Qiongli 2.0: what changes from 1.x

Qiongli 2.0.0 is the first stable release of the Rust-native CLI. It brings the program, research content and MCP implementations into one executable, while keeping the Host and model of your choice. Reading, literature review, study design and writing remain the familiar starting points.

The main change is how those capabilities fit together: fewer runtime pieces to install, one native implementation behind the CLI and MCP, and research records that can be inspected beyond the current conversation. This release covers the CLI and its bundled integrations. It does not mark the broader Desktop replacement, all 1.x migrations or live collaboration acceptance complete.

1.x and 2.0 at a glance

The comparison uses the retained late-1.x implementation, including the 1.19 compatibility baseline. Earlier 1.x releases had different packaging and features; not every difference below applies to every old installation.

Area	Late 1.x	2.0
Main runtime	Python application; npm also provided an asset-management entry	Rust-native CLI with embedded research content and Lite/Full MCP
Direct installation	Requirements varied by entry; native Lite Marketplace bundles already existed	Complete CLI archives for three targets, with no separately installed Python, Node.js or Rust
Package channels	Python and npm entries had different runtime responsibilities	GitHub, npm, PyPI and Cargo deliver the same native CLI contract; each keeps its own installer prerequisites
Plugin and MCP	Native Lite and Python Full followed different runtime paths	Both profiles run natively; Marketplace starts Lite, CLI-installed local Plugins start Full
Installation flow	A larger set of surface, part and provider options	Guided install; install plugin also updates an existing verified export and Host registration
Research guidance	More prescribed stage sequences, roles and generic quotas	Guidance follows the requested outcome, available evidence and applicable protocol; narrow requests can stay narrow
Graph	Citation discovery and expansion supported literature work	Research Graph also links claims, evidence locations, decisions and gaps in canonical project records
Continuity	Workflow artifacts and handoff records	Structured capture, revision checks, Graph snapshots and stage summaries share source identities
Agent work	Workflow roles and orchestration conventions	Explicit task/source scope for Host subagents and handoff packets; the Host still supplies execution
Maintenance	Python runtime plus multiple distribution paths	Shared Rust services and canonical content; package builders bind versions, targets and embedded resources

Install the program, then connect your Host

For the fewest prerequisites, download the matching complete CLI archive from Release v2.0.0, verify SHA256SUMS, extract it and run ./qiongli --help (PowerShell: .\qiongli.exe --help). Supported targets are macOS ARM64, Windows x64 and Linux x64 with glibc 2.35+. Windows includes the C runtime in the executable. GitHub's automatic Source code archives are for building, not this installation path.

npm needs Node.js 18+, PyPI needs Python 3.9+, and Cargo compiles the CLI with Rust 1.97+ and a native linker. Those are channel requirements. They are not requirements for the standalone binary or its exported native Plugin. See the download and package commands.

Run qiongli install to choose a Plugin, standalone Skills or MCP configuration. The recommended local Plugin contains the Skills, its own copy of the program and Full MCP configuration. The Host loads its Plugin cache; copying Skills into ~/.agents/skills is unnecessary. There is no separate Qiongli MCP package or background terminal to maintain.

qiongli install plugin discovers the registered source and offers an update. upgrade plugin and update plugin use the same flow. File changes and official Host registration have separate previews and confirmations. A second Host uses a separate source directory. The completion summary shows the version, source, cache, verified registration and remaining session or Hook checks.

For known Codex Qiongli conflicts, 2.0 can migrate after showing the exact entries to disable and obtaining Host confirmation. It preserves previous sources, caches, unrelated Plugins and model settings. Unsupported identities or changed configuration stop the operation. Claude retains manual conflict handling. If registration fails after an old Plugin is disabled, retry or re-enable it in Codex; the installer reports that state.

Know which tools your Plugin provides

Installation	MCP profile	What it provides
CLI-installed local Plugin	Full, 32 tools	Literature and Zotero operations plus projects, Graph and Host handoffs
Native Marketplace platform Plugin	Lite, 14 tools	Bounded literature, configuration, search-planning and Zotero operations
Standalone Skills export	None automatically	Research instructions and references

Both MCP profiles are compiled into the program. Native Marketplace bundles need no Node/npm bridge, Python interpreter or executable download at launch. Late 1.x already offered independent native Lite bundles; 2.0 extends native delivery to the complete CLI and Full MCP. A common Skill name does not make unavailable tools appear in a Host.

qiongli doctor and qiongli install list inspect local installation state. qiongli mcp check exercises the current CLI's protocol. Open a new Host session to verify actual tool discovery and call qiongli_config_status; check online literature setup separately with qiongli_literature_status. A successful file export alone does not establish a working session. Host applications, provider credentials and Zotero configuration remain separate setup choices.

Give the model direction without prescribing every step

The Skills retain source attribution, method limits and reviewable changes, while removing generic pass counts and compulsory long workflows from small requests. Formal reviews still follow the chosen protocol; a reply about one paragraph does not need a whole-project artifact set.

Journal work keeps two distinct routes: assess and adapt to a journal you have chosen, or assess a manuscript before finding suitable journals. Recommendations must use the actual article type, evidence and current venue requirements. Humanizer guidance improves English and Chinese flow without changing claims, numbers, citations or uncertainty.

Use $no-qiongli in Codex, or say “reply only”, “NoQ问理” or “仅回复”, when you want a conversational answer without model-invoked tools, agents or file work. This is an instruction to the model, not a runtime permission lock. Optional context hooks can remind the Host of continuity and approval boundaries. They are off on first local Plugin installation; enabling them still requires Host trust and a real event check. See collaboration and hooks.

Follow a claim through the Research Graph

The Host reads authorized material and proposes structured research records. After review and approval, Qiongli can rebuild the Graph from those records, retaining claim IDs, citation keys and source locations. Reviewed, proposed and rejected relations remain distinct. Missing evidence stays visible instead of becoming an inferred supporting link.

qiongli project graph view --project-id <prj_id> --save
qiongli project graph view --project-id <prj_id> --open

The first command saves a new private HTML snapshot; the second also asks the operating system to open it. Search and filter records, follow directed links, inspect gaps and copy source commands tied to the exported revision. The page works offline without a server. Export again after changes: it is a snapshot, not a live editor. Exported research may be private; filtering the view does not remove other records from the exported snapshot.

This complements literature citation discovery. It does not automatically understand arbitrary PDFs, prove a claim correct or replace the model's reading. Try the reproducible Graph example, including its unsupported claim, to see the intended behavior.

Stage summaries preserve substantial findings, source links, earlier stages and what changed. They help a later session resume without treating a summary as new evidence. Any retention review lists individual files for the user to select and delete personally; summarizing never silently removes the originals.

Collaborate through the Host you already use

Qiongli describes bounded tasks, source/candidate revisions and how to return findings to a coordinator. Actual subagents come from the Host. A second role in the same conversation is self-review, not independent review. Cross-Host work uses an available authorized communication tool or a packet that you transfer. There is no new background agent scheduler or automatic cross-device service.

The coordinator checks returned evidence and revision identity before integrating changes. Existing preview, approval and compare-and-swap checks still control canonical project writes. Delegation does not grant another agent permission to overwrite research files.

Move from 1.x without losing your way back

1. Keep a backup of research files, configuration and the previous install version. A new executable does not make a data migration reversible.
2. Choose one CLI installation channel. Use type -a qiongli ql on Unix or Get-Command qiongli,ql -All in PowerShell to see which command runs. npm, pip and Cargo do not overwrite one another's installations. qiongli setup offers manual cleanup guidance; you decide which files or packages to remove.
3. Install 2.0.0, check qiongli --version, then run qiongli install plugin. Keep the registered source directory when refreshing a Plugin.
4. Open a new Host session and check the tools. Try a copied project before moving important work. Old shell commands are not universally compatible; use the 2.x command guide and the retained 1.x reference.
5. For rollback, use the saved archive or reinstall the exact previous package, then restore its matching Plugin through the supported Host flow. Research files stay in place; restore a data backup separately if a migration changed them.

Stable npm installation uses qiongli@latest; next can remain on the earlier Beta. PyPI and Cargo can pin 2.0.0. upgrade cli explains the original channel's update command; it does not run a package manager or convert installations.

What the release checks establish

The release pipeline builds and installs the same source on all three supported targets, verifies versions and hashes, exercises extracted binaries with an empty PATH, and checks packaged Lite/Full MCP and registry installs. Cargo source publication has its own checks. Public asset verification belongs to the release workflow; submitting that workflow is not evidence that publication finished.

Rust removes the Python runtime from the native execution path, and shared owners reduce duplicated implementation. These are architectural improvements, not a measured speedup or maintenance-cost result. Full 1.x migration acceptance, live Hook delivery, real browser interactions, all Host/model combinations and cross-Host research journeys retain their own evidence requirements. The retained Desktop and 1.x sources are not retired by this CLI release.