What DeepSeek Harness is
DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI. Upstream, it is described as built around a plugin architecture — “everything is a plugin” — and is licensed MIT. It is currently in developer preview, and the upstream project warns that compatibility-breaking changes are expected during that preview.
What HarnessRouter added
HarnessRouter Community Edition tagged v0.8.0 on 20 August 2026 and added DeepSeek Harness as a version-pinned, developer-preview backend. The tag's conformance commit reports a Full-class UHP run with 52/52 checks executed through a DeepSeek Harness harness, with task checks run on claude-haiku-4.5 via the dsh backend's pi-ai route.
Conformance evidence through DeepSeek Harness
The v0.8.0 evidence is a published Full-class 52/52 UHP conformance report executed through the DeepSeek Harness integration. It demonstrates that HarnessRouter's dsh path satisfies the same Full-class behavioral expectations as its other backends. It does not, by itself, expand the UHP conformance suite beyond 52 checks.
Post-v0.8.0 integration note (HarnessRouter main)
Re-checked 22 August 2026. The following is HarnessRouter implementation work on its unreleased main branch — not a change made by DeepSeek, and not native UHP support.
- The aggregator problem: the deepseek-official adapter sends a
reasoning_effortparameter on every request. DeepSeek's own API accepts it, but aggregators servingdeepseek/*models over the OpenAI request shape refused the whole request (captured via LLMTR by conformance check T-01; one rejection failed four full-class checks). - The handling: HarnessRouter's dsh relay retries such a rejection once without the parameter and strips it for the rest of the turn — a provider that accepts
reasoning_effortkeeps it, and one that refuses it costs a single extra round trip. - Documented result: a live LLMTR run improved from 47/52 (4 failed + 1 skipped) to 52/52 full-class conformant on deepseek-v4-pro, including the artifact-download check that had depended on the failing turn.
This remains point-in-time adapter behavior in front of an upstream developer-preview harness; see the upstream status below.
Upstream status to keep in mind
Because DeepSeek Harness is in developer preview with expected compatibility-breaking changes, integration behavior may shift as upstream evolves. Treat the v0.8.0 integration and its conformance run as a point-in-time result, not a permanent guarantee of the upstream API surface.
What is not proven
v0.8.0 relay handling described above — all HarnessRouter code, not upstream DeepSeek work. The upstream README contains no UHP mention.Why the distinction matters
A reference implementation can adapt an existing harness without that harness vendor changing its own API. Native adoption would be a stronger ecosystem signal: it would mean a vendor, tool or independent runtime exposes or consumes UHP directly, reducing dependence on one adapter implementation.
Related pages
Read HarnessRouter Community Edition, conformance, the adoption tracker, the ecosystem map, and the other harness pages: Codex, Claude Code, Hermes, and Pi.