DeepSeek open-sourced its coding agent harness, and made the trajectory a first-class object

DeepSeek released dsh, its own coding agent harness, as a developer preview with the source under an MIT license. The design claim is that every capability is a swappable plugin: "models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI", all mounted on a kernel called Cordis, and all replaceable from configuration without touching the harness source. The landing page opens with the equation AGENT = MODEL + HARNESS and glosses it: "The model is the soul of an agent. A harness lets an agent understand its environment, use tools, and keep working in real-world settings." That is a frontier lab defining the primitive in the first person.

The instrumentation is explicit: "Everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection." A Trajectory view reads those records by source, and "resume, fork, search, and replay all operate on the same event stream." Minimal mode provides a runtime built for measurement rather than work: a two-tool agent with persistent bash and str_replace_editor only, kept, in DeepSeek's words, "for benchmarking models in a minimal environment". The company's own API changelog shows it using that mode as the framework for the official V4-Flash benchmark numbers published on July 31, with the effort level, top-p and temperature all stated.

The preview warns in capitals that there will be compatibility-breaking changes. The repository stood at 38,600 stars and 3,000 forks when checked at about 7:45pm Eastern on August 13, hours after release, and the submission sat at the top of Hacker News for most of the day. The Cordis paper the README cites was not read for this item.

Why it matters: Treating the session log as the single event stream that resume, fork and replay all read from is a concrete answer to a problem most harnesses handle with ad hoc transcripts. It separates reconstructing a run from re-running it at a chosen point. Minimal mode makes a second point: by shipping a mode that removes its own harness from a measurement, the vendor concedes that harness and model results are not separable by default. Most published agent benchmarks assume otherwise.