Your stack
Models
2 pickedAI coding tools
1 pickedAI frameworks
1 pickedAuth & data
1 pickedInfra & queues
1 pickedCloud platforms
4Frontend & UI
1 pickedLanguages & runtimes
7Observability & evals
2Section order
drag to arrange- OpenAI1
- Anthropic2
- Claude Code3
- Vercel AI SDK4
- Vercel5
- Next.js6
- Drizzle ORM7
Verbosity
How often
Your edition
60 items this week
16 Nov – 22 Nov
OpenAI
5 updates
19 Nov
GPT-Live-1.1 improves multilingual voice conversation quality
GPT-Live-1.1 improves pronunciation, accent consistency, and transcription for multilingual voice conversations. Drop the model name into your existing Live session configuration.
Live API sessions now support standby mode to reduce inference costs
Live API sessions can now pause inference when nobody is speaking and resume automatically when someone speaks, reducing compute costs during inactive periods.
17 Nov
Responses API changes reasoning preservation and instructions handling for GPT-6.2
GPT-6.2 changes how the Responses API preserves reasoning when you edit a conversation prefix and how it processes the top-level `instructions` field. Check the migration guide for required updates.
Background completion windows cut model costs by 30–40%
The Responses API beta adds background completion windows: choose two hours for 30% savings or eight hours for 40% savings on model tokens, useful for non-urgent inference.
16 Nov
CSAM image classifier available to approved organizations
Approved organizations can now call the CSAM Image Classifier API to screen images for child safety risks, returning separate signals for real and AI-generated material at no cost.
Anthropic
1 update
17 Nov
Customers can pay for Claude usage through their own organization
New linked billing feature (beta) lets your customers pay for Claude API usage through their own organization account instead of your billing. Set `billing_policy` per request; customers manage spend limits and revocation in Claude Console.
Claude Code
20 updates
20 Nov
Organizations can enforce settings and cloud providers through sign-in
Organizations managed in Anthropic's Enterprise Admin Console can now enforce settings and route Claude Code through their own cloud provider or gateway when users sign in.
Fixed conversation freezing after compacting very long chats
Claude Code no longer freezes for minutes after compacting a long conversation history.
Built-in chart tool draws graphs in terminal and MCP results
Claude Code now includes a built-in chart tool called easel that renders graphs in the terminal and in MCP metrics results on supported plans and platforms; turn it off in config if needed.
Self-hosted runner can start work in a repository subdirectory
Self-hosted runners can now start Claude's work in a subdirectory of a repository via `--start-dir` or environment variable, with settings and hooks loading as usual.
19 Nov
MCP server connection errors now shown in /mcp
When an MCP server rejects a connection, the `/mcp` page now displays the error message it sent, making it easier to diagnose connection problems.
Session tasks, loops, goals and forks now visible in /tasks
The `/tasks` command now shows scheduled tasks, loop wakeups, the session goal, and conversation branches. You can stop any of them individually by pressing `x` and confirming.
Trust dialog now shows where folder hooks are declared
When Claude Code asks you to trust a folder's hooks, the dialog now tells you where those hooks are defined, making it easier to verify they're safe.
Code Mode now works with OpenTelemetry tracing enabled
You can now enable Code Mode alongside OpenTelemetry instrumentation. Each code tool call will emit its events, though tracing itself remains disabled during execution.
MCP tools with complex schemas now work on third-party providers
MCP tools with complex input schemas or validation errors were silently dropped on third-party providers and gateways. They're now handled: complex schemas work, invalid ones are excluded and Claude is told why.
Deferred tool calls can now be blocked from resuming
Hooks can now defer tool execution. The new flag lets hosts prevent those deferred calls from running when a session resumes, giving more control over what executes in headless runs.
Code Mode scripts can now use top-level await with return statements
Code Mode scripts that combined top-level `await` (waiting for async work) with `return` (sending a result back) were incorrectly rejected. This restriction is removed.
Session metadata and labels for self-hosted runner scheduling
Self-hosted runners can now label sessions with metadata and hook labels for custom scheduling logic. New orchestrator metrics break down session queue wait and startup failures by cause.
17 Nov
Status command lists broken files and configuration errors
The `/status` command now shows a detailed list of configuration problems: broken files, keybinding errors, sandbox issues, permission overrides and hooks that never match. Helps diagnose why the agent is not behaving as expected.
Confirm before installing GitHub app to a repository
The `/install-github-app` command now asks for confirmation before modifying a repository, reducing the risk of unintended changes during agent runs.
Read MCP server JSON from stdin to keep tokens off command line
You can now pipe MCP server JSON to `claude mcp add-json` via stdin, keeping authentication tokens out of command-line arguments and shell history. This only works outside a Claude Code session.
Session billing group attribution via managed settings
Sessions can now be tagged with a billing group name via the `attributionProject` setting, enabling cost attribution in Claude apps gateways that support `admin.billing_group_tag`.
16 Nov
MCP server loading now explains why it stops mid-session
When an admin adds managed MCP configuration during a session, Claude Code now explains why existing servers stop loading instead of silently failing.
Code Mode now works with agent teams enabled
Code Mode, which gives the agent deeper code control, now works alongside agent teams. Sessions with teams already enabled will start in Code Mode by default.
Settings file errors now warn instead of silencing failures
When `/permissions` rules, tool settings, or `/add-dir` folders fail to persist because the settings file can't be written, Claude Code now warns you and names the file instead of silently ignoring the error.
Tool permission rules now validate on startup
Claude Code now warns at startup if your tool permission rules are malformed, catching configuration errors before they affect a session.
Vercel AI SDK
10 updates
20 Nov
Gateway mints Live-bound client secrets for Live models
A new getToken() method on the realtime gateway lets you mint short-lived credentials for Live model connections.
Gateway now routes Live models over WebSocket with normalized protocol
The Gateway can now handle OpenAI's Live models over WebSocket with a normalized protocol instead of rejecting them. Use gateway.experimental_realtime() for gpt-live-1/1.1, or pass { api: 'live' } for other Live model IDs.
Client-side WebSocket connections to Live models no longer require an application relay
If a Live model supports client-secret-websocket, you can now connect directly from the browser or client to the model using a short-lived token, removing the need to relay through your application server.
Messages array copied per step to distinguish tool calls
When prepareStep returned the same messages array for multiple steps, tool approval functions couldn't tell which tool calls belonged to which response. Messages are now copied per step.
Tool approval signature validation prevents replay of denied requests
With experimental_toolApprovalSecret, denied tool calls were being signed, allowing a client controlling message history to replay the request with an approved response. Automatic denials are no longer signed.
Preliminary tool results no longer marked as final in streamText
Preliminary results like OpenAI image generation previews were being recorded as final tool results. The preliminary flag is now forwarded correctly in streamText.
18 Nov
Steering can now interrupt model calls before output starts
Steering now supports an interrupt option, letting you cancel model calls before they generate output—useful for stopping expensive or unwanted inferences mid-request.
17 Nov
Fix streamText tool handling in transform factories
If you use custom transforms with `streamText`, they now consistently receive a tools array (empty when none provided) rather than undefined, fixing type mismatches.
Support output parameter in batch operations
You can now specify structured output schemas in batch calls, with results parsed according to your output definition.
Reduce memory usage when validating UI messages
Validating UI messages with complex nested metadata no longer consumes multiple times the memory of the data itself, fixing crashes on large requests.
Vercel
20 updates
20 Nov
Vercel Functions support request bodies up to 20MiB in opt-in beta
Opt-in beta raises request body limit to 20MiB for serverless functions on Node.js, Python, and Bun (not Edge). Requires 4GB memory setting and an environment variable. Routing Middleware won't see request bodies when enabled, so body-dependent middleware logic breaks.
19 Nov
Agents can search Vercel CLI commands by task description
Coding agents can now search Vercel CLI commands by describing what they want to do rather than guessing command names. The search handles synonyms and typos, and agents can fetch full help text in one call with `--with-help`.
Billing CLI commands now grouped under vercel billing prefix
All billing operations now use a `vercel billing` prefix for consistency; old command names remain as aliases. JSON output available for automation.
Page Integrity monitors and controls client-side scripts without JavaScript
Page Integrity uses Content Security Policy to discover, review, and block unapproved client-side scripts and connections server-side, with alerts when new resources appear.
Firewall rules now match on HTTP response status and headers
Response-based conditions let you apply rate limits and Firewall rules only to specific HTTP responses—useful for counting failed logins (401) or missing pages (404) without affecting successful requests.
Priority Projects adds bot signals and traffic-source labels to Firewall rules
Priority Projects exposes bot categories, verification status, and traffic sources (Tor, VPN, proxies, abuse lists) as Firewall rule conditions and request headers, letting you build smarter traffic filtering without client-side code.
Priority Projects increases per-project Firewall rule and alert limits
Priority Projects increases per-project limits for Firewall rules, custom rules, and alert configurations for teams that need more complex traffic policies.
Liquid AI decision models now available on AI Gateway
Liquid AI's d1.2 and d1.2-ultra decision models are now accessible via AI Gateway. These models return structured answers (Boolean yes/no, category selection, or rubric scores) instead of text, and now support audio input and bounding box detection for image analysis.
GPT-Live 1.1 voice model now available on AI Gateway
OpenAI's GPT-Live 1.1 supports simultaneous listening and speaking, letting callers interrupt or add detail mid-reply. It integrates with AI Gateway's existing Live endpoint, so you only need to update the model ID in your integration.
18 Nov
Vercel Board coordinates work across agent fleets and teams
New experimental primitive for coordinating work across multiple agents and humans. Board acts as a shared task store with channels, threads, and labels; agents can claim tasks, post updates, and sync with Slack, GitHub, and other tools. Includes watchers to trigger agents on events and scheduling for follow-ups.
Vercel Board enables agents to coordinate tasks on a shared message board
Vercel Board is a message board for agents to share context and coordinate work. Agents access it through the CLI or JavaScript SDK, with granular permissions and two-way sync to Slack and GitHub via Vercel Connect.
Vercel Drives reads files up to 6x faster with predictive prefetching
Vercel Drives now uses predictive prefetching to load data before your code requests it, reducing cold-read latency. Automatic for all drives, no SDK changes required.
Gemini 4 Neon model available on AI Gateway
Google's Gemini 4 Neon—faster than earlier Gemini 4 with comparable output quality—is now accessible through Vercel's AI Gateway, supporting text, images, video, documents, and grounding via Search and Maps.
17 Nov
Vercel Blob adds automatic deletion with configurable TTLs
Vercel Blob now supports time-to-live deletion for temporary data like session transcripts and exports. Set TTL on write, modify later with setTtl(), and track costs per 60-day period in Observability.
Native integrations for Slack, Teams, Discord, and web chat with tenant isolation
eve now ships native integrations for eight chat platforms, plus a Chat SDK adapter for others. Slack gets multiplayer sessions and ambient listening. Multi-tenant support lets one agent serve multiple customers with separate credentials and approval rules per tenant.
Agent observability rebuilt on OpenTelemetry with privacy controls
eve's tracing moved to OpenTelemetry, letting you use your existing observability tools. You control what gets collected and where—redact sensitive data per destination, and private conversations send only metadata by default.
eve framework reaches general availability with production-ready agent infrastructure
eve, an open-source agent framework, exits preview with durable execution, multiplayer sessions, and approval controls. You choose the model, tools, and approval rules; the framework manages the loop, persistence, and infrastructure. Deploy on Vercel or elsewhere.
Median agent turn latency drops to 414ms on Vercel through Workflows optimization
Vercel Workflows optimized orchestration and I/O; eve agents on Vercel now see 86% lower latency per turn. The framework also caches session routing and defers tool/skill listings from initial prompts.
16 Nov
Measure agent task success with evals framework
Agents now need the same testing rigor as traditional software. Evals let you measure success rates and efficiency across multiple trials, catching failures that logs miss—like ordering the wrong coffee. Harbor framework and Vercel integration make concurrent eval runs practical.
Vercel Sandbox Python SDK reaches 1.0 with stable API
Vercel Sandbox Python SDK stabilizes at 1.0, matching TypeScript capabilities. Use it to run AI-generated code in isolated sandboxes from Python apps; includes async support and automatic cleanup.
Next.js
1 update
17 Nov
AI-powered security review runs on every pull request
An AI agent reviews every pull request for security flaws by attempting attacks, catching issues that code review and CI miss before they ship.
Drizzle ORM
3 updates
20 Nov
Cast codecs removed; raw queries now respect driver configuration
Drizzle stops adding type casts to SQL, fixing invalid queries on several databases. Raw queries now decode output using the same codecs as mapped queries, making behavior consistent.
SQLite integer codecs accept number, bigint, and text
SQLite integer columns can now receive values in multiple formats without depending on driver-specific safe-integer settings, making code more portable across SQLite drivers.
Output decoding moved from drivers to Drizzle codecs
Drizzle now handles all output decoding through its codec system instead of relying on driver configuration. If you use custom type parsers or safe-integer modes, you must migrate them to `refineCodecs()` with `normalize` and `normalizeParam` functions.
A tool that ships more than twenty items in one edition has the rest follow in the next rather than being dropped.
This is what shipped each week, shaped by the same code that builds the email. A real edition also carries items an earlier one held back, skips anything you have already been sent, and never reaches back past the week you subscribed.