py                Groq/Ollama symbol ranker
    auto_annotator.py       Batch file annotation via Groq
  ui/                       React + Vite dashboard