GitHub Top 5: Thaw tidies your Mac menu bar, 12.5K stars

Every week I go through the repos that gained the most GitHub stars in seven days and pick the top five. This week, Oct 26 to Nov 1, 2026, two of the five teach your coding agent to stop producing the same AI-looking pages, one audits the agents themselves, and one is a learning workspace with a memory you can read. Number one has nothing to do with AI at all: it cleans up your Mac's menu bar.

The video above counts them down in about ten minutes. Below is the written version, with how each repo works, how to set it up, and what to watch out for. Stars this week come from gittrend.io's weekly ranking; totals are as of Nov 3.

#5 iFixAi: an independent auditor for AI agents

iFixAi calls itself the independent auditor for AI agents, and it asks one question: is your agent doing what it's supposed to do? It puts your agent in a simulation of your business and sends it the requests a real day brings, including the ones that push it past its role. A judge then grades every reply, and every finding shows what was asked and what your agent did.

You can see it work in a minute with no API key. This runs a built-in demo agent with canned replies and canned grades, on Python 3.10 or newer:

``bash pip install ifixai ifixai author reference -o reference.yaml ifixai validate reference.yaml ifixai run --provider mock --eval-mode self --fixture reference.yaml ``

Any agent with a chat endpoint can be audited. An OpenAI-compatible endpoint works as it is, and anything else takes a few lines of Python. Any model can be the judge, and iFixAi never picks your agent's own provider. The recommended route is your coding agent: install the plugin or the skill, then ask it to run iFixAi on your agent. It finds the agent, builds the simulation, shows the cost before anything is billed, runs the audit and explains the results.

A default run has 10 inspections, including whether a role only uses its allowed tools, whether the agent can be talked into more access, whether it resists prompt injection, and whether it reports a failure instead of pretending it worked. --full-run covers all 61. You get a pass or fail per inspection with the evidence, plus an A to F grade from five pillars: fabrication, manipulation, deception, unpredictability and opacity. Failing any of three mandatory minimums caps the grade at D.

Two things to know: use --dry-run first to see how many tests and judge calls a run makes, and point it at a staging copy, because the probes are real requests. Apache-2.0 licensed.

#4 Hallmark: a design skill that refuses to look AI-generated

Hallmark is a design skill for Claude Code, Codex, Cursor and OpenCode, made by Together AI. It lints a page in a real browser against the tells that read as machine-made: the eyebrow above every heading, tracked-capital labels, one colored word in a headline, a hero that misses its screen, an email sign-up at the end of every page.

The gates come in three tiers:

• Floor: never ships. Contrast, focus, reduced motion, invented facts, a broken phone layout.
• Reflex: the defaults a model falls into. A build with a reason may overrule one on the record.
• Finish: craft detail that never blocks a ship.

It has fifty themes, and each one is a material: paper, type, color, surfaces, and its own components and texture. A theme never says what a page contains, so two prompts in one theme make two different pages. Pages are built from the job: a booking page books, a shop sells, a newsletter subscribes, and the page ends on that action.

A build runs in four steps. Pre-flight reads the project and asks once for your real material. It deals three themes that fit the prompt. The hero is built alone and reviewed cold. Then the page is checked at eight widths and a subagent reviews it with fresh eyes. Point it at a URL or a dev server and it captures your product's screenshots itself. It also lists 755 openly licensed photos, credited on the page, and avoids stock sites, whose terms bar AI tools from picking their images.

For a page you already have, name it: /hallmark redesign ./src/app keeps the content and the wiring and fails if a link, field or handler goes missing. /hallmark audit ./index.html only reports.

``bash npx skills add nutlope/hallmark -g ``

It needs Node 18 or later, and the browser checks also need puppeteer-core and Chrome. MIT licensed.

#3 DeepTutor: a learning workspace with a memory you can read

DeepTutor is an agent-native learning workspace that connects tutoring, research, creation and practice. Version 1.7.1 shipped on Nov 1. One runtime runs Chat, Quiz, Research, Visualize, Solve, Mastery Path and Reading, and they share session context.

Learning has four kinds of activity:

• Book: an interactive book built from selected sources. You review a chapter outline before content is generated, chapters are typed blocks (quizzes, flash cards, timelines, code, figures, animations), and each has its own chat.
• Mastery Path: a learning goal broken into topics, with mastery evidence.
• Immersive Reading: work beside a document, web page, video or audio source, with source-linked citations.
• Practice: answers, mistakes, recall ratings and the next review date.

Knowledge bases offer a choice of retrieval engines, including LlamaIndex, PageIndex, GraphRAG and LightRAG, or a linked Obsidian vault. deeptutor kb eval scores how well one retrieves against your own questions.

Memory is deliberately not a hidden vector store. It is files you can read, in three layers: L1 traces, L2 curated facts and L3 synthesis across surfaces, with a graph showing the evidence links. Partners are persistent companions with their own memory that can live in channels such as Telegram, Slack, Discord, WhatsApp and Microsoft Teams. The README calls a partner "a chat that has a personality and a phone number." You can also connect a live agent such as Claude Code, Codex or Hermes Agent and consult it from a chat turn, and import past conversations from ChatGPT, Claude Code and Codex.

The fastest trial is Docker:

``bash docker run --rm --name deeptutor \ -p 127.0.0.1:3782:3782 \ -v deeptutor-data:/app/data \ ghcr.io/hkuds/deeptutor:latest ``

Or install the full Web app and CLI (Python 3.11 to 3.14, Node 20.9+) with python -m pip install -U "deeptutor[full]", then deeptutor init and deeptutor start. Either way, you configure your own provider and model in Settings. Apache-2.0 licensed.

#2 Impeccable: the design vocabulary for your coding agent

Impeccable gives your agent words for design: steer with polish and bolder, review with critique and audit, and catch AI slop before it ships. It works with Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI and many more. It started from Anthropic's frontend-design skill.

The README lists what shows up when you ask an agent for a page with no other guidance: Inter for everything, purple-to-blue gradients, cards nested in cards, gray text on colored backgrounds, a rounded icon tile above every heading. Its demo gives the same model the same brief, a fourteen-room guesthouse on the Ligurian coast. Without Impeccable it builds a cream page, a thin serif and pill buttons. With it, the page is built around one picture: the house at night, seen from the water.

``bash npx impeccable install ``

Reload your tool, then run /impeccable init once per project. It reads your code, asks only what it can't infer, and writes PRODUCT.md, which every later command reads. Then name a command and a target, such as /impeccable critique the onboarding flow.

What you get:

• 24 commands to steer (bolder, quieter, distill, typeset), review (critique, audit), finish (polish, harden, clarify) and plan (shape).
• A reviewer. critique scores the design against usability heuristics and ranks what to fix first. audit checks accessibility, performance, theming and responsive behavior, with findings ranked P0 to P3.
• A slop detector. 62 deterministic rules, no LLM and no API key. It runs as a hook on every UI file your agent edits, in your terminal or CI with npx impeccable detect, or as a Chrome extension on any live page.
• Live mode. /impeccable live opens your running dev server with an overlay. Select an element, pick a direction, and it hot-swaps variants in place. Accept one and it is written back to your source.

Two things to read first. In Claude Code, installed hooks run even when you deny the model's own commands, so review them before unattended runs. And in live mode, Apply can fall back to running a CLI agent with approvals and sandboxing off; set IMPECCABLE_LIVE_COPY_AGENT=chat to keep it in your chat session. Telemetry is anonymous and can be turned off with npx impeccable telemetry off; your project name, files, code and PRODUCT.md are never sent. Apache-2.0 licensed.

#1 Thaw: the open source menu bar manager for the Mac

Thaw hides the menu bar items you don't need, brings them back on demand, and finds any of them in a keystroke. It is free, and free to keep: there is no subscription, no account and no tracking, and a new macOS never costs an upgrade. The code is GPL-3.0, written in Swift.

Beyond hiding items, you can organize them into groups, switch between profiles, customize your layout and keep distractions away with Zen Mode. Every action is also a thaw:// link, so Raycast, Floe, Shortcuts or a shell script can drive it.

``bash brew install thaw brew install thaw@beta ``

Or download the disk image from the releases and drag Thaw to Applications. Thaw 2 runs on macOS 15 to 26, and Thaw 3 is for macOS 27. Thaw 3.0.0-beta.5 came out on Nov 2, and the repo lists 75 contributors.

The README carries three supply chain badges (an OpenSSF Scorecard, OpenSSF Best Practices Gold and SLSA Build Level 3) and links a guide to verifying releases. If you want to help, the repo lists ways to contribute, pull requests go against the development branch, and translations go through Crowdin.

That's the top five for this week. Tell me in the comments which one you're trying first.