How to Download, Install & Use
Agent Junky Tokenomics

Unlike rough "word count" estimates, Agent Junky uses real Byte-Pair Encoding (BPE) algorithms matching OpenAI (cl100k_base and o200k_base) and Anthropic Claude 3.7. As you type, edit, or open any file, the telemetry item in your status bar computes:

Transcript Surgery (sessionPruner.ts) is a deterministic AST cleaner. It scans your conversation history, keeps recent turns untouched, and replaces older dead terminal outputs with concise 1-line semantic tombstones.

When you've been working with Claude Code or Cursor for hours and context approaches 100k+ tokens, native compaction triggers. Native compaction uses an LLM to write a summary of the chat. This often forgets edge cases or loses track of what you were just editing.

Agent Junky replaces this with Deterministic State Handover. It queries your actual Git index and IDE memory to generate a structured handover file called SESSION_STATE.md:

Before passing a whole codebase or folder into an agent with @folder or tool execution, audit its token weight. Often, a single auto-generated file (like openapi.json or a test fixture) can silently eat 60% of the entire token allowance.

When an agent needs to know how to call code from another file, it only needs the function signatures, class declarations, argument types, and return types. It does not need 300 lines of internal SQL queries and loops.

export class UserService {
  constructor(private db: Database) {}

  async findUser(id: string): Promise<User> {
    const query = "SELECT * FROM users WHERE id = ?";
    const result = await this.db.query(query, [id]);
    if (!result.rows.length) {
      throw new UserNotFoundError(id);
    }
    return mapRowToUser(result.rows[0]);
  }
}

export class UserService {
  constructor(private db: Database);
  findUser(id: string): Promise<User>;
}