qiongli

Qiongli 2.x

2.0.0 is the first stable release of the native CLI, built from main. 2.x remains the development and prerelease branch. See the release notes for this release's scope and upgrade steps.

Native academic research CLI with embedded Skills, templates and Lite/Full MCP. Keep your chosen Codex or Claude Code Host and model settings. Qiongli supplies research records, source links, reviewable changes and stage handoffs; no Qiongli desktop App is required.

Installation and updates share one guide: run qiongli install in a terminal, or qiongli install plugin to reuse an existing Host registration. Check the local MCP protocol with qiongli mcp check. A Plugin includes the research Skills and its MCP runtime; standalone Skills are an optional export. The 2.x guide explains the choices and how to verify the first Host session.

Explore the offline Research Graph example, then choose the research guidance, reply-only entry or optional context hooks you need.

What changes from 1.x?

Late 1.x already had independent native Lite Plugins. The 2.0 change extends native delivery to the complete CLI and Full MCP. Models and Hosts remain your choice; actual subagents come from the Host. See the detailed comparison and migration guide, including compatibility limits and rollback.

Qiongli 2.x Standalone Downloads

Download, extract, and run. You do not need Python, Node.js, Rust or a package manager.

Choose the complete CLI from GitHub Release v2.0.0. After extraction, run ./qiongli --help (PowerShell: .\qiongli.exe --help). The program already includes the research Skills, templates and Lite/Full MCP resources. You can use it from that folder; installing an App or changing PATH is optional.

Windows 2.0 releases include the C runtime in the executable; no Visual C++ runtime installation is needed. Linux uses system libraries with glibc 2.35+.

Verify with the release's SHA256SUMS before running. See extraction, PATH and MCP setup for step-by-step instructions. Choose these platform archives from Assets, rather than GitHub's automatic Source code downloads. Host applications and online services remain separate.

Package-manager users can use npm install --global qiongli@latest with Node.js 18+, or follow the PyPI instructions with Python 3.9+. Both distribute the native executable. Cargo builds it from source with Rust 1.97+ and a native linker.

cargo install qiongli --version 2.0.0 --locked

Cargo provides qiongli and ql. Choose the archives above to skip compilation.

Run qiongli for help, qiongli project to list projects, or qiongli setup to review installed CLI versions and manual archive/uninstall steps. npm can show the review during a foreground install; pip and Cargo users run it afterward. No files or settings are changed. See the installation review guide.

install plugin and upgrade plugin use the same flow, with file previews and separate confirmation for official Codex/Claude registration. See bundled content installation.

After installation

qiongli --version
qiongli install
qiongli mcp check
qiongli doctor
qiongli setup
qiongli content
qiongli help install plugin

setup reviews visible CLI copies and gives manual removal or archive guidance. It does not uninstall programs, move files or change PATH. qiongli and ql use the same commands; direct downloads provide qiongli, while package-manager installs also provide ql. Use --json in scripts.

install plugin previews bundled files and asks for confirmation before exporting. It then asks separately for official Host registration. Later runs discover and refresh that source; upgrade plugin is an alias. install skills exports .qiongli-skills; use the Plugin installation path to load the workflow in a Host. upgrade cli explains how to update through the original channel without running a package manager.

Follow the installation and upgrade examples. After a Plugin update, start a new Host session and check that its tools are available. The Plugin guide offers optional context hooks (off by default). install plugin --hooks context includes them; --hooks off removes their Plugin configuration. See Hook setup and verification.

Skills, MCP and research records

Use no-qiongli ($no-qiongli in Codex), also triggered by “reply only” / “NoQ问理” / “仅回复” / “no 处理”: answer from the conversation without model-invoked tools, agents or file operations. See reply-only scope and limits.

New Codex Plugin builds have 20 workflow shortcuts and the general $qiongli entry; 82 internal skill cards stay available on demand. Both Codex and Claude also include the independent no-qiongli entry; Claude retains the main research Skill. Native Marketplace platform packages start Lite MCP; CLI-exported local Plugins start Full MCP. A shared Skill name does not imply the same available tools.

Graph improvements are checked through stable identities, multiple sources, resolvable evidence, deterministic rebuilding and stale-read refusal. The Host still needs to normalize authorized material into canonical records. This is not an arbitrary-folder or PDF semantic scanner. See Graph scope and checks.

Development and documentation

Shared research content lives in content/; native services live in packages/qiongli-native/. npm, PyPI and Cargo keep their own launchers and installation instructions while sharing the CLI description and release identity. Generate packages from their canonical sources; installed caches and generated mirrors are not edit targets.

• Quickstart and 2.x command guide
• Architecture and contribution workflow
• 1.x command reference, retained for older installations and migration
• 1.x compatibility materialization, including staged materialization and npm package contract tests

Reuse existing approval, rollback, package-identity and content checks. A Rust rewrite alone does not establish a speed or maintenance-cost improvement; those claims need measurements against the same task and environment.

Credit

The Academic Idea Funnel and Academic Grill Loop are an academic adaptation of Matt Pocock's grill-me pattern for academic idea-discovery, evidence, rival explanations and feasibility. Thanks to the linux.do community for practical discussion and feedback.