Bookmarks

Top 10 GitHub Repos — November 4, 2026

Developer Tools

Spec Kit

GitHub's `specify` CLI drops a fixed sequence of slash commands into whichever coding agent you name at init: constitution, specify, plan, tasks, implement, then `converge`, which loops until it reports the code matches the spec. The newer extensions push the same pattern past greenfield work. `bug` runs assess, fix and test and ends with a verified, partial or failed verdict, and `assess` takes a raw idea through to go, needs-clarification or kill. GitHub's Open Source Friday stream on 2026-10-16 covered the 1.0 release, and the repo was pushed again on 2026-11-03. It has gone from a spec template to a framework you extend, and the extension catalog is unaudited, so read community code before installing it.

Thinkrail

It's JetBrains, but it isn't IntelliJ. Thinkrail is a small Bun-served IDE built around the `pi` coding agent, with one agent session per chat, each in its own git worktree and branch. It keeps `SPEC.md` files next to the code and gives the agent `spec_*` tools to read and update them. Non-read-only MCP calls need your approval, and PR checks, review state and merges go through your own `gh` login. The repo was created on 2026-07-08 and pushed minutes before this digest ran. A large IDE vendor shipping a deliberately small, agent-first editor with no plugin system yet is worth watching.

PR-Agent

The oldest repo here (created 2023-07-05) is trending because of who owns it now. Qodo has donated PR-Agent to a community org, named @naorpeled as its first outside maintainer, and says the project is moving to an open-source foundation. The tool works as before: `/describe`, `/review`, `/improve` and `/ask` each make one LLM call through LiteLLM, and it runs on GitHub, GitLab, Bitbucket, Azure DevOps and Gitea as a CLI, Action, container or webhook. A vendor-owned AI reviewer becoming community-owned matters more than another feature release would.

AI Agents

Superpowers

Jesse Vincent's skill set is opinionated in a way most skill packs aren't. The TDD skill deletes any code that was written before its test. Plans are broken into tasks of two to five minutes, each with exact file paths. Each task goes to a fresh subagent and gets two reviews, one for spec compliance and one for code quality. Recent additions include a `diagnosing-superpowers` skill that reads your session transcripts and packages a scrubbed bug report, plus hooks that re-inject the bootstrap skill after context compaction on harnesses that support it. Most popular skill packs offer tips. This one enforces a process, and it now supports a long list of harnesses, from Antigravity to Kimi Code.

OpenBot (CopilotKit)

CopilotKit's take on "AI coworkers" is mostly a permission system. Each bot gets its own container with Chromium and a workspace, but every browser, file, shell and MCP action passes through one gateway that checks a CEL policy and writes an audit row first. Deny rules are evaluated before allow rules, and a missing policy permits nothing. When a bot hits a 2FA prompt, a human takes over the browser and the bot is locked out until they hand it back. The README calls it a template rather than a product, and it's in beta. It's the clearest example in this week's feed of designing the audit trail before the agent.

Security

Cloudflare Security Audit Skill

The design assumes the model will make findings up. A skeptic agent tries to disprove each candidate, separate agents re-check every source claim, and anything they can't confirm goes into `NEEDS-VALIDATION.md` rather than the report. Findings are written to a schema-checked `findings.json`, coverage is tracked unit by unit in a ledger, and the skill won't execute target code unless an OS-enforced sandbox with networking off is in place. Cloudflare says it is the seed of its internal vulnerability-discovery harness. The repo has only 14 commits and was last pushed on 2026-09-14, but it is on this week's trending feed anyway. It treats an LLM security audit as untrusted output to be checked, and few audit tools do that.

AiSOC

AiSOC is a whole SOC stack: 91 connectors, about 3,000 executable detection rules, alert correlation into incidents, and LLM triage agents. Those agents call typed tools and can't write SQL. Any verdict that cites an indicator missing from the evidence gets sent back to a human. Prompts containing raw logs or secret-shaped strings are refused before they leave the box, playbooks run in preview until someone approves them, and a bundled llama3.2:3b means triage works without an API key. The version badge reads 22.0.0. It has the most guardrails of any agent project in this week's feed, in a field where a confident wrong answer does real damage.

CLI

Worktrunk

Running five agents on five branches means five worktrees, and git's own commands make that painful. Worktrunk is a Rust CLI that fixes the tedious parts. `wt switch` creates or jumps to a worktree by branch name, `wt merge` squashes, rebases, fast-forwards and cleans up in one step, and `wt switch pr:123` checks out a pull request. Smaller features show it was built by someone who actually works this way: copy-on-write copying of `target/` and `node_modules/` into new worktrees, and a `hash_port` filter so each worktree's dev server gets its own port. It handles the build caches, ports and cleanup that come with running agents in parallel.

AI Tools

Graphify

Graphify turns a repo into a knowledge graph without an embedding model or a vector store. Code is parsed locally with tree-sitter in about 40 languages, and an LLM is only used for docs, PDFs and media. Every edge is tagged as extracted or inferred, so you can tell what the parser saw from what the model guessed. It also turns `# WHY:` comments and ADR references into nodes. The project now lives under the Graphify-Labs org, carries a YC S26 badge, and works from a `v8` default branch with schema-versioned graph output. You can check where each edge came from, and that suits agents that need to navigate unfamiliar code.

MCP

FFmpeg Skill

Giving an agent FFmpeg usually means giving it a shell and hoping. This repo offers 42 standard-library Python scripts with typed argparse flags instead. Nothing goes through a shell and callers can't pass a raw filter graph. A generated JSON contract feeds an MCP server (12 tools by default, all 42 behind an env flag), `doctor` checks which encoders and filters your build has, and a saved `--plan` won't render if any input file has changed since it was made. Version 3.0.0 dropped FFmpeg 5.x and Python 3.9. It's a small, local tool that makes a dangerous command safe to hand to an agent.