Signals / deepseek-harness profile

deepseek-harness

Sourced signals for deepseek-harness: release changes, governance moves, runtime shifts, and operator consequences. Each links to the run that produced it. The deepseek-harness profile carries the current evergreen state.

August 2026

  1. 2026-08-20 / deepseek-harness

    rc.8 still has no privileged core, and the Web UI still has no login

    • Still a developer preview. Do not expose the port.
    • Read architecture.md at 141eb6fe before installing a plugin. Inspect a bundle's permission mode; do not attribute it to Claude Code or Codex alone. Pin bundle packages at 0.1.0-rc.8; npm latest on those packages is still 0.0.1-rc.1.
  2. 2026-08-17 / deepseek-harness

    Everything is a plugin, including the components that enforce the limits

    • If you write a dsh plugin, understand that your plugin is not governed by the approval gate in the way a tool is governed by a permission system. It sits in the same layer the gate does.
    • Ask of any plugin-everything architecture: which component is not replaceable? If the answer is none, then the enforcement story is a convention rather than a boundary.
    • This is developer preview and the whole project ships to a prerelease. Nothing here is an upgrade instruction; it is a design question to settle before this reaches a stable channel.
  3. 2026-08-17 / deepseek-harness

    Nothing authenticates the Web UI, and the code fence says so in its own comment

    • Do not expose that port. Loopback-only is the entire access control, so any tunnel, container port mapping, or reverse proxy in front of it is the whole security boundary.
    • Treat a third-party plugin as code you are running in-process with no declared permissions, because that is what it is. Read it before you install it.
  4. 2026-08-17 / deepseek-harness

    The lab's own harness ships a supported path to run rival models and drive rival harnesses

    • A harness that runs any model and drives other harnesses is a hedge against the thing its own vendor sells. Read it as evidence about where that vendor thinks the lock-in is, and note that the answer is not the model.
    • For an operator, this is the cheapest available test of whether a harness is really model-agnostic: run it against a competitor's model on the supported path and see what breaks.

<- All signals