AI4XE/Parallax

Parallax consists of seven skills: qstart first conducts research and aligns requirements through one question at a time; qgo writes a plan, executes it in a clean new session after passing bcheck review (score above 4); rcheck verifies each item post-execution; and xwrap finalizes the work and commits to Git by scope. The default solo mode is executed by the current session; team mode spawns a separate execution session, automatically handing off when context usage reaches approximately 60%. It is installed on DeepSeek Harness as the npm plugin parallax-dsh, which includes native handoff hooks.

Parallax consists of seven skills: qstart first conducts research and aligns requirements through one question at a time; qgo writes a plan, executes it in a clean new session after passing bcheck review (score above 4); rcheck verifies each item post-execution; and xwrap finalizes the work and commits to Git by scope. The default solo mode is executed by the current session; team mode spawns a separate execution session, automatically handing off when context usage reaches approximately 60%. It is installed on DeepSeek Harness as the npm plugin parallax-dsh, which includes native handoff hooks.

Why it matters The agents performing the work and those checking it operate in different contexts: plans must pass independent review before execution, and completion claims are verified item by item by a separate party afterward. It is also usable on DSH.

Stars 53

Reasons

  • The repository was created on 11/15, under the MIT license; the README has been reviewed and lists verified versions for Claude Code, DSH 0.2.2-rc.1, and VS Code
  • The README states that it does not self-elevate permissions and does not treat "handoff" as authorization

Usage

  • In the DSH web interface: Plugins › Add Plugin › enter parallax-dsh, enable it, and restart the configuration; or run dsh plugin --profile <config_name> add parallax-dsh, then say qstart in your project

Risks

  • The Copilot version recommends setting chat permissions to "Allow All," which expands the scope of automatic execution; macOS has not yet been verified
  • Multiple agents for review and execution consume significantly more quota than a single conversation

Open

GitHub

Explore the latest AI Radar updates