Manifesto
Today's issue is about the receipt. A proof that arrives with a file you can run. An agent that has to pass your own test before it says it finished. A reviewer handed your team's rules instead of generic advice. And a token, being taken away, because it let a publish skip its second factor.
Ephemeris.
Ten stuck problems, and a certificate for each one.
An internal version of OpenAI's Astra model produced proofs for ten problems in mathematics and theoretical computer science that had seen no progress on the main result for at least a decade — among them the existence of non-sofic groups, a disproof of Connes's rigidity conjecture, and sphere-packing bounds pushed down to the Cohn–Elkies threshold. Each result ships with a Lean 4 formalisation in a public repo, so the argument can be machine-checked rather than taken on trust. What is missing from the write-up is the denominator: how many problems were attempted, and what the prompts were.
- Results
- 10
- Fields
- 8
- Checked in
- Lean 4
- Token cost
- < $2,000 ea.
- Human role
- Formatting
- Failures shown
- None
Make it run your check before it says done.
A verification loop is the cycle where an agent runs tests, linters or custom checks against its own output and fixes what fails before moving on. Claude Code ships a /verify skill and hooks into your toolchain, but the part worth your afternoon is converting the checks you still run by hand into skills — a markdown file under .claude/skills/ with frontmatter and instructions written the way you'd hand them to a new teammate. The examples are the project-specific rules a generic linter will never know: every error log carries a request ID, and no log ever contains a request body.
Standalone
Invoked deliberately, for cross-cutting passes — a security sweep, an accessibility audit.
Embedded
Appended to a producing skill's instructions, so it fires as part of the work itself.
Chained
One skill calls the next on completion. End-to-end, and it runs without you.
PR-wide
The same verification runs against every pull request, not only the ones you remember.
Your reviewer can finally read your standards.
Copilot code review now supports agent skills and MCP servers on every plan — Pro, Pro+, Business and Enterprise. Put a SKILL.md in a subdirectory of .github/skills and the reviewer argues from your team's conventions instead of general advice; point it at an MCP server and it pulls context from the issue tracker, docs site or service catalogue you already run. Every MCP call made during a review is read-only, and the GitHub and Playwright servers are on by default. Anyone who wired this up during the public preview changes nothing.
- Skill file
- .github/skills/<name>/SKILL.md
- MCP setup
- Repo settings → Copilot → MCP servers
- Credentials
- Settings → Secrets and variables → Agents
- Tool scope
- Read-only
- On by default
- GitHub, Playwright
The token that skipped 2FA just lost its rights.
npm granular access tokens set to bypass two-factor authentication can no longer create or delete tokens, change package access, maintainers or trusted-publishing config, or manage organisation and team membership. Those operations now demand an interactive 2FA challenge. Publishing still works from such a token — until January 2027, when direct publish goes too. If your CI publishes with a bypass token, the migration is to trusted publishing over OIDC, or to staged publishing, and you have five months. GitHub PATs, GitHub App tokens and GITHUB_TOKEN in Actions are untouched.
One CLI, and the agent stops paying to read the web.
Agent-Reach gives an agent a single command line for reading and searching Twitter/X, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu, without a platform API bill behind any of them. Each site gets a first-choice backend and a fallback — twitter-cli, then OpenCLI, then bird — layered over yt-dlp, Jina Reader and Exa, with an agent-reach doctor command that reports which paths actually work on your machine right now. It sits at 64,100 stars and was last updated on 25 July.
No login needed
- Web pages
- YouTube
- RSS
- GitHub
- Bilibili
Login required
- Twitter / X
- XiaoHongShu
That's Sunday.
Five picks for 2 August 2026, drawn from eighteen sources read this morning.
Fed today's issue
OpenAI, Claude Code, GitHub Changelog, and @TochkiNadAI on Telegram.
Read, nothing taken
The Batch, The Zvi, Import AI, Jeff Bullas, Anthropic, Cloudflare, Sentry, Vercel, PostHog, fly.io, Next.js, and four more Telegram channels.
Dropped as repeats
Hugging Face's forensic timeline of the agent intrusion — that story ran here on 20 July. A talk claiming 80% of Claude Code's prompt was deleted, which ran on 27 July from Anthropic's own write-up. Vercel's adoption of the 2026-07-28 MCP specification, after GitHub's ran on 23 July.
Dropped as too old
Anthropic on containing agents (25 May), fly.io on agent sandboxes (8 June), and the Olmo 3 energy accounting paper (1 May) — all past the thirty-day cap, the last one surfaced by a newsletter that made it look current.
Note
sentry.engineering now redirects to blog.sentry.io/engineering. Nothing there was newer than 23 July.
The rubric
Tools worth adopting this week, creative software, dev tools and agentic coding, privacy and security when the signal is high, research with a practical kernel. Maximum five, and no reprints.