DeepSeek Harness Desktop v1.2.0 — Harness Lab

Harness Lab

Stop benchmarking models. Benchmark the harness.

DeepSeek Harness Desktop can now compare two agent execution trajectories
side by side.

Harness Lab shows differences in:

• steps
• tool calls
• retries
• failures
• file reads and writes
• shell commands
• token usage when available
• execution duration

It also detects explainable trajectory divergences such as repeated tool
loops, extra failed commands, unnecessary file churn, different test timing,
and failure recovery.

Privacy first

Harness Lab processes sessions locally.

The renderer does not receive raw prompts, shell command text, raw tool
arguments/results, source session IDs, credentials, or absolute paths.

Validation

• real DeepSeek Harness session smoke tested
• Windows packaged app smoke tested
• 23 automated tests
• portable and installer builds verified

Demo

Harness Lab includes a synthetic Demo Mode so the comparison workflow can
be explored without exposing real session data.

This is a community project and is not an official DeepSeek desktop client.