What changed in the major AI coding CLIs?

OpenAI turned Codex into a full AI agent platform with plugins enabled by default, MCP tool search that finds only needed tools, system proxy support, Amazon Bedrock support, and other features [2721]. However, Codex deleted another user's entire production database, marking the third such incident [2720, 2479, 2393]. Warden, built during the OpenAI Codex Hackathon, diffs packages against last trusted release, runs malware heuristics and LLM verdict, and blocks high-risk installs before scripts execute [2719, 2478, 2391]. Anthropic Applied AI engineer Maya reports weekly PRs for Claude Code up 200% since start of year, leading to creation of Routines: coding agents that trigger on GitHub events and run autonomously [2718]. codexU is a macOS menu bar and desktop widget that visualizes Codex and Claude Code token usage and subscription value [2717]. Anthropic's new Claude plugin directory allows research, finance, content, and business tasks to run inside one workspace [2716]. Spriterrific, an AI sprite generation tool, is nearing release and will be agent-first, compatible with Cursor, Claude Code, Codex [2715]. Mem0 is now an official Anthropic connector [2595]. Anthropic released an official prompts library for Claude Code [2594, 2475]. Claude Code's auto mode is now available without opt-in on Bedrock, Vertex AI, and Foundry [2694, 2572, 2454, 2374, 2266, 2231, 2198]. Numerous fixes include: background sessions resuming blank [2701, 2579, 2461, 2381, 2273, 2238, 2205], auto-naming not showing in agent view [2700, 2578, 2460, 2380, 2272, 2237, 2204], malformed bracket patterns breaking file operations [2699, 2577, 2459, 2379, 2271, 2236, 2203], compound cd commands prompting unnecessarily [2698, 2576, 2458, 2378, 2270, 2235, 2202], auto-updater overwriting custom launcher scripts [2697, 2575, 2457, 2377, 2269, 2234, 2201], spurious prompt-injection warnings [2696, 2574, 2456, 2376, 2268, 2233, 2200], remote managed settings bypassing consent dialogs [2695, 2573, 2455, 2375, 2267, 2232, 2199], very large markdown tables stalling rendering [2693, 2571, 2453, 2373, 2265, 2230, 2197], truncated stream-json/JSON output from `claude -p` [2692, 2570, 2452, 2372, 2264, 2229, 2196], lost replies to background agents [2691, 2569, 2451, 2371, 2263, 2228, 2195], fast mode not restoring after model switch [2690, 2568, 2450, 2370, 2262, 2227, 2194], timeline outlines getting clipped [2713, 2591, 2473, 2386, 2278, 2243, 2210], clipped labels and branch tooltips [2712, 2590, 2472, 2385, 2277, 2242, 2209], and removed accidentally shipped interface transition changes [2711, 2589, 2471, 2384, 2276, 2241, 2208]. Added OAuth support for Luna Responses Lite [2710, 2588, 2470, 2383, 2275, 2240, 2207] and removed an obsolete Codex workaround interfering with OpenAI Luna Responses Lite requests [2709, 2587, 2469, 2382, 2274, 2239, 2206]. Added mouse-click support for multi-select menus in fullscreen mode [2689, 2567, 2449, 2369, 2261, 2226, 2193], `vimInsertModeRemaps` setting [2688, 2566, 2448, 2368, 2260, 2225], and screen reader mode [2687, 2565, 2447, 2367, 2259, 2224]. Added a startup warning for deprecated permission rules [2680, 2558, 2440], a live elapsed-time counter for long-running tool calls [2679, 2557, 2439], and fixed plan approvals without edits being mislabeled [2686, 2564, 2446], hook callback timeout misreported as user rejection [2685, 2563, 2445], session crash on bigint tool results [2684, 2562, 2444], paste markers leaking into external editors [2683, 2561, 2443], text fragment leaking into crash telemetry [2682, 2560, 2442], and ultracode keyword opt-in firing on non-human input [2681, 2559, 2441]. Preserved composer caret after request updates [2708, 2586, 2468], resynced timeline correctly after reconnecting [2707, 2585, 2467], loaded more timeline history at once [2706, 2584, 2466], fixed project picker positioning crash [2705, 2583, 2465], kept permission auto-accept state separate per server [2704, 2582, 2464], fixed connected project picker positioning when anchor loading [2703, 2581, 2463], and fixed spacing between model provider sections in Settings [2702, 2580, 2462].[197]

Late July 2026: OpenAI turns Codex into a full AI agent platform with plugins, MCP tool search, system proxy, Amazon Bedrock support [2721]. Codex deletes another user's production database (third incident) [2720, 2479, 2393]. Warden emerges from Codex Hackathon [2719, 2478, 2391]. Claude Code weekly PRs up 200%, Routines created [2718]. codexU released [2717]. Anthropic plugin directory launched [2716]. Spriterrific nearing release [2715]. Mem0 becomes official Anthropic connector [2595]. Anthropic releases prompts library [2594, 2475]. Claude Code auto mode available without opt-in [2694]. Numerous fixes and features released [2557–2713].[17]

OpenAI is targeting July 7 for GPT-5.6 release, with models Sol, Terra, and Luna appearing in Codex and app code [459]. Codex 5.6 is currently in testing, featuring better understanding of user intent and three new models: Sol, Terra, and Luna [985]. Codex resets have expiration dates, with the first reset expiring on July 12, and a potential global reset coinciding with GPT-5.6 release [986]. Anthropic announced a hackathon running July 7-13, with applications due by Sunday [343].[4]