DeepSeek Rewards dsh-layerfs Dev: 166x Storage Deduplication for Multi-Agent Coding

DeepSeek Rewards dsh-layerfs Plugin Developer

DeepSeek officially rewarded a DSH plugin developer. Community member Yifan Xu publicly reported receiving a 999 yuan airdrop after DeepSeek noticed their dsh-layerfs plugin. This signals DeepSeek's expansion beyond open models and Harness infrastructure to actively engage third-party developers to enrich the agent tooling ecosystem.

dsh-layerfs: LayerFS-Powered Storage for Multi-Agent Development

The dsh-layerfs plugin leverages the underlying LayerFS to organize a project as a shared storage history with multiple independent working branches. This addresses a key challenge: Git manages code versions, but agents need a complete, executable, isolated environment that can efficiently share historical data.

Architecture: Shared History, Incremental Changes, Deduplication

When a main agent receives a complex task, it creates a project and delegates to child agents. Each child agent creates a branch from the current project version and works in an isolated workspace. LayerFS does not copy the entire project per branch; instead it shares committed history, saves incremental changes, and deduplicates storage, drastically reducing disk usage.

The plugin also introduces Sandbox as an execution environment. Agents mount their workspace in a sandbox, run install, build, or test commands, and commit verified changes back to their branch. A JavaScript SDK and CLI ( layerfs) allow agents to create branches, manipulate files, run sandboxes, and manage commits. The plugin connects via DeepSeek Harness's bash execution environment without modifying Harness core.

Visualization: Work and History Tabs

Two new tabs were added to DeepSeek Harness:

Work – shows ongoing agent work: file operations per branch, terminal execution logs, and task progress per workspace.

History – displays version history: branches, commits, merges, and releases.

A sidebar LayerFS project management page provides file tree, history, storage usage, and sandbox status.

Real-World Experiments: Massive Storage Savings

The developer published two concrete multi-agent experiments.

Experiment 1: Wayfare (Node.js Travel Planner)

8 agents, 56 minutes, 147 commits, 31 branches, 13 layers.

All 334 tests passed.

Snapshot logical size: 91.5 MB; actual LayerFS disk usage: 1.21 MB → 27× deduplication .

Experiment 2: Quarry (Rust SQL Database vs SQLite)

8 agents, 4 hours 16 minutes, 913 commits, 56 branches, 9 layers.

891 tests passed, including 800 sqllogictest tests.

Snapshot logical size: 4.93 TB; actual LayerFS disk usage: 8.90 GB → 166× deduplication .

Installation Guide

The recommended installation uses a Coding Agent with the following prompt:

Set up LayerFS × DSH for me: the dsh-layerfs plugin bundle in DeepSeek Harness (dsh) web, on real LayerFS.
1. Read README.md in Ephemeral-AI-Lab/dsh-layerfs, the "Install" section and its "Install step by step" steps. It is the guide: follow its steps in order and for this machine's OS (macOS, Linux, or Windows; Windows uses the WSL2 setup below).
2. Check what it needs first: Node 22.19 or newer, pnpm, Docker (Docker Desktop on macOS), network access, a LayerFS checkout layerfs at the commit in layerfs.json, beside this repository as ../layerfs or named by LAYERFS_SOURCE, and a built deepseek-harness checkout deepseek-harness beside this repository; if there is none, clone it there and run pnpm install && pnpm run build in it. If something is missing that you cannot install, stop and tell me which.
3. Step 1, Install LayerFS: in the LayerFS checkout, build layerfs-daemon and the SDK's native addon in the rust:1.85.1-bookworm image, gather them into the bundle folder, build the runtime image and note its image ID.
4. Step 2: in this repository, run pnpm install and pnpm run build, with LAYERFS_SOURCE set to the LayerFS checkout and DSH_SOURCE to the deepseek-harness checkout (absolute paths).
5. Step 3: make a fresh profile with its own DSH_HOME (not ~/.dsh) and add only dsh-layerfs with dsh's plugin command. Do not add dsh-layerfs-demo.
6. Step 4: write layerfs-host's settings into the profile's cordis.patch.yml: the image ID, plus the bundle folder on macOS or the daemon path on Linux.
7. Step 5: start dsh web with that DSH_HOME and give me the printed http://127.0.0.1:…/?token=… link.
8. Step 6: run pnpm test in this repository, and check that GET /layerfs/status reports no missing settings.
Report what you ran and what each step returned. A long cargo build may be run again until it prints Finished, as the guide says; if anything else fails, stop and give me the exact error, and don't retry it on your own.