Key Signals

• Copilot CLI ships two feature-packed releases with a new /chronicle command and major UX improvements. Version 0.0.419 introduces the experimental /chronicle command for generating standups and tips from session history, new keyboard shortcuts, a --mouse/--no-mouse flag, and npm-style MCP server naming with dots, slashes, and @ characters. Version 0.0.420 follows up by extending auto-update to cover binary executables and adding automatic 502 retry logic. Together, these releases meaningfully improve both the daily workflow ergonomics and the MCP integration surface for Copilot CLI users. [1][2]

• Copilot SDK v0.1.28 introduces breaking security defaults that require immediate attention from SDK integrators. The release changes the permission model to deny all permissions by default and now requires a permission handler on session creation — a significant shift for anyone building Copilot-powered agents or extensions. It also ships agent selection and session compaction APIs, permission checks for custom tools, an OpenTelemetry GenAI setup guide, and bumps minimum Python to 3.11. Developers upgrading will need to update their integration code to explicitly handle permissions. [3][4]

• Gemini CLI delivers a massive v0.31.0 stable release alongside a v0.32.0 preview with sub-agent classification and a generalist agent. The v0.31.0 release is one of the largest Gemini CLI updates to date, adding parallel function calling for read-only tools, MCP progress updates, session-based SDK architecture, Gemini 3.1 Pro Preview model support, A2A API key authentication, project-level policies, and a /extensions explore command. The v0.32.0-preview.0 follows with an experimental Gemma Router using LiteRT-LM, Kind.Agent for sub-agent classification, and adaptive plan complexity. This positions Gemini CLI as an increasingly mature competitor in the AI coding CLI space. [5][6]

• GitHub Copilot metrics reaches general availability with new enterprise CLI telemetry. Copilot usage metrics is now GA, providing dashboards and APIs at enterprise, organization, and individual user levels covering code completion activity, IDE usage, model breakdowns, and code generation statistics. Simultaneously, a separate update adds CLI-specific telemetry to enterprise metrics, including daily active CLI users, request/session counts, and token usage totals. This gives engineering leaders the data foundation to track Copilot adoption across all surfaces and plan rollout strategies. [7][8]

• OpenAI closes a $110 billion funding round and launches a Stateful Runtime Environment for Agents on AWS Bedrock. The round — consisting of $50B from Amazon, $30B from Nvidia, and $30B from SoftBank at a $730B valuation — is one of the largest private funding rounds in history. As part of the Amazon partnership, OpenAI is building a new stateful runtime environment on Amazon Bedrock with persistent orchestration, memory, and secure execution for multi-step AI workflows. OpenAI committed to consuming at least 2GW of AWS Trainium compute. This dramatically expands the infrastructure available for OpenAI-powered coding agents and enterprise AI applications. [9][10]

• Vercel releases an AGENTS.md skill with 40+ React/Next.js performance rules designed for AI coding agents. The open-source react-best-practices repository encapsulates over a decade of Vercel's engineering knowledge into rules prioritized from CRITICAL to LOW across eight categories. Rules compile into a single AGENTS.md document installable into Claude Code, Cursor, Codex, and OpenCode via npx skills add vercel-labs/agent-skills. With 21k+ GitHub stars and 150k+ weekly installs, this represents a significant step toward standardizing how human engineering knowledge is packaged for consumption by AI agents. [11]

• An experienced developer's detailed writeup on AI agent coding with Claude Opus 4.5 demonstrates both the current capabilities and limitations of agentic development. Max Woolf, a self-described agent skeptic, documents building multiple Rust and Python projects using Claude Code and GitHub Copilot with Opus 4.5, finding it dramatically better than Sonnet 4.5. Key insights include the critical importance of well-crafted AGENTS.md files as the main differentiator for agent quality, and that agents work best when users have "approximate knowledge of many things" with sufficient domain expertise to audit results. The post provides concrete prompting methodology and open-source project examples. [13]

AI Coding News

• OpenAI introduces a Stateful Runtime Environment for Agents in Amazon Bedrock, enabling persistent multi-step AI workflows on AWS. As part of the broader $110B funding announcement and Amazon partnership, OpenAI is developing a new runtime where OpenAI models will run on Amazon's Bedrock platform with persistent orchestration, memory, and secure execution capabilities. Amazon CEO Andy Jassy stated that "our unique collaboration with OpenAI to provide stateful runtime environments will change what's possible for customers building AI apps and agents." OpenAI also committed to building custom models for Amazon consumer products and expanding its AWS compute partnership by $100B. This creates a significant new platform for developers building agentic applications with OpenAI models on cloud infrastructure. [9][10]

• Microsoft open-sources Evals for Agent Interop, a starter kit for benchmarking enterprise AI agents across realistic workflows. The kit provides curated scenarios, representative datasets, and an evaluation harness that measures schema adherence, tool call correctness, and AI judge assessments for coherence and helpfulness. Initially focused on email and calendar interactions, it includes a leaderboard concept for comparing agents built with different stacks and model variants. Deployed as a Docker Compose set of three images, it gives teams a repeatable baseline for evaluating agentic AI before broad enterprise rollout. [12]

• A detailed blog post chronicles an AI agent coding skeptic's journey from dismissal to cautious optimism with Claude Opus 4.5. Max Woolf documents building four Rust projects (icon-to-image with pyo3 bindings, browser-based word clouds via WASM, a terminal MIDI DAW, and a terminal physics simulator handling 10,000+ balls) and multiple Python projects using Claude Code and Copilot. He identifies AGENTS.md as the critical success factor, shares his methodology of writing elaborate prompts in tracked Markdown files, and notes that the leap from Sonnet 4.5 to Opus 4.5 was widely recognized across developer communities. The post provides unusually concrete, reproducible evidence for evaluating current agent capabilities. [13]

Feature Update

• GitHub Copilot CLI v0.0.419 adds the experimental /chronicle command and comprehensive keyboard/mouse controls. The /chronicle command generates standups, tips, and improvement suggestions powered by session history. New keyboard shortcuts include Ctrl+F/Ctrl+B for alt-screen page scrolling, Ctrl+G for external editor integration, and Home/End for scroll buffer navigation. The --mouse/--no-mouse flag and config option let users disable mouse mode in alt screen. MCP server names now support npm-style naming with dots, slashes, and @ characters. The AUTO theme now uses the terminal's actual ANSI color palette for accurate theming. [1]

• GitHub Copilot CLI v0.0.420 extends auto-update to binary executables and adds automatic 502 retry. Auto-update now updates both the JS package and the binary executable, ensuring the full installation stays current. Plugin and marketplace git repos update correctly after force-pushes and tag-based installs. The 502 bad gateway errors that previously crashed sessions with raw HTML output are now retried automatically. Users can type # to reference GitHub issues, pull requests, and discussions directly in the CLI. [2]

• Copilot SDK v0.1.28 ships breaking security changes: all permissions denied by default with mandatory permission handlers. This is a significant release that changes the default permission posture across all SDK languages — all permissions are now denied by default, and a permission handler is required on session creation. New features include agent selection and session compaction APIs, permission checks for SDK-registered custom tools, clientName in SessionConfig, and E2E scenario tests for all SDK languages. The release also adds Azure Managed Identity guide for BYOK, Microsoft Foundry Local documentation, OpenTelemetry GenAI semantic conventions guide, and bumps minimum Python to 3.11. [3]

• Copilot SDK v0.1.29 updates the @github/copilot dependency to v0.0.420 and changes exception handling. The CopilotClient.stop() method now raises an exception group instead of returning a list of exceptions, aligning with modern Python exception handling patterns. The Go SDK also received a corresponding v0.1.29 release with the same copilot dependency update. [4]

• Claude Code v2.1.62 fixes a prompt suggestion cache regression that reduced cache hit rates. This is a targeted bugfix release addressing a regression that degraded prompt suggestion cache performance. The previous day's v2.1.61 fixed concurrent config file corruption on Windows, and v2.1.59 introduced auto-memory, the /copy command with interactive picker, and improved MCP OAuth token refresh handling. [14]

• OpenAI Codex publishes five Rust alpha builds (v0.107.0-alpha.2 through alpha.6), signaling active Rust rewrite development. Five alpha releases were tagged on the same day, indicating rapid iteration on the Codex Rust rewrite. No detailed changelogs were provided for these alpha builds, but the volume of releases suggests significant development velocity on the Rust port of the Codex CLI. [15]

• Gemini CLI v0.31.0 delivers a major stable release with parallel function calling, session SDK, and Gemini 3.1 Pro support. Key features include parallel function calling for read-only tools, MCP progress updates, session-based SDK architecture refactor, A2A API key authentication, project-level policy support, macOS run-event notifications, /extensions explore command, pr-address-comments skill for PR feedback, TrueColor detection, and Gemini 3.1 Pro Preview model support. Security improvements include rate-limiting web_fetch to mitigate DDoS via prompt injection and stripping deceptive Unicode characters from terminal output. [5]

• Gemini CLI v0.32.0-preview.0 introduces sub-agent classification, a generalist agent, and an experimental Gemma Router. The preview adds Kind.Agent for sub-agent classification, a generalist agent mode, robust A2A streaming reassembly, interactive shell autocompletion, model steering in workspaces, and plan mode improvements including external editor support and complexity-adaptive planning. An experimental Gemma Router uses a LiteRT-LM shim for the Composite Model Classifier Strategy. Extensions now load in parallel and support policy engine integration. [6]

• GitHub Copilot metrics is now generally available with dashboards and APIs at enterprise, organization, and user levels. The GA release includes a Copilot usage dashboard for code completion activity and IDE usage, a code generation dashboard quantifying lines suggested/added/deleted across completions, chat, and agent features, and fine-grained access controls via custom enterprise roles. APIs are available at enterprise, organization, and individual user levels. Data residency support is included for enterprises with residency requirements. [7]

• Enterprise Copilot usage metrics now includes Copilot CLI activity data. Copilot enterprise usage metrics has expanded to include CLI-specific telemetry: daily active CLI users, CLI request and session counts, and token usage totals including average tokens per request. This allows enterprise administrators to track CLI adoption alongside IDE usage and compare usage patterns across development environments. [8]

• Vercel releases react-best-practices, an open-source AGENTS.md skill with 40+ performance rules for AI coding agents. The repository organizes rules across eight categories prioritized from CRITICAL to LOW, focusing on eliminating async waterfalls and reducing bundle size as the highest-priority categories. Rules compile into a single AGENTS.md document and can be installed into Claude Code, Cursor, Codex, and OpenCode via npx skills add vercel-labs/agent-skills. The framework covers higher-level architectural decisions like request waterfalls and bundle composition that traditional linting tools do not address. [11]

• Microsoft open-sources the Evals for Agent Interop Starter Kit for systematic evaluation of enterprise AI agents. The starter kit ships with declarative evaluation specs in JSON, a harness measuring schema adherence and tool call correctness, and calibrated AI judge assessments. It includes a leaderboard for comparing agents across different stacks and model variants, and is deployed as three Docker Compose images. The kit is designed to bring repeatable, transparent evaluation baselines to teams building agentic AI for enterprise workflows. [12]