Agent Skill Kit

ask-parallel-auditor

An orchestrator skill designed to bypass single-agent context limits by chunking repositories and spawning parallel subagents to conduct massive audits.

Purpose

When executing a repository-wide security scan, architectural audit, or massive refactor, a single AI agent will quickly blow out its context window if it tries to read hundreds of files sequentially.

ask-parallel-auditor implements the Hierarchical Multi-Agent System (HMAS) pattern. It acts as the Strategy Layer Orchestrator. It does not read the files itself. Instead, it chunks the repository and spawns multiple, isolated subagents (like ask-owasp-security-review) operating in parallel across isolated Git worktrees.

Usage

You invoke this skill when you need a comprehensive scan of a large project.

1. Tell the agent to “Run a parallel audit on the src/ directory”.
2. The orchestrator chunk the src/ directory.
3. The orchestrator spawns background bash/python scripts to kick off identical subagents assigned to different chunks.
4. The orchestrator waits for the subagents to return their JSON/Markdown findings.
5. The orchestrator merges the results into a massive PARALLEL_AUDIT_REPORT.md.

Features

• Chunking Logic: Groups files logically (by directory or by size) before delegation.
• Git Worktree Isolation: Ensures subagents operate in isolated physical directories so they don’t corrupt the main contextual workspace or conflict over file locks.
• Subagent Routing: Delegates to specialized subagents for the actual labor.
• Result Aggregation: Compiles massive subagent outputs into a unified report.

Examples

Before (Single Agent Failure)

The agent opens 80 files sequentially, forgets its core instructions around file 40, hallucinates code, and crashes with a token exhaustion error.

After (Parallel Orchestration)

The orchestrator splits the 80 files into 4 chunks. It spawns 4 background workers. 30 seconds later, it receives 4 JSON reports and synthesizes them for you perfectly.

Best Practices

• Make sure subagents being spawned are highly constrained (e.g., disallowedTools: Write, Edit) to prevent rogue autonomous destruction.
• Use ask-context-janitor on the aggregated results if the final report is still too massive.

This site is open source. Improve this page.