Claude Report — 2026-09-06

🚀 Anthropic Official

Function Hooks proposal — "make plugins 10x more powerful" (Sep 3, ongoing)

Anthropic engineer Alice Poteat opens a public RFC for a new hook type: instead of shell scripts reacting to preset events, plugins would write typed TypeScript functions that wrap Claude Code's own behavior Express/Koa-style, with a next() continuation model and a single $ capability object for auditable side effects (file I/O, network, UI rendering).

• Already shipped behind a flag in build 2.1.260, but unreleased pending feedback — Poteat states "the response from the community likely dictates whether this ships or not."
• Includes a full technical architecture doc and short demo videos attached to the issue.
• Drew 113 comments by Sep 6, making it one of the most active open discussions in the claude-code repo this week.

Formalizing Fermat's Last Theorem in Lean (Sep 4)

Anthropic Research reports that Claude — using an internal research model "roughly comparable to Claude Fable 5.1" — produced the first end-to-end, machine-checked formalization of Wiles' proof of Fermat's Last Theorem in the Lean theorem prover.

• Took "a little under two weeks" (11 days) and about six billion output tokens, coordinated via Prove2Me, an open platform that maintains a dependency graph of theorem statements across many parallel Claude agents.
• The formalization spans 13 million lines of Lean code and 30,300 proved theorems (29,500 used in the final proof) — over 5x the size of Mathlib, Lean's main community library.
• Lean verified the result from only its three standard axioms; a companion project formalized Vinogradov's Three Primes Theorem in three days using consumer Claude Max subscriptions.

Claude Code v2.1.263 (Sep 6)

Ships as a bug-fixes-and-reliability-improvements-only release, with no changelog detail beyond that line.

Claude Agent SDK TypeScript v0.3.263 (Sep 6)

Bumps to parity with Claude Code CLI v2.1.263 with no new API surface; the Python SDK remains stuck at v0.2.152 (bundling CLI 2.1.259, Sep 2), widening the gap between the two SDKs.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

tt-a1i/archify

Ships a fix rejecting non-HTML output file overwrites without restricting output directories (#322, Sep 6) and merges a fix that flushes large artifact receipts before the checker process exits (#311, Sep 5–6).

Unchanged in window: anthropics/skills (last commit Sep 3, already reported), agentskillexchange/skills (routine sync/regeneration commits only, no content change), K-Dense-AI/scientific-agent-skills (last push Sep 2, already reported), alirezarezvani/claude-skills (no change since Aug 26).

🤖 Agents & Subagents

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — ships v0.3.263 (Sep 6), a parity-only bump tracking Claude Code v2.1.263 with no new agent-facing behavior.

Unchanged in window: nicobailon/pi-subagents (still v0.65.1, released Sep 4, already reported), claude-agent-sdk-python (still v0.2.152), VoltAgent/awesome-claude-code-subagents (README-only churn, no content change), AMAP-ML/LongHorizon-Harness (no push since Aug 20).

🔗 MCPs & Integrations

oraios/serena

Fixes concurrent Serena instances contending for the same Kotlin language-server storage directory — a non-blocking file lock now gives a second concurrent instance its own storage instead of having its requests cancelled by the first (fixes #1966, Sep 6), following the already-reported SolidLSP position fix (Sep 4).

DeusData/codebase-memory-mcp

Hardens semantic search: rejects non-string elements in array queries instead of silently dropping them, and makes a failing vector scan fail closed with an explicit error instead of rendering it as zero matches (Sep 5–6).

Unchanged in window: upstash/context7 (last commit Sep 4, already reported), github/github-mcp-server (last commit Sep 3, already reported), modelcontextprotocol/typescript-sdk (last commit Sep 3, already reported), modelcontextprotocol/servers (routine memory-server hardening Sep 3, no major new feature), punkpeye/awesome-mcp-servers (last commit Sep 1).

💡 Community — Workflows & Ideas

The 1930 poetry book that Anthropic tried to censor

A writer transcribing Stanley Kunitz's public-domain 1930 collection Intellectual Things into text repeatedly hit Claude's content filter, which flagged the dense literary use of words like "body," "death," and "blood" as unsafe — a concrete case of automated safety filters misfiring on canonical, rights-clear text (Sep 4).

📰 Quick Mentions

• Anthropic IPO launch shifts toward mid-October (Sep 5) — Reuters reports the listing timeline has slipped from late September.
• Authors wrangle with publishers over $1.5B Anthropic settlement (Sep 5) — a dispute over how the copyright payout gets split.
• reverify (active Sep 5) — an MCP server and CLI that checks AI-generated claims against deterministic tools instead of trusting the model, past 900 stars.
• okf-agent-memory (Sep 5) — a Git-native, embedded-MCP memory server implementing Google's OKF v0.2 spec with sub-300µs BM25 search, no external database.
• codenotch (Sep 5) — a macOS app that pins Claude Code, Cursor, Codex, and Antigravity usage limits to a screen edge.
• Show HN: claude-transplant (Sep 6) — moves Claude Code sessions between accounts inside Claude Desktop.
• Show HN: Stemma (Sep 6) — generates one source of truth for CLAUDE.md, AGENTS.md, and Copilot instructions.
• claude-adhd (Sep 5) — a Claude Code plugin that surfaces gentle reminders for unfinished threads from past sessions.