Key Signals

• GitHub's Copilot SDK reaches its final beta before GA, introducing a Rust SDK and remote/cloud sessions. Version 1.0.0-beta.8 is the last beta release before the planned general availability in approximately one week. The addition of a Rust SDK completes multi-language coverage, while remote sessions let developers share agent sessions via URL and cloud sessions offload execution entirely to GitHub's infrastructure. A comprehensive API review introduced breaking changes across all SDKs to establish the final 1.0 surface, including a new preMcpToolCall hook for intercepting MCP tool invocations. [1]

• Google is migrating Pro, Ultra, and free-tier developers from the open-source Gemini CLI to the closed-source Antigravity CLI, effective June 18. Announced at Google I/O, Antigravity CLI is positioned as a "premier agent-first development platform" with server-side orchestration and multi-agent capabilities. However, early adopters report drastically lower token quotas — some hitting limits in 6–7 prompts — and the platform lacks feature parity with Gemini CLI. Only enterprise users and those with API keys retain access to the original Gemini CLI. [2]

• A critical vulnerability dubbed "BadHost" (CVE-2026-48710) in Starlette threatens millions of MCP servers and AI agent infrastructure. A single character injected into the HTTP Host header bypasses path-based authorization in Starlette (325M weekly downloads), the routing core underlying FastAPI, vLLM, LiteLLM, and most OpenAI-shim proxies. Because MCP servers store credentials for external systems, the vulnerability creates a direct path to credential theft. The fix is available in Starlette 1.0.1. [3]

• OpenRouter's $113M Series B at a $1.3B valuation signals the multi-model future is already the production reality for AI agents. The AI gateway now processes 100 trillion tokens per month across 400+ models — a 5x increase in just six months — serving 8 million global users. The raise, led by Alphabet's CapitalG, suggests enterprises are actively avoiding model-vendor lock-in and treating AI models as swappable inference engines, a pattern accelerated by the shift from training to agentic workloads. [4]

• GitHub Copilot Memory gains CLI-native controls and a repository-level off switch, expanding organizational governance. The /memory command in Copilot CLI persists across sessions, while the new repo-level switch lets admins disable memory entirely for specific repositories. The store_memory permission prompt now explicitly distinguishes user-level preferences from repository-level facts, addressing a key transparency gap for teams. [5]

• OpenAI Codex 0.134.0 unifies profile management and adds conversation history search. The --profile flag is now the single entry point for profile selection across CLI, TUI permissions, and sandbox flows, with legacy profile configs rejected through migration guidance. Read-only MCP tools can now execute concurrently when they advertise readOnlyHint, improving throughput for multi-tool agent workflows. [6]

AI Coding News

• GPT-5.5 now leads Google's Android Bench as the top AI model for building Android applications. The May 18 leaderboard update dethroned the previous co-leaders (Gemini 3.1 Pro and GPT 5.4), while also adding open-weight models and new columns for latency, tokens, and cost. Android Bench evaluates models against real-world issues and pull requests from public GitHub Android repositories, testing challenges ranging from breaking-change resolution to Jetpack Compose migrations. Zencoder CEO Andrew Filev praised the initiative but cautioned that "public repositories leak into training" and private benchmarks can shift rankings by up to 26 percentage points. [7]

• A new governance framework called AC/DC proposes four stages — Guide, Generate, Verify, Solve — for managing AI coding agents at scale. The framework argues that verification, not generation speed, is the key differentiator for teams adopting agentic development. A controlled study by Sonar found that agents working in higher-quality codebases used 7% fewer input tokens, 8% fewer output tokens, and re-read files 34% less often — suggesting code quality is becoming an AI infrastructure efficiency variable. The framework positions "Guide" as the first layer of control, emphasizing that stronger models actually increase the need for explicit architectural boundaries. [8]

• GitHub introduced targeted model rules allowing enterprise owners to control which Copilot models are available per organization. This public preview feature moves beyond the single enterprise-wide model setting, enabling fine-grained rules that allow specific models for specific organizations. A refreshed interface lets admins set each model's availability to "Enabled" or "Optional". This addresses a common enterprise governance need as the number of available Copilot models continues to grow. [9]

Feature Update

• Copilot SDK v1.0.0-beta.8 ships the Rust SDK, remote/cloud sessions, and a comprehensive breaking-changes API review ahead of imminent GA. The Rust SDK bundles the Copilot runtime by default — add the crate and build. Remote sessions produce shareable URLs for monitoring agents from GitHub web/mobile, while cloud sessions offload execution entirely. The new preMcpToolCall hook lets applications inspect, modify, or block MCP tool calls before invocation. All SDKs gained simplified RuntimeConnection config, plain-string send(), typed PermissionDecision factories, and getMessages() renamed to getEvents(). Provider model overrides now support modelId, wireModel, maxInputTokens, and maxOutputTokens for BYOK users. [1]

• Copilot CLI v1.0.55-1 improves theme contrast and extension visibility. Selection background contrast increases across all color themes, and /env now shows loaded extensions with their status and source. Bug fixes address terminal bell sounding on turn completion, blank rows in the /resume picker, a stuck "Cancelling" state when Task tool agents are running, and runaway vote_memory tool call bursts. CJK and supplementary-plane Unicode clipboard support is fixed on Windows. [10]

• GitHub Copilot Memory adds deletion guidance, repository-level off switch, and the /memory command in Copilot CLI. When asking Copilot to forget something, it now directs users to the correct deletion interface and down-votes the memory. Repository admins can disable Copilot Memory entirely for their repo from settings. The /memory on|off|show command in the CLI persists across sessions, and the store_memory prompt now clarifies whether an entry is a user-level preference or a repository-level fact. [5]

• OpenAI Codex 0.134.0 adds conversation history search, unifies --profile, and enables concurrent read-only MCP tools. Case-insensitive search across local conversation history includes result previews. MCP setup gains per-server environment targeting and OAuth options for streamable HTTP servers. Connector tool schemas are more reliable through preserved local $ref/$defs structures and best-effort compaction of oversized schemas. Extension tools now receive conversation history context, and hook inputs include subagent identity. [6]

• Kiro is now a HIPAA eligible service and ships a v2.4.2 bugfix patch. Kiro's inclusion in the AWS Business Associate Addendum opens the door for healthcare organizations to use the AI coding tool while meeting compliance requirements. The v2.4.2 patch fixes a crash on Windows caused by koffi/createRequire in terminal input handling. [11]