Should I commit the .workflows directory to Git?

Usually, you should add .workflows to your .gitignore as it represents transient working state. However, you can version it if your team requires a permanent audit trail of AI research.

How does this skill prevent context exhaustion?

Instead of agents returning thousands of tokens to the main chat, they write their full report to a local file and return only a 2-3 sentence summary, keeping the parent context window clean.

Can I run multiple agents in parallel with this pattern?

Yes, it is designed for batch dispatching. You can run 10-15 agents in the background, then use a synthesis step to read all their findings from disk once they are complete.

Where are the agent output files stored?

Files are stored in a .workflows/ directory at your repository root. This location ensures that research artifacts survive worktree transitions and are shared across different coding sessions.

Does this require special tools or setup?

It relies on standard file-writing tools available to Claude Code and follows a specific directory convention to ensure consistency across different types of tasks like research, planning, and review.

Disk-Persisted Agent Pattern

Name: Disk-Persisted Agent Pattern
Author: adamfeldman

byadamfeldman

•

生産性とワークフロー

Optimizes multi-agent workflows by persisting detailed subagent outputs to disk to prevent context exhaustion and ensure session recovery.

The Disk-Persisted Agent Pattern is a specialized workflow architecture for Claude Code designed to handle large-scale agentic tasks without hitting context window limits. By mandating that subagents write their full analysis and code findings directly to disk while returning only a concise summary to the parent context, this skill maintains a lean conversation state. It provides a standardized directory convention for research, planning, and code reviews, ensuring that all AI-generated insights are traceable, recoverable, and accessible for synthesis even after context compaction or session restarts.

主な機能

01Context exhaustion prevention via disk-based output storage

021 GitHub stars

03Batch dispatching pattern for managing 10-15 parallel agents

04Automatic session recovery and traceability for complex workflows

05Synthesis-ready architecture for aggregating distributed agent findings

06Standardized $WORKFLOWS_ROOT directory structure for multi-session persistence

ユースケース

01Executing massive codebase audits where dozens of files need individual security or performance analysis

02Running multi-stage research and planning phases that require extensive documentation without bloating the active chat

03Maintaining a permanent, disk-based record of agent findings for team traceability and future project context

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add adamfeldman/compound-workflows disk-persist-agents

For use in Claude.ai and ChatGPT

主な機能

01Context exhaustion prevention via disk-based output storage

021 GitHub stars

03Batch dispatching pattern for managing 10-15 parallel agents

04Automatic session recovery and traceability for complex workflows

05Synthesis-ready architecture for aggregating distributed agent findings

06Standardized $WORKFLOWS_ROOT directory structure for multi-session persistence

ユースケース

01Executing massive codebase audits where dozens of files need individual security or performance analysis

02Running multi-stage research and planning phases that require extensive documentation without bloating the active chat

03Maintaining a permanent, disk-based record of agent findings for team traceability and future project context