Repository navigation
Releases: jiangjunw15-web/uplus-mcp
Releases · jiangjunw15-web/uplus-mcp
Release list
v0.5.2 - Security & logic defect fixes + CI fix
What Changed
Five audited defects are closed in this patch release, with 18 regression tests guarding them.
Fixes
| # | Priority | Problem | Fix |
|---|---|---|---|
| 1 | High | Custom profile directory could be recursively deleted | _copy_profile now refuses when src/dst overlap or target lacks Local State |
| 2 | High | Password values echoed back inconsistently | fill(), snapshot, and batch extract all mask password fields as .... |
| 3 | High | Submit gate had two opposite bugs (empty list passed 0==0; answered-only items unbalanced) | Empty question list now refuses submit; previously-answered objectives counted in total |
| 4 | Medium | patrol(submit=True) had no effect |
auto_answer+submit now wired through rule_based_answer with full-coverage gate |
| 5 | Medium | Unknown course ID widened query to all courses | Unknown/expired course ID now raises UplusError instead of falling back |
Regression Tests
tests/test_fixes.py - 18 tests covering all five fixes (positive + negative cases), all passing.
CI Fix
pom.xml: replaced hardcoded0.5.0filenames with${project.version}so artifact paths follow the version bump automaticallypublish-packages.yml:rm -rf maven-package/artifactsbefore copy to prevent stale files from previous releases
Full Changelog: v0.5.0...v0.5.2
v0.5.0 - Cross-platform generic MCP install, per-agent AI onboarding guide
What's new
- Cross-platform, machine-agnostic install: the launcher probes the installed Microsoft Edge or Google Chrome on Windows, macOS and Linux, with a PATH fallback; no Windows-only executable path is hard-coded.
- Configurable browser engine, CDP port, dedicated profile directory and auto-launch switch, overridable through
UPLUS_BROWSER,UPLUS_CDP_PORTandUPLUS_PROFILE_DIR; the legacyauto_launch_edgeconfig field maps automatically onto the new generic settings. rename.md: a step-by-step playbook any MCP-compatible AI agent follows after installation (environment check, install, client registration, two-mode sign-in, patrol, objective-question answering, human takeover).- Login detection prefers the session token before page text, fixing a first-navigation hydration timing miss that could report a signed-in session as logged out.
- README, README.zh and
docs/mcp-clients.mdare generalized for Windows / macOS / Linux first-time users, with bare Python anduplus-mcpconsole-script registration options.
Tooling & quality
- 28 tools: 10
uplus_*business tools and 18browse_*generic browser tools. - 73 unit tests pass; a clean virtualenv install was verified; a real-platform read-only regression (courses, homework, experiments, attachment question classification) passed without answering or submitting anything.
Scope & integrity
- Objective questions (single choice, multiple choice, true/false, fill-in) can be filled and submitted after the model's answers are reviewed; subjective answers, programming, attachments, experiment reports and group deliverables are always handed back to the user.
- The server stores no passwords, cookies or session snapshots in the repository.
Getting started
Follow the Quick start in README.md, then register the stdio server in your MCP client; read rename.md for the AI onboarding script.