DeepSeek Harness + REA: How to Reverse-Engineer an Open-Source Project with Proof and Zero GuessworkNote · digest for November 20
Reverse-engineering open-source code with the REA plugin in DeepSeek Harness using evidence.md proof
Note3 sections≈ 5 min

DeepSeek Harness + REA: How to Reverse-Engineer an Open-Source Project with Proof and Zero Guesswork

It sounds like the easiest task in the world: grab a great open-source repo on GitHub, feed it to an AI agent, and say 'figure out how this works and build me an equivalent for my needs'. In practice, you almost always get a shiny front end with hollow fake logic underneath.

To stop the model from making things up, developers install the REA — Reverse Engineer Anything plugin (MIT license, 470,025 stars) inside DeepSeek Harness (an open-source agent app under MIT, 284,219 GitHub stars). DeepSeek Harness is built around Cordis plugins, so the local REA 6.9.0 MCP server released on November 12, 2026 plugs straight in.

Why evidence.md Matters—and Its Most Honest Section

The workflow follows four steps: Study -> Evidence -> Rebuild -> Polish. Study the source, log facts in evidence.md, rebuild your own version, and polish the UI. Under the hood runs DeepSeek V4.1 Flash, a fast model that also reads screenshots to check layouts itself.

REA's main strength is its proof layer. Only files, functions, line numbers, and call chains the agent actually read can enter evidence.md. At the end sits a section for 'what could not be proven'. When an agent admits where it couldn't reach, you can trust the rest of the report.

'Wait—What If the Project Isn't in JavaScript? Will the Analyzer Break?'

REA's built-in static code analysis inspects a program without running it, mapping functions and how they connect. Out of the box, static analysis supports JS/TS, Electron, .NET, websites, and native binaries (via Ghidra, IDA, and Hopper). Yet the pairing handles Python and Ruby too, as three examples show.

On the visual editor Puck (TypeScript, MIT), REA went full throttle. DeepSeek bundled the core into one JS file, and the analyzer scanned 32,000 code nodes, producing 81 findings and over 1,500 call edges. It pinpointed two core functions—insertAction (adds a block to the data state) and moveComponent (handles drag-and-drop). Following that 'blocks -> data -> render' pattern, it built a working landing-page builder.

With Hugging Face's Python library smolagents, the behavior was even more honest: the JS analyzer reported 0 JS files and didn't fake a thing. DeepSeek read the Python files itself, pinned down 7 core files of the agent loop (driver, planner, step generator, executor, final answer), and built a client reporting agent—though it invented a 'local clinic' for the demo, which had to be swapped for a coaching business.

For the Discourse forum (Ruby), the agent received only the relevant slice: a 1,700+ line search class and the Solved plugin. Inside, it found the function sorting threads into 3 statuses (resolved, in progress, unanswered) and used that logic to build an 'unanswered questions radar'.

Setup and Two Early Traps

You need DeepSeek Harness (dsh) installed along with Node.js ^22.23.4, ^24.22.1, or >=26.12.1. Install the rea-agents package from your terminal:

# Connect the REA MCP server and skill to DeepSeek Harness
npx rea-agents@latest setup --client dsh

# Verify the integration (DeepSeek Harness should show healthy)
npx rea-agents doctor --client dsh

# Check available analysis tools (optional)
npx rea-agents capabilities

# Launch the DeepSeek Harness web UI
dsh web

Once running, create a workspace folder and put source inside strictly in read-only mode so the agent can read code without editing it. A shallow clone—downloading a Git repository with only the latest file snapshot and no commit history—is all you need there.

Watch out for two traps early on. First, don't feed in a massive repository all at once: the report gets diluted, so slice out the specific module you need like with Discourse. Second, watch out for context bloat. REA registers roughly 139 MCP tools, which waste model memory on normal tasks; community wrappers like dsh-rea add a quick toggle for that. Also note an important warning from the README: install only from official sources (morluto/rea, rea.tools, npm rea-agents), as fake clones are already circulating.

Five rules for a safe teardown

  • Check evidence.md: the report should cite real files and functions.
  • Read the 'what could not be proven' section and dig deeper if needed.
  • Run the build and click through every button by hand.
  • Verify that the discovered logic actually runs under the hood, not a stub.
  • Keep source in read-only and check the original license (MIT, GPL/AGPL, or proprietary).

That last item is not a formality. Someone else's code is protected by its license, so the right approach here is a clean-room rebuild: write down the core idea in a spec and code your own implementation from scratch without copying the original product.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 4