I was tired of AI agents hallucinating fixes, so I built Agent Superpowers

说明与用途

作者创建了 Agent Superpowers 套件，以消除 AI 编码助手的错误修复，提供确定性本地工具和严格的守护规则。

问题

AI coding assistants often hallucinate fixes by guessing changes on static code, causing silent regressions, token budget waste, and missed edge cases.

结果

Agent Superpowers mitigates hallucinated fixes by enforcing deterministic local tooling, reducing regressions, and improving code quality.

工作流

• 1. Agent creates isolated minimal test reproducing the defect (ghost-runner).
• 2. Agent applies surgical patch only after test fails (RED) and verifies it passes (GREEN).
• 3. Pre-commit audit runs adversarial-duelist to fuzz proposed functions with boundary and extreme inputs.
• 4. Token-squeezer compresses codebase context by 50‑70% using AST parsing before prompt injection.
• 5. api-synthesizer ingests network traffic to generate typed SDKs.
• 6. arch-archaeologist maps inter-module dependencies and produces Mermaid.js diagrams.

证据

They make assumptions, introduce silent regressions, blow through token budgets, and miss catastrophic edge cases.

来源

Each skill follows strict behavioral guardrails and runs deterministic local tooling.

来源

Prohibits the agent from editing production code until it creates an isolated minimal test reproducing the reported defect.

来源

An offensive auditor that attacks proposed functions with boundary overflows, null-byte injections, zero-divisions, and extreme inputs before allowing a merge.

来源

Shrinks heavy codebase context by 50-70% via AST parsing, extracting clean type interfaces and eliminating token bloat before prompt injection.

来源

Ingests raw network traffic (cURL commands or browser HAR dumps) and automatically generates typed Python/TypeScript client SDKs.

来源

Traces inter-module dependencies across messy repos, detects uncalled zombie functions, and renders living Mermaid.js architecture diagrams.

来源