Real Harness interaction test

August 17, 2026 ยท View on GitHub

Test date: 2026-08-17.

Environment

  • DeepSeek Harness: @deepseek-ai/dsh@0.1.0-rc.7 (next at test time)
  • Node.js: v24.14.1
  • OS: Windows x64
  • Probe: Web startup on a random loopback port, 12-second timeout, 350-ms post-readiness settle window

Fixture

Three local bundle packages were installed into an isolated web profile using the official dsh plugin --profile web add flow:

  • dsh-reducer-fixture-a registers a global tool with a fixed name.
  • dsh-reducer-fixture-b registers a second global tool with the same name.
  • dsh-reducer-fixture-c is a no-op.

A and B each start successfully alone. Together they trigger Harness's duplicate tool registration error. C is irrelevant. This specifically tests a failure that naive one-at-a-time diagnosis can miss.

Result

#1 FAIL       3 active: fixture-a, fixture-b, fixture-c
#2 PASS       0 active: (none)
#3 PASS       1 active: fixture-c
#4 FAIL       2 active: fixture-a, fixture-b
#5 PASS       1 active: fixture-b
#6 PASS       1 active: fixture-a

Minimal failing set (2):
  - dsh-reducer-fixture-a
  - dsh-reducer-fixture-b
Verified 1-minimal: yes; 6 configuration(s) tested.

The source profile fingerprint was identical before and after the run. The generated report contained no drive letter, username, workspace path, or common credential token; local roots in stack traces were replaced with stable labels.

The checked-in fixtures live under test/real-fixtures/. They are intentionally not part of the normal CI profile installation because a full Harness install is large; the algorithmic interaction case is covered by the network-free integration test on every CI run.