One changelog for all your AI dev tools.

26 Aug

• Compliance API now covers Claude Science and Microsoft 365 sessions↗

  The Compliance API can now retrieve session transcripts from Claude Science and Microsoft 365 integrations, expanding audit coverage for enterprise deployments.

• Admin API available in CLI and all major SDKs↗

  The Admin API is no longer curl-only; it's now available in all major SDKs and the CLI, covering organization and workspace management, identity federation, and encryption key administration.

26 Aug

• Sub-agents now fall back to parent session's model on first-call 404↗

  Sub-agents that fail on their first model call now fall back to the parent session's model chain instead of dying. The error returned to the parent is also richer: it includes error type, status, request ID, and which model failed.

• Claude can draft feedback reports when sessions fail↗

  When something breaks in a session, Claude now drafts a structured feedback report you can review and send. Turn it off with the `feedbackDrafts` setting if you prefer.

• Prevent session wedge from hook or background agent error spam↗

  Fixed a crash where a hook or background agent dumping huge error output could overflow the conversation buffer and lock the session. Error output is now bounded.

• Profile and optimize your Claude API spend per project↗

  New `/claude-api cost-optimize` command profiles your existing project's Claude API costs and walks you through optimization levers one measured change at a time: caching, token hygiene, batching, effort level, and model selection.

25 Aug

• Auto mode classifier rules now editable in permissions interface↗

  You can now view and edit auto mode classifier rules directly in the `/permissions` interface instead of managing them elsewhere.

• Bash wildcard permission rules now warn about matching options before subcommands↗

  Claude Code now warns at startup if your Bash permission rules use wildcards before the subcommand, since they unintentionally match options placed before it.

• Background sessions now start reliably after sleep, deletion, or slow startup↗

  Background sessions now reliably start even if your machine slept, the working directory was deleted, or another Claude Code process is reinstalling npm at the same time.

• Transcript rendering fixed for very long single lines and fullscreen resizing↗

  Fullscreen mode no longer goes blank after resizing the terminal, and very long lines in diffs (like base64 strings) now render truncated instead of causing slowdown.

26 Aug

• Fix stream cancellation before inner stream registration↗

  Cancelling a stitched stream before its inner stream is set up no longer causes errors.

• Validate persisted tool calls against current schemas↗

  When you load a chat with tool calls from history, the SDK now validates them against your current tool definitions. If a tool has changed or been removed, incompatible calls are treated as dynamic tool parts rather than exposing unvalidated data under static types.

• Chats continue after tool approval denials↗

  If you have a tool approval flow that denies a tool call, the conversation can now proceed automatically instead of getting stuck waiting for manual intervention.

• Add operation-level outcomes to message stream callbacks↗

  When a message stream ends, your callback now gets operation-level outcome information, giving you more detail about what happened during the message processing.

• Tool approval reasons now flow through approval requests↗

  When you require manual approval for tool calls, you can now attach a reason explaining why, and that reason flows through your approval workflow to human reviewers—separate from their response reason.

25 Aug

• Batch completion webhooks for text generation↗

  You can now receive webhooks when batch text completions finish, enabling async workflows without polling. Direct Anthropic and OpenAI providers warn if you pass the option.

• Structured output parsing without explicit finish reason↗

  Some providers return text without a finishReason field; structured output parsing now works in these cases instead of failing.

26 Aug

• Security checks run in CI and agents can read and apply fixes automatically↗

  The CLI command outputs structured JSON in CI environments, letting agents read failures, apply fixes like enabling 2FA or replacing credentials with OIDC, and re-check automatically.

• Gemini 3.5 Transcribe model available on AI Gateway↗

  Google's Gemini 3.5 Transcribe is now available on Vercel's AI Gateway, supporting both streaming transcription over WebSocket and batch processing. The model auto-detects language across 85+ languages and accepts custom vocabulary.

• Security Dashboard flags misconfigurations across all projects and accounts↗

  New Security Dashboard available on all plans shows misconfigurations across your entire account in one place. Findings link directly to settings to fix them, and results export to CSV for reporting.

• GLM 5.3 Flash model available on AI Gateway↗

  Vercel's AI Gateway now supports Alibaba's GLM 5.3 Flash model, a multimodal option with 1M token context and function calling for inference routing and cost tracking.

• Python projects can define routing rules at the CDN layer↗

  Python projects can now use Vercel's routing rules to rewrite requests and set headers at the CDN layer before they reach your app. Rules publish immediately across all regions without a new deployment and can be rolled back from the History tab.

• Muse Image model available on AI Gateway↗

  Meta's Muse Image, a unified text-to-image and image-editing model, is now available through Vercel's AI Gateway. Call it with generateImage for creation or editImage for modifications, optionally passing reference images to guide generation.

• Qwen 3.8 Flash model available on AI Gateway↗

  Qwen 3.8 Flash, optimized for coding and agent workflows, is now accessible through Vercel's AI Gateway with a 1M-token context window and 65k-token response limit.

25 Aug

• Vercel Image Optimization protected from Next.js AVIF vulnerability↗

  Vercel disabled AVIF optimization on its managed service after a critical vulnerability in libheif was disclosed. Hosted apps are protected automatically; self-hosted Next.js apps must upgrade.

• Speed Insights free tier now works across all projects and plans↗

  Speed Insights free tier is no longer locked to Hobby or single-project limits—it works on any plan and any number of projects, with 10,000 events per team per month. Paid tier renamed Speed Insights Plus but unchanged otherwise.

• Run SDK sandboxes agent-generated code with host function boundaries↗

  Agents can now safely run generated TypeScript programs against your services without exposing secrets or internal clients. Host functions let you control what code can do and pause execution for human approval or authentication, resuming without repeating completed work.

• Interrupted agent runs resume without replaying completed work↗

  If an agent's code hits a sensitive operation and you interrupt it for human approval, you can resume the run later with a token. Already-completed function calls don't run again—only the interrupted function and anything after it.

• Vercel Connect exits beta with 100+ preset connectors and custom OAuth support↗

  Vercel Connect is now generally available with 100+ connectors for common services and support for custom OAuth and MCP servers, plus per-environment attachment and one-command revocation.

• Agent programs can combine multiple operations before returning results↗

  Agents can write programs that fetch from multiple services, filter or combine results locally, and return only the useful output. This reduces the amount of data fed back into the model's context, making workflows more efficient.

• Agents can request scoped credentials at runtime instead of storing long-lived tokens↗

  Agents now request short-lived, task-scoped credentials at runtime instead of storing long-lived tokens, eliminating the need to rotate secrets or risk leaks from a single compromised token.

• Webhook events flow to agents without storing signing secrets↗

  Agents receive webhook events from providers like Slack without storing signing secrets or bot tokens; Vercel Connect verifies and re-attests each event server-side.

• Audit logs and RBAC let teams control and track connector access↗

  Teams can now restrict who creates and manages connectors, and audit logs track all authorization and connector activity across projects for compliance.

• Vercel Connect adds managed Linq connector for iMessage, RCS, and SMS↗

  Apps and agents can now send and receive messages over iMessage, RCS, and SMS through a managed Linq connector. Vercel handles account setup, credentials, and token exchange; the channel maintains session state across messages and supports read receipts and typing indicators.

• Agents can join Notion comment threads via Chat SDK adapter↗

  The Chat SDK Notion adapter lets the same agent you run on Slack or Discord participate in Notion page comments without a separate codebase. Replies stay threaded; mentions, editing, and file attachments work; keyword or all-comment triggers available where @-mentions aren't supported.

• Webhooks work without stored secrets via signature verification↗

  Provider webhooks are verified server-side and re-attested using OIDC identity before forwarding to your project, so you don't store webhook secrets and protection stays enabled.

• Chat SDK Slack adapter supports Enterprise Grid org-wide installations↗

  The Slack adapter now supports Enterprise Grid org-wide installations, storing by enterprise ID and routing events correctly across workspaces while keeping tenant data isolated.

25 Aug

• Patch critical remote code execution on Windows-hosted servers↗

  A vulnerability affecting Windows-hosted Next.js applications using both the Pages Router and App Router without Cache Components can lead to unauthenticated remote code execution. Linux and macOS are unaffected. Patch immediately; no workaround exists.

• Patch critical remote code execution in Image Optimization with AVIF↗

  A vulnerability in the libheif library used by sharp can allow unauthenticated remote code execution when Next.js optimizes attacker-controlled AVIF images. AVIF optimization is now disabled until upstream fixes the issue.

• Two critical security vulnerabilities patched in Next.js 16.3.3 and 15.5.24↗

  Next.js 16.3.3 and 15.5.24 patch two critical security issues. Upgrade as soon as available; full advisories will accompany release.