• Sources: primary, repository
  • Summary: The page states the harness is MIT licensed with source included, built on the Cordis plugin system, with models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI all supplied as plugins that are selected or replaced in configuration rather than in the harness source. It records every system prompt, reasoning step, tool call and result, subagent scheduling decision, and context injection in an append-only session log that resume, fork, search, and replay all read from. Four runtime modes are named, including a minimal mode carrying only a persistent bash tool and a str_replace_editor, which the DeepSeek-V4-Pro-0813 model card names as the mode behind its code-agent benchmark numbers.
  • Why it matters: The harness behind a published code-agent benchmark number is now readable, so the number can be reproduced rather than taken on the vendor's word.
  • Follow-up: DeepSeek states the core plugins and APIs will keep changing while the developer preview lasts, so track the point at which they stabilise.

send feedback on this story