The part of this I actually want is the session log. Every agent I’ve built or debugged eventually comes down to the same question: what did the model actually see? Harness answers it by design — an append-only record of system prompts, tool calls, results, and every context injection, with resume, fork, and replay running off the same event stream.
The “everything is a plugin” architecture is the second draw for a self-hoster: models, sandboxes, storage, even the UI get swapped in configuration, not by forking source. And Minimal mode — persistent bash plus a file editor, nothing else — looks like an honest way to benchmark models without harness sugar inflating the results.
My plan: npx @deepseek-ai/dsh web, poke at the Trajectory view, keep it off anything load-bearing. It’s a developer preview, and they say the core plugins and APIs will keep moving.
The story — DeepSeek has released DeepSeek Harness in developer preview, source code included, built on the Cordis plugin system: every capability — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI — is a plugin that can be swapped or extended in configuration without changing the source. Every run is recorded in an append-only session log, inspectable in a Trajectory view. Four runtime modes ship: Standard, Code, Minimal, and Creator. It starts with npx or from source, and core plugins and APIs will continue to evolve. (Source)