Parallax is a lightweight enhancement layer on top of your coding-agent harness. You own the goal; the agents own the work, context management, and independent acceptance.

Install Parallax, then just wish it. Make it real. Say what you want, make a few choices, hand it off — an agent team does the work under independent review and in a clean session, audits the result, and commits it clean.

中文文档

The three install badges jump to the install section of this README — they are not one-click marketplace installs.

Three steps

Plugin trigger words: wish · Parallax · qstart

Wish — say what you would like, as casually as you like. You do not need to have it figured out; add more in several messages and say "that's all" when you are done.
Choose — the agent researches first, then aligns with you one question at a time; you just pick.
Hand off — the agent team takes it from there: plan → independent review → new-session execution → independent audit → scoped commit → wrap-up.

Two modes

solo (default): you work directly with the execution session; the current session is the executor. Suited to everyday use.

team: your conversation always stays with the orchestrator while the executor runs in its own session. Suited to large and very long tasks.

Automatic context management

At task start: once the plan passes review, execution continues in a new session.

Long-running tasks (team tier): context is handed off automatically at about 60%, keeping the AI clear-headed with no interruption to the task.

https://github.com/user-attachments/assets/c85a9a89-2afa-44b0-a4f5-bf6db1db3001

Platform support

Claude Code — Linux ✓ · Windows (WSL) ✓ · macOS ?
DeepSeek Harness — Linux ✓ · Windows ✓ · macOS ?
GitHub Copilot (VS Code) — Windows ✓ · Linux ? · macOS ?. Currently solo only.

? means the code is written for compatibility and portability, but has not been actually tested; please open an issue with what you see.

Install

Claude Code

Pick one of the following — the skills are plain folders under platforms/claude-code/skills/:

Option 1: Installer (recommended):
```
git clone https://github.com/AI4XE/Parallax.git
bash Parallax/install.sh /path/to/your-project
```
Installs all seven skills into <project>/.claude/skills/, atomically, and preserves your local *.conf edits on re-install/upgrade.
Option 2: Manual copy: copy each folder from platforms/claude-code/skills/ into <your-project>/.claude/skills/. Fully functional without the installer; re-copying later overwrites local *.conf edits.
Option 3: Claude Code plugin marketplace: /plugin marketplace add AI4XE/Parallax, then /plugin install parallax. Slash commands get namespaced (/parallax:qstart); natural-language triggering ("qstart") still finds them. This path is structurally validated in the repository (plugin manifest, skill paths, marketplace list) but has not been exercised end-to-end against a remote marketplace.

Then, inside that project in Claude Code, just say: qstart (or wish / Parallax).

Permissions: the team track starts a separate headless executor process. Use the permission mode you already trust for autonomous work; Parallax does not raise permissions on its own and never treats "hand-off" as approval.

Tip: running Claude Code inside tmux together with a plugin is the most compatible setup; you can also use tools such as cc-connect to connect the session to a chat app for remote follow-up.

DeepSeek Harness

Pick one of the following:

Option 1: Plugin bundle (preferred — one command, no copying, re-run to upgrade):
```
git clone https://github.com/AI4XE/Parallax.git
node Parallax/platforms/deepseek-harness/bundle/build.mjs
dsh plugin --profile <profile> add "file:<absolute path to Parallax>/platforms/deepseek-harness/bundle"
```
Requires the Node runtime supplied by dsh and pnpm visible on the dsh process PATH (e.g. via corepack). Windows can run both commands natively; quote paths containing spaces. Remove with dsh plugin --profile <profile> remove parallax-dsh. Restart the selected profile to load an installed update. Zero extra setup: the bundle registers the skills and the native handoff hook and applies subagent.maxDepth: 3; do not edit rules, profiles, patches or permission settings for handoff.
Option 2: Installer (skills fallback): bash install.sh --platform deepseek-harness <your-project-dir> — copies the seven skills into <project>/.agents/skills/ (atomic swap, preserves local *.conf). DSH discovers them by directory scan (hot reload, no restart). This is skills-only, not a full-bundle substitute: it has no native handoff hook or depth patch, and the complete qgo handoff chain is unsupported.
Option 3: Manual copy: copy each folder from platforms/deepseek-harness/skills/ into <your-project>/.agents/skills/.

Use normal workspace-write permissions, not bypass/full access. Keep ordinary shells: confined environments that block named-pipe capture cannot run some subprocess tools.

Then, in a DSH session with your project as the working directory, just say: qstart (or wish / Parallax).

Build safety: build.mjs is the cross-platform source of truth; build.sh is only a convenience wrapper. Only explicit skill resources are copied, a verified manifest protects generated files, and unknown files, links or changed managed output stop the build without deleting them.

Copilot

Pick one of the following:

Option 1: Installer (recommended):

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

Option 2: Manual copy: copy platforms/copilot/skills/* into <your-project>/.github/skills/, plus platforms/copilot/agents/* and platforms/copilot/instructions/copilot-instructions.md into the matching .github/ folders. Do not overwrite an existing copilot-instructions.md — merge it.

Then in VS Code, with that folder open:

Run Developer: Reload Window once (skills and agents are discovered on reload).
Trust the workspace once when prompted (remembered per folder).
Recommended: set the chat permission mode to Allow all (or Autopilot) — setting key chat.defaultConfiguration — so cross-session hand-off and subagent reviews run without approval prompts. The default Manual mode also works; you approve each tool type once per session.

Then just say: qstart (or wish / Parallax).

Skills

Seven skills that turn "please build this" into a supervised, self-reviewing loop:

Skill	Role
qstart	Discussion entry: requirements → research → one-question-at-a-time alignment → complete overview.
qgo	Hand-off: re-checks the discussion, writes the plan, passes the bcheck gate (score ≥ 4/5), then a clean-context agent executes autonomously → rcheck audit → scoped commit.
bcheck	Independent adversarial review of the plan before execution. The doer never grades their own homework.
gclear	Saves a distilled brief and hands it to clean model context, with delivery verification; the platform adapter supplies the native mechanism.
rcheck	Independent reconciliation audit after execution: every "done" claim is verified against reality.
xwrap	Wrap-up: obvious-problem report, todo update, wrapup doc, smart scoped git commit.
wish	One-line wish entry (许愿创造 / 许愿编程) that forwards to qstart.

One line: qstart (align the wish) → qgo (release; plan gated by bcheck → clean-context execution → rcheck audit → commit) → xwrap (wrap up).

Two execution tracks: solo is the default (one agent does it all after the hand-off); say team to add an orchestrator watching a separate executor. ver / nover select versioned vs non-versioned tasks.

How the hand-off works on each harness:

Claude Code: gclear drives /clear and verifies the injection against the session transcript; the team track runs a headless stream-json executor in its own process, started and watched by the bundled driver. The CLI tmux requirement below still applies to gclear's detached launcher.
DeepSeek Harness: the full bundle provides parallax_handoff, which replaces model-visible history with a concise handoff at a native safe step. The SID, agent, append-only user transcript, Web tab/composer and child ownership stay intact, and work resumes automatically. In solo you keep talking directly to the executor; in team the original window stays the orchestrator and manages native child executors via list_agents / send_message / interrupt_agent. The bundle sets subagent.maxDepth: 3, so an executor can spawn its own independent reviewers.

Tested versions

Claude Code 2.1.322, 2.1.140
DeepSeek Harness (dsh) 0.2.2-rc.1
VS Code 1.146.0 with bundled GitHub Copilot Chat 0.74.0

Philosophy

The harness is the key. Real quality comes from agent collaboration under harness constraints — not from a magic prompt or a bigger model.
Three steps, one hand-off. Wish → choose → hand off. You own the goal and the boundaries; the agents execute, manage context, and independently verify and cross-review, so you steer instead of supervising every step.
The loop. Input → analysis → plan → execution → acceptance → delivery. Every step gets independent review, scoring, and iteration — the one doing the work and the one checking are different entities with isolated contexts.
Three pillars: phase isolation (strategy vs tactics) · adversarial acceptance (anti-sycophancy: the reviewer's job is to break it) · TDD (see it fail before you see it pass).

Name: from Greek parallaxis — para- "beside" + allassein "to change". Viewed from beside, the position shifts: change the viewpoint and what you see changes. Self-review from inside cannot; an outside look can.

Compatibility

Git: recommended — the workflow uses Git roots, scoped commits, and hand-off checkpoints; qstart runs one cheap git --version during research and asks once whether to install it if missing (it does not block).
Claude Code: Linux · macOS · Windows WSL. Bash 4+, Python ≥ 3.11 (stdlib only; no third-party packages), tmux for the CLI hand-off launcher.
DeepSeek Harness: Runtime uses only dsh's Node and injected native services — no Python, Bash, zstd CLI, tmux or patched harness.
GitHub Copilot: pure Markdown skills, zero runtime dependencies. Solo only.

Configuration

Everything works with zero configuration. Optional knobs, each a tiny local file (kept across upgrades by the installer):

skills/bcheck/skill.conf, skills/rcheck/skill.conf — review_model= : model for the independent review subagent. Empty = session default (never up- or down-shifted).
Claude Code only: skills/gclear/gclear.conf — MODEL_ALIASES="" : space-separated alias=full-id pairs for gclear --model; plus timeout/retry tuning keys.
Default reply language per skill (trigger-message language and spoken requests override it): Claude Code reads the display_lang frontmatter key; DeepSeek Harness reads metadata.display_lang.

Upgrading

Re-run the installer (bash install.sh [--platform <p>] <project>) — it swaps skills atomically and keeps your local *.conf files. Claude Code plugin installs: /plugin update parallax. DeepSeek Harness bundle installs: re-run node platforms/deepseek-harness/bundle/build.mjs, then dsh plugin --profile <profile> add "file:<bundle path>" and restart that profile. Each skill versions itself in its SKILL.md frontmatter.

Contact

Questions or feedback: open an issue or email [email protected].

MIT — see LICENSE.