DeepSeek Harness: The End of Claude Code?

Overview

This video reviews DeepSeek’s “Harness” developer preview, highlighting two standout design ideas:

1. Everything is modular “plug-ins”
2. Full run transparency / traceability

---

Core Concepts / Product Features

Everything is a plug-in architecture

• UI elements (e.g., the UI sidebar) are implemented as plugins.
• System-level components—such as tool use capabilities, the system prompt, and other features—can be activated/deactivated.
• Plugins can be managed via configuration (at least in the preview), with the goal of becoming more user-friendly later.

---

Traceable execution (“every run is traceable”)

The harness exposes detailed logs for:

• Every message
• Every tool call
• Every skill loaded

The video emphasizes that this approach is the opposite of Anthropic/Claude-style systems that obscure the “thinking process.”

It also includes a trajectory view with:

• Step-by-step graphical analysis
• Tool call payloads and results
• Timing and run statistics

Additionally, sessions can be exported as a zipped log (e.g., containing session.jsonl).

---

Formal underpinning / “calculus” idea

• The harness repository is said to be based on Cordis.
• A linked paper describes a formal paradigm where actions are reversible (i.e., via inverse/revert actions), along with:
  • Dynamic composition / temporal composability
  • Spatial composability through dependency management between components

---

Practical Guide / Setup Tutorial Shown

Install from source

• Clone the repository.
• Install using pnpm.
• Run build.
• Start the web UI (example: pnpm dsh web).

---

Configure API / model providers

• First-time setup asks for a DeepSeek API key.
• You can also connect other providers via:
  • settings → models → add provider

Example providers shown:

• OpenCodeGo models (selecting a model like GLM5.2)
• Custom Ollama
  • Runs locally on hardware
  • Models are fetched and then selected

---

Run a workflow and inspect traceability

• Example workflow: generate a simple Flask to-do app.
• The demo shows inspecting:
  • System prompt vs user prompt
  • Context loaded
  • Assistant “thinking” / outputs
  • Bash tool calls, such as directory listings and command payloads/results
• It then runs the generated app using commands such as:
  • activating a virtual environment
  • running python app.py
• Finally, it visits the app in the UI/browser.

---

Plugin System Demonstration (Creator Mode)

Disable a plugin via config

• The sidebar plugin can be disabled by editing a YAML file (e.g., a cordis patchy.yaml under the repo’s profiles directory).
• Example config behavior:
  • plugin id: UI sidebar
  • disabled: true
• Reloading removes the sidebar; reverting the config restores it.

---

Dynamically add plugins at runtime

• In creator mode, the system loads a Cordis plugin development skill.
• New UI/tool behavior can be added during an active session.
• Examples created on the fly include:
  • A cat overlay animation at the bottom
  • A calculator overlay (bottom-right), controlled via UI overlay prompts

The demo notes that created UI may have limitations (e.g., imperfect button display/interaction), but the core takeaway is real-time plugin loading and control.

---

Model / Provider Flexibility Shown

• The harness supports multiple inference backends and models via its provider system.
• Examples include:
  • DeepSeek via API key
  • Ollama locally
• The trace/trajectory UI is presented as consistent across providers.

---

Main Speakers / Sources

• Speaker: The video’s creator/host (name not given in subtitles)
• Primary referenced sources/tech:
  • DeepSeek Harness (developer preview)
  • Cordis (foundation referenced in the repo)
  • The linked Cordis paper (formal math explanation referenced)
  • Mentioned comparison target: Anthropic/Claude (contrast regarding hiding the “thinking process”)