GitHub Trending Top 10: AnyPS5 tops the chart at two months old, claude-mem rebuilds memory search, and HyperFrames ships 23 releases in a week (Oct. 4–Oct. 11, 2026)

The board for October 4 to October 11 gave its top slot to a repository created two months ago and its eighth to a framework with 179,801 stars. A weekly chart measures attention during the week, and this week it weighted that far above a project's standing: the top entry gained 25,151 stars in seven days, 83% of everything it has ever collected, while the eighth gained 680.

Almost nothing on the board stood still. All ten took commits inside the window, and six cut a release. The four that did not still moved — 4,295 commits in the loudest case, eight in the quietest, and 38 in the one that has never published a release at all. Last week's chart split the same way, six and four, but two of last week's four silent repositories took no commits whatever.

Seven of the ten exist to give a coding agent something it cannot do on its own: memory, a test suite, a plugin format, a video renderer, a CAD kernel, a team of seats, a set of engineering instructions. The other three are two native systems projects written in C and C++, and Flutter.

The ranking below is GitHub's own weekly Trending view, captured on October 11, 2026 at 8:00 p.m. ET. "Stars this week" is the gain GitHub attributes to the trailing seven days; "stars captured" is the running total at that same moment. Both measure attention, so production readiness stays a separate question, and each entry closes on the adoption boundary that decides the star. 1

Current signal: Rank 1 with 25,151 stars this week and 30,158 captured, against 178 open issues and 406 open pull requests. GPL-2.0, commits through October 11, and 4,295 commits since October 4. No release fell inside the window: the newest is v0.1.1 from September 28, and the repository itself dates from August 3. 123

Problem solved: A PlayStation 5 game exists as an executable built for that console's hardware and its system libraries, and it runs nowhere else.

Implementation: AnyPS5 is C++, and its core is a relinker: it rewrites the executable into the target system's native format and links it against reimplementations of the console's system libraries. There is no emulator and no separate runtime process. Shaders are recompiled to SPIR-V and validated with SPIRV-Tools when the build enables it, and input runs through SDL, including analog sticks and triggers, mapped by an anyps5-input.ini file. The project is blunt about its failure mode: an unsupported state throws std::runtime_error, prints the message to stderr and ends the process. Its one verified game, the 2D platformer Dreaming Sarah, holds a stable 60 fps on a GTX 1050 Ti and an i5-7500. 2

Differentiation: An emulator reproduces the console and runs its code underneath. AnyPS5 rewrites the binary and links it against libraries written for the host, so the game ends up as an ordinary native process. The README frames the work as interoperability, research and preservation, and ships no firmware, keys or proprietary libraries. 2

Verdict: Star it to watch native binary translation move this fast. Three boundaries: GPL-2.0, a compatibility list that currently fits on one page, and a design that terminates rather than degrades the moment it meets an unsupported state. You supply the binaries and accept the legality question yourself. 4

Current signal: Rank 2 with 6,440 stars this week and 9,046 captured, against 30 open issues and 75 open pull requests. Apache-2.0, 125 commits inside the window, and thirteen package releases in two batches, on October 6 and October 9, led by e2e@0.19.0. 156

Problem solved: An end-to-end test has to restate a user's intent as selectors, waits and assertions, so the flows everybody understands are the slowest to write and the first to break.

Implementation: A test describes a goal in plain language and an agent drives the app, while locators and assertions check the result in the same file. What makes that affordable is the replay cache: an agent step that a later assertion verifies records its actions, and the next run replays them with no model calls until the app changes. A test with no agent step calls no model at all, and you bring your own key, subscription or local model. The web engine drives Chromium, Firefox and WebKit through Playwright, the mobile engine drives simulators and emulators, and @e2e-dev/smol branches a warm Chromium microVM for each attempt. Version 0.19.0 added locator.dblclick() as an alias of doubleTap(), so a suite ported from Playwright keeps its double clicks, and replaced the Playwright trace archive with a Markdown page per test naming the node each action landed on. 56

Differentiation: A selector framework records nothing about why a step exists. Here the agent writes the step once and the recording carries it afterwards, and the same runner covers a browser and a phone.

Verdict: Star it if you maintain a web suite and a mobile suite and want one host for both. Boundary: the replay cache only pays off while the app holds still, agent steps need a model you pay for, and the CLI sends anonymous usage data by default. 57

Current signal: Rank 3 with 10,330 stars this week and 286,227 captured — the largest total on the board — against 153 open issues and 14 open pull requests. MIT, 46 commits through October 9, and v1.3.0 and v1.3.1, both published on October 4. This repository last appeared here for July 19–26, at 189,603 stars. 189

Problem solved: An agent fails before it writes code: it misreads the request, lacks the project's vocabulary, or produces a change nobody can review.

Implementation: Plain instruction files, one per skill, copied into a project or installed as a managed plugin — a self-updating Claude Code plugin, a Codex marketplace entry, GitHub Copilot, Gemini CLI, or npx skills@latest add mattpocock/skills. A one-time setup command asks which issue tracker you use, which labels you triage with and where documentation should live, so generic instructions become local working agreements. The window's release graduated implement-spec into the Engineering bucket: it reads tickets as a task graph, runs implementer subagents in their own git worktrees across the ready frontier, lands everything on one integration branch, and closes with a code review. The goal is now the integration branch rather than a pull request. 89

Differentiation: GSD, BMAD and Spec-Kit own the whole process. These files stay small enough to edit in place when one of them is wrong for your team.

Verdict: Star it as a method you will adapt, and expect to prune. Boundaries: every skill runs inside your agent with your credentials, 153 open issues is the surface area of a corpus this size, and the project added 96,624 stars since its last appearance here. 10

Current signal: Rank 4 with 3,803 stars this week and 99,744 captured, against 51 open issues and 148 open pull requests. Apache-2.0, Node 20 or newer, 225 commits inside the window, and nine releases between October 4 and October 8, led by v13.35.0. 11112

Problem solved: An agent starts each session blank and pays again for context it had last week.

Implementation: Five lifecycle hooks capture a session, a local worker service managed by Bun serves a viewer and the search endpoints, SQLite stores sessions, observations and summaries, and a Chroma vector store answers hybrid semantic and keyword search. The window's headline release replaced the memory interface with one tool, mem_search, that discloses memory in bounded steps — a compact index, then context around the anchors the agent picks, then a single batch of details — with an automatic mode that runs all three stages in one call. Tool replies are now short readable text instead of raw JSON, save_memory writes explicit notes, and a hook bridge supplements Claude Code's and Codex's own memory lookups. 1112

Differentiation: Most memory layers embed a session and return the nearest matches. The bet here is the shape of the answer: a search the agent can stop early, which the README prices at roughly ten times fewer tokens than fetching details up front.

Verdict: Star it when the failure you are debugging is forgetting rather than reasoning. Boundaries: hooks and a background worker run inside your coding sessions, memory lands in a local SQLite file and a vector store on every host you install it on, and the compression itself costs model calls. 13

Current signal: Rank 5 with 794 stars this week — the second-smallest gain on the board — and 8,396 captured, against 273 open issues and 47 open pull requests. MIT, commits through October 7, and eight commits inside the window. No release fell inside it: the newest is v0.9.29-alpha from September 30. 11415

Problem solved: Debugging a huge C++ program on Windows runs into the debug information itself — PDB files that overflow their own 32-bit tables, and link times measured in minutes.

Implementation: A native, user-mode, multi-process graphical debugger, in alpha, for Windows x64 with PDBs. Beside it the project is building two things. RAD Debug Info is a debug format the debugger parses directly instead of PDB or DWARF, with a radbin utility that converts and dumps it. The RAD Linker produces x64 PE/COFF binaries and reports 50% faster link times on the project's own gigantic test cases, with large memory pages taking another 25%; its command line is compatible with MSVC. This week's eight commits were the work an alpha collects: a DWARF parser fix, a Linux window manager falling back to 60 Hz when it cannot read the display's refresh rate, and build-script fixes for older shells. 14

Differentiation: The README still says Windows x64 only, and the commits show a Linux port already under way. Against WinDbg and Visual Studio the claim is throughput on builds too large for the existing toolchain.

Verdict: Star it for the linker if you compile enormous C++ on Windows; the debugger is the part to watch. Boundaries: alpha, a Windows-only supported path today, 273 open issues, and a week whose unit of change was eight commits rather than a version. 16

Current signal: Rank 6 with 4,223 stars this week and 60,977 captured, against 80 open issues and 179 open pull requests. Apache-2.0, 259 commits inside the window, and 23 releases between October 4 and October 11 — the busiest release feed on the board — the newest v0.8.149 on October 11. Last covered here for September 27–October 4, at 56,774 stars. 11718

Problem solved: A generative video model renders a slightly different picture every time, which is the wrong tool when the output has to repeat exactly.

Implementation: HTML, CSS, media and seekable animations compiled into a deterministic MP4. Animations run through GSAP and are captured frame by frame under Puppeteer, with ffmpeg for the encode, so the same input yields the same file. It is used three ways — a CLI, a set of skills that coding agents load, and the rendering core beneath a hosted authoring app — and the README now shows a native macOS and Linux application. The window's last release exposed transcribe and removeBackground from hyperframes/api as importable functions with typed results and typed errors, so an app like HyperFrames Desktop no longer has to run the CLI as a child process. A release the day before made a stalled render name the step Chrome hung in. 1718

Differentiation: Against model-based video tools the claim is reproducibility: a template, a chart animation or a product cut that must ship identically tomorrow. Agents can also drive it without a GPU.

Verdict: Star it if you need the same video twice. Boundaries: 23 releases in seven days is a surface that moves under you, and the Studio authoring path reaches for HeyGen's own voice service when you use its media features. 19

Current signal: Rank 7 with 1,140 stars this week and 10,900 captured, against 68 open issues and 158 open pull requests. TypeScript, 38 commits through October 11, and no releases at all — the repository has never cut one, so commits are the only unit of change. It last appeared here for August 16–23. There is no license file at the repository root, and GitHub reports no license detected. 120

Problem solved: Every coding agent expects its own layout for plugins and skills, so a workflow written for one tool does not reach the others.

Implementation: The repository publishes the plugin specification and a catalog of official plugins. Each plugin is a standalone directory at the root with its own .cursor-plugin/plugin.json manifest. The catalog mixes first-party entries — a teaching plugin, a branch-review plugin called Thermos, an orchestration plugin that fans a large task across parallel cloud agents with planners, workers and verifiers, a Cursor SDK plugin, and a scaffolder for new plugins — with third-party ones contributed under their authors' names, among them a team kit for CI and code review, a plugin about designing CLIs that agents can run, and integrations for Gmail, Drive, Calendar, Docs, Sheets and Slides. 20

Differentiation: Anthropic's plugin directory is the model for Claude Code. Cursor's is a manifest format plus a market, and the third-party entries show the format is being adopted outside the vendor.

Verdict: Star it to read the manifest spec and to see what a plugin catalog looks like while it is still forming. Boundaries: no license file, no releases to date, and an entry's quality depends on the author who contributed it.

Current signal: Rank 8 with 680 stars this week — the smallest gain on the board — and 179,801 captured, against 12,705 open issues and 633 open pull requests. BSD-3-Clause, 117 commits inside the window, and 3.47.7, published on October 8, inside the window. 12122

Problem solved: Shipping one product to iOS, Android, the web and four desktop platforms usually means several codebases and several visual results.

Implementation: Flutter compiles one Dart codebase to ARM machine code, JavaScript and WebAssembly, and draws through the Skia and Impeller graphics libraries rather than through each platform's own widgets. It ships Cupertino and Material widget sets, offers stateful hot reload, and updates the stable channel quarterly with hotfixes in between. The window's release is one of those hotfixes: two iOS accessibility bugs, in which accessibility frames for an OverlayPortal were shrunk by the device pixel ratio and anchored to the top left, so VoiceOver focus and XCUITest taps missed the visible controls, and VoiceOver focus was incorrectly shifted to the top left of the screen inside a ReorderableListView. 2122

Differentiation: React Native maps to each platform's native components. Flutter draws every pixel itself, so one codebase looks the same everywhere and text input and platform-specific controls have to be re-implemented.

Verdict: This is the board's only mainstream framework, and 680 stars is a chart placement rather than news. Upgrade if your app shows either of the two iOS accessibility bugs. Beyond that, 12,705 open issues describes the cost of maintaining a framework at this scale. 23

Current signal: Rank 9 with 2,508 stars this week and 19,259 captured, against 1 open issue and 30 open pull requests. MIT, Python 3.11 or newer, 35 commits inside the window, and ten releases between October 4 and October 11, from v0.7.12 to v0.7.21. 12425

Problem solved: Getting a part out of a chat model is easy. Getting a file a machine shop will accept, with the checks that decide whether the part can be made at all, is the part that stalls.

Implementation: A plugin that gives an agent local workflows for generating 3D models as STEP, GLB, STL or 3MF files, running design-for-manufacturing checks, producing engineering drawings, and handing off to 3D printing, sheet-metal and CNC services. The geometry runs on build123d 0.11 over the Open CASCADE 7.9 kernel, installed through uv and published on PyPI as cadgen, with a local server, cadgen mcp, that gives the agent a viewer and exposes the same operations as MCP tools. The window's ten releases were last-mile repairs: a STEP tail crash, a race on the first browser snapshot, console windows flashing open on Windows workers, and a second round of build telemetry. 2425

Differentiation: Most text-to-3D tools stop at a mesh for rendering. This one targets the interchange format and the manufacturability checks, which is a narrower promise and a more useful one if you intend to build the part.

Verdict: Star it if your models have to be manufacturable rather than merely visible. Boundaries: telemetry is on by default and reports usage and crash data under a random ID, and the CAD runtime downloads on first start, so the first run needs a network connection. 2526

Current signal: Rank 10 with 2,028 stars this week and 6,977 captured, against 101 open issues and 60 open pull requests. Apache-2.0, 311 commits inside the window, and four releases between October 6 and October 10, v0.6.6 through v0.6.9. 12728

Problem solved: Once several coding agents are installed, each one is a separate terminal, and nothing carries the work, the context or the decisions between them.

Implementation: Node.js 22 or 24 plus tmux, on macOS or Linux, with Windows supported only through WSL2. You define the team — roles, shared context, owned work — in YAML and boot it with one command; a daemon runs underneath, and a terminal interface shows the rig first as a graph, then as a table of seats with runtime, model, context and state. Claude Code and Codex can occupy seats in the same rig. The window's releases worked on the first hour: one release made setup open the terminal interface and the kernel's operator in front of you, and 42 of that version's 120 changes came from ten outside contributors. 2728

Differentiation: Where a control plane like Paperclip wraps the agents in a service, OpenRig keeps them as the terminal processes they already are and makes the team itself a configuration file.

Verdict: Star it to study the team-as-configuration pattern. Boundaries: launching a rig writes provider hooks and workspace-trust settings on your machine, the daemon must be restarted after a CLI upgrade because it runs database migrations on first start, and the project's automated tests do not run on WSL2 yet. 29

The chart did two things at once this week. It ranked attention, which put a two-month-old binary translator on top with 25,151 stars against Flutter's 680, and it recorded cadence, which split the board six and four. Those four are worth reading together: AnyPS5, raddebugger and cursor/plugins published nothing, and mattpocock/skills' newest version landed twelve hours before the window opened. All four took commits anyway, from 4,295 down to eight.

The six that shipped worked mostly at the edge of the product. HyperFrames turned two media operations into importable functions so other programs could stop shelling out to its command line. OpenRig spent a release on the first hour a new user sees. E2E replaced its trace archive with a Markdown page per test. Claude-Mem redesigned search to disclose less at a time. Text-to-cad spent ten releases on its last mile, and Flutter fixed two iOS accessibility bugs. That is the work that decides whether a tool survives contact with a user.

Every verdict this week turned on the edge of the product. AnyPS5 terminates on an unsupported state and has one verified game. RAD Debugger is alpha, Windows-only, and spent its week on eight commits. Cursor's plugin catalog has no license at all. Text-to-cad's telemetry is on by default. OpenRig writes hooks and trust settings into your machine. Claude-Mem puts a worker and a database on every host. E2E sends anonymous usage data unless you turn it off. HyperFrames moves 23 times a week.

For a first experiment, pick the entry closest to a task you already have — the test suite, the memory, the video, the part — read its license, and test the boundary that applies to you. A star here records that the project moved this week. Whether it deserves to stay in your toolchain is a question about the edge, and the edge is the one thing a star count never measures.