Phantom Agent — AI skill for Claude Code

Phantom — Autonomous Agent for BitGN PAC1 Challenge

[Русский](docs/README_RU.md) | [中文](docs/README_ZH.md)

An autonomous file-system agent built with [OpenAI Agents SDK](https://github.com/openai/openai-agents-python) that solves the [BitGN PAC1 Challenge](https://bitgn.com/challenge/PAC) — a benchmark for AI agents operating in sandboxed virtual environments.

**Current score: ~86% (37/43 tasks)**



What is PAC1?

[BitGN](https://bitgn.com) runs agent benchmarks where autonomous agents solve real-world tasks inside isolated sandbox VMs. Each task gives the agent a file-system workspace and a natural language instruction. The agent must explore, reason, and execute — no human in the loop.


PAC1 covers 43 tasks across:

• CRM operations — lookups, email sending, invoice handling
• Knowledge management — capture, distill, cleanup
• Inbox processing — with prompt injection traps and OTP verification
• Security — detecting and denying hostile payloads

Learn more: [bitgn.com/challenge/PAC](https://bitgn.com/challenge/PAC)

Architecture

User task → LLM Classifier (picks skill) → Agent(system_prompt + skill_prompt + task)
  → ReAct loop: LLM → tool call → result → LLM → ... → report_completion

• 12 specialized skills with hot-reloadable prompts (edit .md files, no restart needed)
• Dual classifier — LLM-first with regex fallback and override logic
• Self-correcting agent — can call list_skills / get_skill_instructions to switch workflows mid-task
• Auto grounding refs — tracks read/written files, injects references if model forgets
• Retry on empty — retries up to 3x if model returns text without tool calls
• Live dashboard — React + Vite with SSE streaming, heatmap compare, token tracking

Quick Start

Prerequisites

• Python 3.12+
• [uv](