Key Signals

• GitHub Copilot gains multimodal vision across all plans, marking its most significant perception upgrade since launch. Copilot can now reason about images and PDFs attached directly to chat prompts in VS Code, github.com, and the CLI. This removes the previous requirement for Copilot Business/Enterprise users to have the Editor Preview Features policy enabled—vision is now on by default for everyone from Free through Enterprise tiers. For developers, this means Copilot can now interpret screenshots of UI bugs, architecture diagrams, and design mockups as first-class context alongside code. [1]

• Kimi K2.7 Code becomes the first open-weight model selectable in GitHub Copilot, signaling a shift toward model marketplace diversity. Hosted by GitHub on Azure and billed at provider list pricing, this gives developers a lower-cost alternative without leaving the Copilot ecosystem. The model is rolling out to Pro, Pro+, and Max plans first; Business and Enterprise admins must explicitly opt in. This precedent—offering open-weight models alongside proprietary ones—suggests Copilot is evolving into a model-agnostic platform rather than a single-model product. [2]

• Claude Code v2.1.198 advances autonomous coding workflows by making background agents commit, push, and open draft PRs automatically. Previously, background agents would stop and ask for permission when code work was complete; now they carry the task through to a PR without manual intervention. The release also makes Claude in Chrome generally available, adds background agent notification hooks, and inherits extended thinking configuration into subagents for improved quality on delegated tasks. [3]

• The Copilot SDK v1.0.5 introduces MCP OAuth host token handlers and per-session AI credit limits, establishing critical infrastructure for enterprise agent deployment. The new onMcpAuthRequest callback enables SDK applications to handle OAuth challenges from MCP servers transparently, supporting initial auth, refresh, and upscope flows. Combined with the new sessionLimits option for capping per-session spend, these features address two key enterprise blockers: secure tool connectivity and cost governance. The Java SDK also gains annotation-based tool registration, significantly reducing boilerplate. [4][5]

• OpenClaw, Anthropic, and OpenAI are converging on a persistent-agent architecture where mobile devices become authenticated endpoints rather than execution environments. OpenClaw's new iOS/Android apps connect to agents running elsewhere rather than running AI locally, mirroring patterns in Claude Cowork with Dispatch and OpenAI Codex. This architectural convergence suggests the industry has found a model that solves real engineering constraints—agents that persist independently while phones serve as notification surfaces, action approval points, and sensor inputs. [6]

• GitHub is consolidating its AI platform: Models fully retiring July 30, while enterprise governance matures with managed-settings.json GA and auto model selection defaults. The GitHub Models playground, catalog, inference API, and BYOK endpoints will cease on July 30 (brownouts on July 16 and 23), directing developers to Azure AI Foundry or Copilot itself. Simultaneously, enterprise managed-settings.json reaches GA, letting admins define AI governance standards enforced across VS Code and Copilot CLI. [7][8]

AI Coding News

• OpenClaw launched native iOS and Android apps that treat phones as remote controls for persistent AI agents, not as execution environments. The apps let users approve actions, receive notifications, share camera input, and converse with their agent—but all AI processing happens on a persistent runtime elsewhere. This mirrors Anthropic's Claude Cowork with Dispatch and OpenAI's Codex patterns, where long-running agents operate independently and can be checked from multiple clients. The convergence of multiple companies on this same architecture suggests it solves fundamental engineering problems around battery constraints, memory limits, and always-on availability. [6]

• Google's Gemini Spark agentic assistant launched on macOS, competing directly with Claude Desktop, Copilot, and OpenClaw for the desktop AI agent market. Available in beta for Google AI Ultra subscribers in the U.S., Spark can work with local files, sort documents, and turn invoices into spreadsheets via Google Workspace. New features include real-time topic tracking, integration with Google Tasks and Keep, third-party app connections, and upcoming custom Model Context Protocol support that will let users connect arbitrary apps into their agent. [9]

Feature Update

• GitHub Copilot vision is now generally available across all plan tiers without admin action. Users can attach JPEG, PNG, GIF, WebP images and PDF documents to chat prompts in VS Code, github.com Copilot Chat, and Copilot CLI. The feature was previously gated behind the Editor Preview Features policy for Business and Enterprise users—that requirement is now removed. Image and PDF attachments are retained approximately 24 hours for Business/Enterprise users. [1]

• Kimi K2.7 Code is generally available in GitHub Copilot as the first open-weight model in the model picker. The model is hosted by GitHub on Azure with usage-based billing at provider list pricing. Initial rollout targets Copilot Pro, Pro+, and Max in VS Code 1.127+, Visual Studio 17.14.6+, Copilot CLI, cloud agent, github.com, GitHub Mobile, JetBrains 1.9.1-251+, Xcode, and Eclipse. Business and Enterprise organizations must enable the policy explicitly—it's off by default for those tiers. [2]

• Claude Code v2.1.198 ships Chrome GA, automatic PR creation from background agents, and the /dataviz skill. Background agents in worktrees now commit, push, and open a draft PR when code work completes. The built-in Explore agent inherits the main session's model instead of running on Haiku. Other fixes include network drop retry with backoff, agent team recovery when teammates die on API errors, and background task completion states in web/desktop/VS Code panels. The /agents wizard is removed in favor of direct .claude/agents/ editing. [3]

• GitHub Copilot CLI v1.0.68 adds Kimi K2.7 Code model support and 30+ reliability improvements. The MCP config form now marks focused fields with a chevron indicator. IDE tools remain available during transient disconnects, returning a clear error while disconnected and auto-recovering on reconnect. Tab completion shows slash command aliases inline. Other fixes cover Thai/Devanagari rendering, sandbox path conflict detection, Windows extended-length paths, and custom agent tool filter inheritance in nested subagents. [10]

• GitHub Copilot SDK v1.0.5 introduces MCP OAuth host token handlers, session spending limits, and Java annotation-based tools. The onMcpAuthRequest callback handles 401 WWW-Authenticate challenges from MCP servers across all SDK languages. Three new session options—enableCitations, excludedBuiltinAgents, and sessionLimits—enable finer control. The Java SDK gains @CopilotTool and @CopilotToolParam annotations with a compile-time processor, plus ToolInvocation parameter injection for accessing session context without LLM exposure. [4]

• GitHub Copilot SDK v1.0.6-preview.0 adds experimental GitHub telemetry forwarding across all SDKs. A new optional onGitHubTelemetry callback in ClientOptions lets host applications receive per-session telemetry events. When registered, the client auto-opts sessions in via enableGitHubTelemetryForwarding and dispatches gitHubTelemetry.event notifications. The feature is intentionally hidden from IDE completion]in C#,#[doc]` in Rust). Companion releases landed simultaneously for Go, Rust, and Java SDKs. [11]

• AI credit session limits are now available in public preview for Copilot CLI and SDK. The /limits command in interactive sessions lets users view, set, or adjust limits mid-session without restarting. For noninteractive runs, --max-ai-credits bounds a single run. Usage tracking spans model calls, subagents, and background compaction. Session limits are a soft cap—a response already underway finishes before the agent stops. Requires CLI 1.0.66+ or SDK 1.0.5+. [5]

• Enterprise managed-settings.json reaches general availability for GitHub Enterprise Cloud customers. The copilot/managed-settings.json file in a .github-private repository defines governance standards enforced in VS Code and Copilot CLI. Supported keys include extraKnownMarketplaces, enabledPlugins, strictKnownMarketplaces, disableBypassPermissionsMode, and model. Configuration takes precedence over user-set values, is fetched on authentication, stored in memory, and refreshed hourly. [8]

• GitHub Models will be fully retired on July 30, 2026, affecting all customers including those with active usage. The playground, model catalog, inference API, and BYOK endpoints will become unavailable after that date. Brownouts are scheduled for July 16 and 23 to help teams prepare. GitHub recommends Azure AI Foundry for model catalog access and GitHub Copilot for AI-powered development workflows. [7]

• New C++ language server config skill for Copilot CLI automates compile_commands.json generation. The Microsoft C++ Language Server is now installable as a Copilot plugin and includes a generate compile commands skill for CMake, MSBuild, and custom build systems. Users can say "regenerate compile commands" or "load project" to configure their compilation database, with the language server auto-watching for changes. [12]

• Enterprises can now default to auto model selection in Copilot by setting model to auto in managed-settings.json. This makes Copilot auto model selection the default for new conversations across all licensed users, while still allowing per-conversation model switching. Requires VS Code 1.126+ and a configured source organization for enterprise governance. [13]

• Kiro ships IAM Roles for sandbox AWS access and standardized third-party service authorization. The sandbox can now assume a configured IAM role for short-lived AWS credentials that refresh automatically during tasks and are removed on completion. Third-party Powers now use a standard OAuth flow—authorize once in Agent settings, and tokens are encrypted at rest and never exposed to the agent. [14]

• Kiro introduces add-on credits for individual paid plans, replacing the previous overage model. Credits cost $0.04 each, come in packs from $5 (125 credits) to $100, allow up to 5 concurrent packs, expire after 12 months, and are consumed only after monthly plan credits are exhausted. The earliest-expiring pack is consumed first. [14]

• OpenCode v1.17.13 fixes reasoning mode for OpenAI-compatible deployments and stale GitHub Copilot response IDs. The reasoning mode fix ensures reasoning settings apply reliably on custom deployments. Desktop improvements include a searchable v2 model picker, session tab hover previews showing project/path/branch/server, streamlined WSL setup, and window-scoped tab restoration. [15]

• Gemini CLI v0.51.0-nightly.20260701 adds a Cloud Run webhook ingestion service and fixes at-reference path resolution. The caretaker feature gains a new webhook ingestion endpoint for Cloud Run, while the core tools fix resolves defensive path resolution for at-reference files and corrects macOS test failures. [16]

• OpenAI Codex v0.142.5 patches a security-sensitive trace logging issue for WebSocket requests. Full Responses WebSocket request payloads were being written to trace logs and are now suppressed. A parallel alpha release (0.143.0-alpha.32) also shipped without detailed user-facing changes. [17]