Claude Report — 2026-09-25

🚀 Anthropic Official

Claude Code v2.1.282 (Sep 24)

A fix-heavy release on top of v2.1.281's gateway/MCP work: adds a maxProseWidth setting for wide terminals, a startup notice plus /status/claude doctor entries listing ignored or disabled telemetry variables, an allowClaudeInChromeWithManagedMcp managed setting, and a gateway readiness_grace_seconds option so /readyz survives a short Postgres failover. Fixes include a 400 error on conversations with undecryptable web-search history, resumed sessions (--continue/--resume) silently dropping earlier reasoning, a "redacted_thinking" API error now handled by dropping and retrying thinking blocks, compaction retrying on a fallback model when summarization is refused, and auto mode switching to a server-side classifier by default when telemetry is off. Also bundles a large batch of Claude Tag (Slack) fixes: correct auto-join on Enterprise Grid org-wide installs, threads on retired models now move to a working model instead of falling back every reply, and file caption formatting fixes.

claude-agent-sdk-typescript v0.3.282 (Sep 24)

Ships same-day parity with Claude Code v2.1.282: adds strictKnownMarketplaces/blockedMarketplaces support for host-managed settings, a new lightweight @anthropic-ai/claude-agent-sdk/core entry point for apps that bundle their own zod/MCP SDK, and alpha prewarm()/SpareProcess.claim() APIs to start a Claude Code process before its session/folder binding is known. Also fixes readMcpResource() leaking CLI-reserved _meta keys.

API changelog — refusal billing expansion (Sep 24)

Anthropic now bills pre-output refusals whose stop_details.category is "bio", "frontier_llm", or "reasoning_extraction" at the normal rate of the model that ran, since these categories show low false-positive rates; other pre-output refusal categories remain unbilled. Applies across all platforms (direct API, Bedrock, Vertex).

API changelog — cache diagnostics out of beta (Sep 23)

Cache diagnostics no longer requires the cache-diagnosis-2026-04-07 beta header; POST /v1/messages responses now always include a diagnostics field (null unless requested).

Claude Tag supports personal connectors in channels (Sep 24)

Claude in a shared Slack channel can now use an individual's own connectors (calendar, drive, CRM, staging deploys) instead of only admin-attached ones, with per-response review or an "auto" posting mode; rolling out on Team plans now, Enterprise to follow.

🔌 Claude Code Plugins

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

🛠️ Skills

addyosmani/agent-skills

Extends its skill validator to run on macOS and Windows, holds command frontmatter to the same YAML rules as skills, validates reference links inside skills' own reference files, and restores security rules dropped in an earlier condensation (Sep 22-23).

anthropics/skills

Updates the claude-api skill to link pre-output refusal billing to the new docs (Sep 24), its first commit since Sep 10.

Unchanged in window: cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills (last substantive commit Sep 13).

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships 8+ commits on Sep 24 alone: fixes async lifecycle paths, makes global agent discovery non-blocking at session startup, aligns dispatch-contract docs, fixes stale pi-ai state during tool activation, stops TUI widget flicker, and loads the council protocol via a guide topic.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Sep 21, README only), anthropics/commerce-agents (dormant since Aug 31, single commit).

🔗 MCPs & Integrations

oraios/serena

Fixes a duplicated "IS" string in the CLI's ignored-path check output (Sep 24), following up on the Sep 23 PyJWT CVE fix and Astro language-server support.

upstash/context7

Adds a Docs7 custom-domains documentation page and flattens the Docs7 sidebar (Sep 24), continuing the Sep 23 Search API docs work.

DeusData/codebase-memory-mcp

Merges a service-pattern-boundaries distillation and a further Windows endpoint directory-creation race-condition fix, distinct from the one shipped Sep 23 (Sep 24).

Unchanged in window: ChromeDevTools/chrome-devtools-mcp (still v1.10.1, Sep 23), github/github-mcp-server (last commit Sep 16), modelcontextprotocol/servers (last commit Sep 22).

💡 Community — Workflows & Ideas

Whiteboard (YC W26) — an open-source IDE for thoughtful software design

Plugs into Claude Code and Codex, giving the agent an SDK to draw specs and diagrams on an in-app canvas that link directly to the underlying code; built on CodeOSS, day's top HN post (233 pts, 88 comments) (Sep 24).

davila7/claude-code-templates

Adds a new "ai-release-hunter" skill (#985), improves the seo-specialist agent based on automated review (#984), and fixes the jev-model-router mod leaving a bare slash command unclassified (#983), all Sep 24.

jev-pilot

A new Claude Code plugin that has "Jev" auto-pick reasoning effort, subagent model, and skill per prompt, then learns from a decision ledger every 20 turns via /jev tune; shown on HN Sep 24.

4% of Sessions, 65% of the Bill: 30 Days of Claude Code Telemetry

An OTel-based analysis of real Claude Code usage finds cost concentrated in a small fraction of sessions, informing where teams should focus cost/risk controls (Sep 24).

Radix — visual UI for agentic programming

Lets an agent generate an interactive workspace widget for a task that persists locally on disk, shown on HN Sep 25.

📰 Quick Mentions

• Show HN: Critic — lets AI agents present and annotate their own code changes so reviewers can talk directly to the agent that wrote it (Sep 24).
• Launch HN: Canary (YC W26) — independently boots and tests the app a coding agent changed, pulling context from GitHub, Linear, Sentry, Datadog and Slack (Sep 24-25).
• public-browser v3.0.0 — new MCP browser-automation server for Claude Code/Cursor claims 33% fewer tokens and 24% fewer tool calls than agent-browser in a 30-test benchmark (Sep 24).