ClaudeMap

Parallax

A seven-skill enhancement layer for Claude Code (also DeepSeek Harness and VS Code Copilot): wish, choose, hand off — the plan passes an adversarial pre-review, a clean-context executor does the work, an independent audit reconciles every done claim

Plugins & Agentspluginskillshandoffadversarial revieworchestration

Parallax is a lightweight layer on top of a coding-agent harness, installable for Claude Code, DeepSeek Harness (dsh) and GitHub Copilot in VS Code. Seven skills turn 'please build this' into a supervised, self-reviewing loop: qstart (requirements -> research -> one-question-at-a-time alignment), qgo (hand-off: re-check the discussion, write the plan, pass the bcheck gate scoring 4/5 or higher, then a clean-context agent executes autonomously), bcheck (independent adversarial review of the plan before execution — the doer never grades their own homework), gclear (distills a brief into clean model context with delivery verification; on Claude Code it drives /clear and verifies the injection against the session transcript), rcheck (independent reconciliation audit after execution), xwrap (wrap-up report, todo update and smart scoped git commit), and wish (one-line entry that forwards to qstart). Two tracks: solo (default — your session is the executor) and team (an orchestrator session watches a separate headless executor, with context handed off automatically around 60 percent on long tasks). Installation for Claude Code: the installer script atomically swaps all seven skills into the project and preserves local skill.conf edits across upgrades, manual folder copy, or the plugin marketplace path (structurally validated in the repo but not exercised end-to-end against a remote marketplace). Trigger words: wish, Parallax, qstart — the Chinese docs add 许愿创造 and 许愿编程. Tested versions at writing: Claude Code 2.1.322 and 2.1.140, DeepSeek Harness 0.2.2-rc.1, VS Code 1.146.0 with Copilot Chat 0.74.0. MIT.

What it does

  • Seven skills, one loop: qstart align -> qgo hand-off -> bcheck pre-execution adversarial plan review (score >= 4/5) -> clean-context execution -> rcheck post-execution reconciliation -> xwrap scoped commit
  • gclear hand-off: distills the discussion into a brief, drives /clear on Claude Code, and verifies the injection against the session transcript before work resumes
  • Two tracks: solo keeps you talking to the executor; team keeps an orchestrator session watching a separate headless executor with automatic context hand-off around 60 percent on long tasks
  • Cross-harness: same skill set installs for Claude Code, DeepSeek Harness (the dsh bundle registers skills plus a native handoff hook and sets subagent.maxDepth: 3) and GitHub Copilot (solo only)
  • Permission-honest: Parallax never raises permissions by itself and never treats hand-off as approval — you run the permission mode you already trust for autonomous work
  • Installer preserves local skill.conf edits (review model overrides, gclear model aliases, reply language) across upgrades

When to use it

Use when a task deserves plan review and post-hoc verification rather than a single-context run — mid-size features, refactors and migrations where an adversarial bcheck pass and an rcheck audit would catch the doer's blind spots — or as a ready-made orchestrator/executor split for very long tasks that would otherwise drown one session's context.

Getting started

Claude Code (pick one):

git clone https://github.com/AI4XE/Parallax.git
bash Parallax/install.sh /path/to/your-project

or copy each folder from platforms/claude-code/skills/ into your-project/.claude/skills/, or add the marketplace: /plugin marketplace add AI4XE/Parallax then /plugin install parallax (slash commands get /parallax: namespacing). Then, inside that project, say qstart (or wish / Parallax; the Chinese docs also trigger on 许愿创造 / 许愿编程). The team track starts a separate headless executor process — use the permission mode you already trust for autonomous work. DeepSeek Harness: build the bundle with node Parallax/platforms/deepseek-harness/bundle/build.mjs, then dsh plugin --profile <profile> add "file:<absolute path to Parallax>/platforms/deepseek-harness/bundle". VS Code Copilot: bash Parallax/install.sh --platform copilot /path/to/your-project, then reload the window once and trust the workspace.

Things to know

Platform support is not uniform: Claude Code and DeepSeek Harness are tested on Linux and Windows (WSL for Claude Code), with macOS written for compatibility but not actually tested (marked ? in the README); GitHub Copilot is solo-only. The marketplace install path is structurally validated but, by the README's own note, not exercised end-to-end against a remote marketplace. Requirements on the Claude Code path: Bash 4+, Python >= 3.11 (standard library only), tmux for the CLI hand-off launcher. Copilot setup recommends Allow-all chat permission mode for unattended cross-session hand-offs — a real permission widening you opt into. Contact in the README is an email address (MatClaw at agentmail.to). MIT license. 40 stars as of writing (2026-11-16).

Repository

AI4XE/Parallax

Charted

Related on the map

  • clawd-conduit

    Plugin / Agent

    Wire all 15 Claude Code hook events into the Clawd on Desk desktop pet — and turn permission requests into a blocking desktop card.

    • hooks
    • desktop
    • notifications
    • claude-code
  • CometixCode

    Plugin / Agent

    An unofficial Rust 1:1 reimplementation of the Claude Code terminal UI, ported file-for-file onto an iocraft TUI framework.

    • terminal-ui
    • rust
    • reimplementation
    • unofficial
  • Abide

    Plugin / Agent

    Hook-level enforcement for the natural-language rules in AGENTS.md and CLAUDE.md — every edit is judged and violations are repaired in the same turn.

    • hooks
    • code-review
    • rules
    • agents-md