Claude Report — 2026-09-20

🚀 Anthropic Official

Projects redesigned: from folder to conversation (Sep 17)

Rebuilds Claude Code Projects around a single persistent conversation instead of a folder-plus-chat model: Claude scopes the goal, delegates work across parallel cloud-session threads (each its own branch/repo copy), reviews outputs, and assembles results, backed by shared memory and a library of files/artifacts. Beta starts with select Pro/Max users on cloud sessions with no existing projects, widening over the following week.

Claude Code v2.1.275 – v2.1.278 (Sep 17–19)

• v2.1.275 (Sep 17): Adds a send-now key (ctrl+enter) that interrupts the turn and flushes queued messages, syncs claude.ai account skills/plugins into terminal sessions (syncClaudeAiSkills/syncClaudeAiPlugins to opt out), adds /plugin install <plugin> --marketplace <source>, and fixes dozens of resume/crash/sandbox bugs.
• v2.1.276 (Sep 18): Same-day hotfix for a v2.1.275 regression that failed every request with 400 … Input tag 'advisor_20260301' behind a proxy/gateway.
• v2.1.277 (Sep 18): Adds AGENTS.md fallback support when no CLAUDE.md exists, frames subagent output so it can't masquerade as top-level instructions, strips invisible Unicode formatting from prompts, and closes 100+ other bugs.
• v2.1.278 (Sep 19): Defaults auto mode to the server-side classifier for Claude API/Enterprise, Bedrock, Vertex, Foundry and gateway users (no classifier-overhead billing); adds CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out and an Auto mode server row in /status.

claude-agent-sdk-typescript v0.3.275–v0.3.278 (Sep 17–19)

Fixes deferred tool-call result key names, getSessionMessages()/forkSession() dropping assistant and queued messages, and resumed/forked session cost totals; adds a builtin field on SlashCommand, pasted_content on SDKUserMessage, remote-session latency fields, and a 'userSettings' updateSettings() source, then syncs to Claude Code v2.1.278.

claude-agent-sdk-python v0.2.154–v0.2.156 (Sep 17–18)

Bundles the Claude CLI forward through v2.1.274, v2.1.275, then v2.1.276 across three same-week point releases, trailing the TypeScript SDK and CLI by roughly one to two versions.

Measuring the pace of AI development inside frontier labs (Sep 17)

Publishes three transparency metrics: Claude now "leads" ~26% of Anthropic's AI R&D (up from under 1% in February) with over 90% of tasks at "AI collaborates" level or above; roughly 30,000 research agents run concurrently, with online monitors blocking just 0.002% of agent decisions; about 6% of R&D compute goes to safety work (12% within AI-driven AI R&D specifically).

Introducing the Life Sciences Verification Program (Sep 17)

Launches a verification program giving vetted life-science organizations Standard or High-risk Use grants that unlock model capabilities blocked by default, covering drug discovery, research biology and clinical development for Team/Enterprise plans first.

Compliance API adds Claude in Chrome session transcripts (Sep 18)

Extends the Compliance API's local-session endpoints to return transcripts from Claude in Chrome sessions under a new claude_in_chrome product-surface value, in beta for Claude Enterprise orgs using the existing Compliance Access Key scope.

🔌 Claude Code Plugins

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

🛠️ Skills

eugeniughelbur/obsidian-second-brain

Merges a fix for coexistence with another vault plugin and names which SessionStart hook and schema takes priority when both are installed (Sep 18–20).

addyosmani/agent-skills

Hardens its skill-authoring/grading toolchain — adds authoring rules against model-specific workarounds, binds grader results to declared expectations by ID, fixes Stop-hook file expansion to prevent data loss, and portable-izes reference-link checks for Windows (Sep 15–18).

Unchanged in window: cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills, tt-a1i/archify, ayghri/i-have-adhd, affaan-m/ECC, agentskillexchange/skills.

🤖 Agents & Subagents

nicobailon/pi-subagents v0.70.0

Adds a subagents.defaultSubagentOnlyExtensions setting and per-agent frontmatter to restrict which agents children may launch, keeps background workflows visible through detaches/interruptions, and fixes 20+ workflow-state and cross-platform issues (Sep 20).

Unchanged in window: VoltAgent/awesome-claude-code-subagents, anthropics/commerce-agents (no commits since Aug 31).

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Fixes two data races the worker threads exposed in memory/spill handling, corrects TypeScript awaited-generic-call edge tracking and a Cypher unbound-variable count bug, adds a memwaste-sanitizer feature, and bumps CodeQL to v4.38.0 (Sep 19–20).

ChromeDevTools/chrome-devtools-mcp

Ships a chunked trace-buffer parser for large recordings (#2721), records a sanitized client-name field for telemetry (#2757), lets the CLI's upload_file accept multiple paths (#2668), and updates the SKILL.md to use get_css_styles (#2760) (Sep 17).

upstash/context7

Clarifies Bedrock regional setup guidance in its docs (Sep 17) — a minor doc-only fix after two weeks without a functional release.

Unchanged in window: oraios/serena (Sep 17–19 MCP SDK v2.2.0/LanguageBackendRegistry work already reported), modelcontextprotocol/ext-skills, github/github-mcp-server.

💡 Community — Workflows & Ideas

25 Prompt Templates Worth Stealing This Month

Organizes 25 reusable prompts around a Role–Context–Task–Format–Constraints structure, with 8 adapted from Anthropic's official Claude Code library (reviewing diffs, investigating production incidents) and 17 original templates for writing and decision-making, each annotated with why the pattern works (Sep 17).

📰 Quick Mentions

• Researchers used Claude to hack into OpenAI — Hacktron AI chained two vulnerabilities via a Claude Opus 5-built exploit pipeline to reach OpenAI's internal repo in under 72 hours, earning a $6,500 bug-bounty reward (Sep 18).
• Piebald-AI/claude-code-system-prompts — tracks Claude Code's system prompt swinging +6,153 tokens (v2.1.274), −1,237 (v2.1.275), +5,234 (v2.1.277), then unchanged through v2.1.278 (Sep 17–19).