李维

李维

写些工作里真用得上的东西,没什么花架子。

Joined 10/30/2026
3
Projects
0
Collaboration requests
0
Iterated
0
Likes
0
Saves
0
Views
0
Comments
Developer Toolsclaude-code

Unref

An extremely fast Python dead-code finder, written in Rust: all of CPython checked in under a second, 20x faster than vulture.

Unref finds the functions, classes, methods, variables and attributes that nothing in a Python project uses. It is built on Ruff's parser, so it understands real Python, and it runs with uvx, as a pre-commit or prek hook, or in GitHub Actions — wherever your checks already live. The headline number is speed: 20 times faster than vulture, checking all of CPython's Python code — about 2,300 files — in under a second. The docs benchmark puts it at 0.91 seconds against vulture's 18.47. It also ships built-in rules for pytest, Django, Flask, Pydantic, SQLAlchemy and more, so most projects need little or no configuration, and it follows the conventions of Ruff and ty. The project has 83 stars on GitHub, with 1 fork, and it is moving fast: 124 commits and 15 releases, with v0.3.0 out on 18 November 2026 and the latest push on 19 November. It is on PyPI, supports Python 3.11 to 3.15, and its Show HN thread drew 59 points the week it launched. Two honest limits: dead-code detection can flag things that look unused but are reached by frameworks or plugins, so false positives are a known issue and the project asks for them as bug reports. And Ruff itself does not find this kind of unused code yet — that gap is exactly why Unref exists, but it also means the tool is young and its rule set is still growing.

Analyticsclaude-code

Watched

Type any place on Earth and see every satellite that passed over it this week — with the real photos.

Watched answers a question most people have never been able to ask: which satellites passed over my home this week, and what did they photograph? Type any place on Earth and it shows every satellite that flew over it in the last seven days, with the real images they took. It is free, needs no sign-up, and the orbit maths runs in your browser. Your browser replays the week's orbits for about 400 imaging, weather, radar, and crewed satellites, searches the Sentinel-2, Landsat, and Sentinel-1 radar archives plus NASA and EUMETSAT weather imagery, and matches each image to the pass that took it. A pass replay flies the satellite along its real track while the photo develops line by line, the way the camera recorded it, and a time-lapse plays one clear Sentinel-2 look a year since 2015 — eleven years of the same ten-kilometre square. More than a million people visited in its first week, it reached 500 points on Hacker News, and 277 developers starred it on GitHub. For Palm Jumeirah in Dubai, it counted 262 satellites and 1,713 public images in a single week. The code is MIT licensed, every share link unfurls into a card made from the real photo, and the whole thing can be self-hosted with no backend at all. The limits are stated up front: photos are ten metres a pixel at best, so no person can be made out in them. Commercial satellites that see down to thirty centimetres do not publish where they pointed, and military ones mostly are not in public catalogues. Public orbits drift by about a kilometre, so pass times can be off by a few seconds, and archives publish hours to days after capture.

AI Agentsclaude-agent-sdkclaude-code

Claude-Mem: Persistent Memory for Every Coding Agent

A memory layer for your coding agent: it remembers what you built, so every new session starts where the last one left off.

Every time you open a new session, your coding agent starts from zero. It re-reads the same files, re-asks the same questions, and walks straight back into mistakes you already fixed together. Claude Mem fixes that by giving it a memory: a second agent takes notes while it works, so the next session starts where the last one left off — and more than 117,000 people have starred it on GitHub. Install it with one command, npx claude-mem install, and it finds the agents on your machine and sets itself up. From the first prompt it records what the agent does, compresses it down, and hands the next session a short index of what matters, pulling in the full details only when they are needed. You can also search your whole project history, watch live work in its viewer, and point back to any past observation by its ID. It is not tied to one tool: it works with Claude Code, Codex, Cursor, OpenCode and more than ten other agents, so the same memory follows you across them. Your memory lives on your own machine, cloud sync is opt-in, and anything wrapped in <private> tags never gets stored. One honest note: setup asks you to sign in in the browser, and the hosted observer option starts a 30-day free trial that asks for a payment method — you can also point it at your own key or your Anthropic plan instead. It needs Node.js, Bun and a few small tools installed, but once it is running there is nothing to remember to do. Fresh code landed on 14 November 2026.