GitHub AI Weekly · Issue 15 (Nov 21, 2026) · AI Beyond
AI Beyond Editor2026 week 47
As of November 21 at 09:44 Beijing time, this issue selects 10 AI projects by weekly star gains on GitHub Trending, covering coding, agents, infrastructure and multimodal tools. It reviews updates, uses and limits, with deeper readings of AG-UI event ordering and FFmpeg Skill output boundaries.
Repository authors describe it as: Build your own AI office with Univer Office SDK. Bring editing, calculation, collaboration, and file processing into your product for people and AI agents. Built on an extensible open-source core.
Stars
30,952
Forks
2,826
Open issues
201
Last commit
2026-11-21
Univer released 1.5.2 on November 16; subsequent main-branch work fixes drawing undo and font handling. Its useful angle is giving agents an editable spreadsheet or document inside a product, rather than asking a model to return table-shaped text.
Developers with existing business data and permissions can start with one spreadsheet workflow and compare formulas, formatting and undo against the original file.
The open-source repository does not include the whole commercial SDK. Collaboration, file conversion and some AI extensions are separately distributed and licensed; the demos exceed the default open-source installation.
Repository authors describe it as: Keep your coding agent up to date with the latest web best practices
Stars
9,539
Forks
770
Open issues
384
Last commit
2026-11-20
This week Modern Web Guidance added a file limit and partial-coverage reporting to legacy-scan. Retrieving guidance for a particular problem also makes it possible to see which code has not been checked, instead of treating a large prompt as complete coverage.
Maintainers of older frontends can try native browser features on one dialog or form, retaining keyboard-operation tests before and after the change.
Newer platform features still need fallbacks for the browsers the product supports. A repository commit is not proof that the change has reached the installed package.
We examined Open Code Review in Issue 06. This week, version 1.13.12 fixes non-zero exits for partial operational failures, while 1.13.13 fixes context compression budgeting. Once review runs in CI, recognizing a failed review matters as much as presenting its findings.
Teams already using it in CI can simulate a file-read failure, check that the run is not recorded as a pass, and verify that findings retain complete location information.
Results depend on the model service, repository and rules. Author benchmarks are not measurements of this site or your code; comments need reproduction and do not replace human review of permissions and business logic.
Repository authors describe it as: An open source design system that's fully customizable and agent ready
Stars
31,356
Forks
2,457
Open issues
923
Last commit
2026-11-21
Astryx 0.7.17 shipped on November 21 Beijing time, adjusting swipe actions, keyboard occlusion and menu focus restoration. These interaction fixes show the work required when a design system meets real mobile use.
Teams on React 19 can try a mobile form on a real device, checking typing, scrolling and focus after dismissing the keyboard before migrating more components.
The project remains in Beta, and Lab components are not the same as published packages. Older React applications face migration costs; generated examples do not establish accessibility or touch acceptance.
Repository authors describe it as: AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
Stars
71,462
Forks
6,336
Open issues
29
Last commit
2026-11-20
Version 4.4.0 of last30days-skill fixes literal handling of scheduled topics, local persistence with save=false, and scopes GitHub research evidence to public repositories. Scheduled research makes input boundaries and stored artifacts worth inspecting alongside the summary.
Product researchers can start with a familiar topic, separating announcements, opinions and secondary reporting, then check whether missing sources are explicitly reported.
Platforms have different credentials, services and availability. Engagement is not factual reliability: follow claims back to primary sources, and do not treat market discussion or prediction odds as events that have happened.
Repository authors describe it as: AG-UI: the Agent-User Interaction Protocol. Bring Agents into Frontend Applications.
Stars
20,823
Forks
2,125
Open issues
599
Last commit
2026-11-20
The November 17 AG-UI TypeScript 1.0.5 release fixes message order and snapshot boundaries across overlapping streams. For applications running concurrent tools, matching each result to its task is essential to meaningful human intervention.
Developers connecting backend agents to a web or terminal app can test two interleaved tool calls and check that each result remains attached to its own task.
A shared event representation does not implement authentication, authorization or persistence for the application. Check each framework adapter and its supported versions separately from the core protocol version.
Deep dives
A simple chat UI can append text as it arrives. Concurrent tools, human approval and shared state make that approach misleading: parallel work becomes a seemingly linear story. Typed AG-UI events let a frontend represent messages, tool calls and state separately, reducing custom integration code between frameworks.
The 1.0.5 fixes illustrate the details. A mid-stream snapshot must not incorrectly overwrite later deltas during compaction, and interleaved tool results must retain their relationships. A chat screenshot will not reveal these failures. Deliberately delay one tool, introduce a snapshot, and check whether the rendered sequence still matches the work.
For an existing backend, a small replayable workflow is a useful trial. Record events and expected views, test disconnection, cancellation and duplicates, then assess adapter maintenance. The protocol supplies a shared language; the product still decides which actions require approval and which state must be undone after rejection.
Repository authors describe it as: The workspace to stay in the loop and on top of agents - shape, decide, plan, annotate, review code, share with team, and send all feedback to agents with one click.
Stars
13,906
Forks
1,390
Open issues
227
Last commit
2026-11-21
Plannotator 0.29.0 shipped on November 18 with a redesigned two-pane code review window, image diffs and semantic diffs. It addresses the practical problem of finding decisions that require a person when several agent tasks finish together.
Engineers reviewing agent plans or PRs can use a change containing renames and image updates to check whether feedback returns to the correct file and context.
A review surface organizes evidence; it does not establish correctness. Plugins, hooks and agent integrations need configuration, and reviewers should inspect what files are shared with their feedback.
Repository authors describe it as: Claude Code skills for equity investors and traders — market analysis, technical charting, economic calendars, screeners, and trading strategy development.
Stars
7,439
Forks
1,470
Open issues
54
Last commit
2026-11-20
Claude Trading Skills combines market review, journaling and risk workflows. The verified commits this week mainly update English documentation and evidence, with no GitHub release records. Attention to a workflow says nothing about trading performance.
Readers organizing their research can begin with the fictional examples and trace conclusions to inputs and records, rather than turn the output directly into orders.
Workflows use different data services, whose information may be incomplete or stale. The project disclaims profit promises and is not an automated execution platform. This review concerns software workflow, not investment selection.
Repository authors describe it as: Chrome DevTools for coding agents
Stars
58,544
Forks
4,884
Open issues
138
Last commit
2026-11-20
This week the Chrome DevTools MCP main branch fixed CLI parsing of negative numbers. The latest release, 1.13.0 from November 11, already includes logpoints and page-scoped network mocks. Keeping commit and release dates separate avoids assuming an installed package contains a newer fix.
Frontend engineers can reproduce a network failure with a test account, let the agent inspect console and request records, and compare the diagnosis with observed page behavior.
Browser contents are exposed to the client. Usage statistics are enabled by default and performance tools may query CrUX; inspect those settings and use an isolated test browser.
Repository authors describe it as: Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex. Cut, captions, reframe, loudness, platform exports. No cloud, no API keys.
Stars
8,870
Forks
812
Open issues
39
Last commit
2026-11-21
FFmpeg Skill 4.2.2 rejects output paths that expand into multiple files; 4.3.0 adds corrected commands after unambiguous calling errors. Output boundaries and recoverable mistakes both matter when an agent repeatedly edits footage.
Creators processing their own footage can copy a short clip, inspect cuts, captions, sound and export properties, then consider applying the workflow to more material.
Local execution requires FFmpeg, storage and processing time; whether the calling model client goes online is a separate matter. Automated checks cover specified metrics, not editorial correctness or protection against every unwanted cut.
Deep dives
Writing an FFmpeg command is easy compared with recovering from a wrong one after files have changed. The 4.2.2 notes describe a concrete boundary: playlist paths and numbered filename patterns can produce additional files that a check of the visible output path does not cover. The release rejects these outputs before execution, including dry runs.
Automation integrations need a contract for actual side effects, not just the path visible in the command line. Version 4.3.0 addresses the other end by suggesting corrected commands only when mistakes have a clear answer. That is different from repeatedly trying commands until the exit code is zero: restrictions and correction make the call sequence reviewable.
Try reframing and loudness processing on a copy, retaining the input, command, output and measurements. Then watch for captions covering the subject and cuts losing meaning. Correct dimensions and loudness do not establish editorial quality. Small acceptance tests help repetitive transformations; pacing and aesthetic decisions still need human viewing.
Signals this week
Last week focused on handing specifications and evidence to the next participant. This selection moves closer to execution: ordering concurrent messages, reporting a failed review, and identifying every file a video command can affect. Across frontends, review and editing, the common concern is leaving a record that a person can still understand and verify.
There are no dedicated RAG or model-inference entries in this selection. That does not establish declining importance; it reflects the AI subset of this Trending snapshot. A published fix also does not prove why a project gained stars. Start a trial with a small task whose correct outcome is known, preserving failures and original files so the conclusion remains checkable.