AI Coding Tools Roundup: Claude Code v2.1.282, Codex Ships GPT-6 Sol and Luna
Three separate corners of the AI coding stack shipped meaningful updates in the last day: Claude Code jumped to v2.1.282 with a dense list of reliability fixes, OpenAI's Codex CLI hit 0.157.0 with new GPT-6 Sol and Luna models plus Amazon Bedrock support, and GitHub Copilot for JetBrains landed AI-assisted tool approvals and a conversation rewind feature. Here's what changed since yesterday's roundup.
Claude Code v2.1.282 Tightens Gateway Reliability and Squashes a Long List of Bugs
Anthropic's Claude Code v2.1.282, released September 24, is a dense bug-fix release layered on top of v2.1.281's Claude apps gateway work. It fixes conversations that failed with 400 errors when history contained undecryptable web search results, stops continued or resumed sessions from re-sending earlier messages in a changed form, and resolves sessions crashing with "Invalid data in redacted_thinking block" errors. Extended thinking no longer gets silently dropped during slash commands, and a recursive rm running unprompted in auto mode under certain conditions has been fixed — a meaningful safety fix for anyone running agentic sessions unattended.
The release also adds a maxProseWidth setting to cap prose width in wide terminals while keeping tables and code blocks full-width; a startup notice plus /status and claude doctor entries listing ignored telemetry variables; an allowClaudeInChromeWithManagedMcp managed setting for running claude --chrome alongside an exclusive managed-mcp.json; and store.readiness_grace_seconds for the Claude apps gateway to ride out short Postgres outages. Multiple vim-mode cursor and editing bugs, Windows credential-read error messages, and fullscreen-mode rendering issues round out the fix list.
OpenAI Codex CLI 0.157.0 Ships GPT-6 Sol and Luna With Amazon Bedrock Support
OpenAI's Codex CLI rust-v0.157.0, released September 25, introduces the GPT-6 Sol and Luna models with Amazon Bedrock support and migration prompts for anyone still on older models. The release also turns on fullscreen transcripts by default, adds shift-click to extend text selections, and enables automatic background-server startup for eligible interactive sessions, with recovery choices when server settings are incompatible.
A new f keyboard shortcut forks a conversation into another app while preserving drafts and queued prompts, and /import now works in remote sessions and local background-server sessions. Bug fixes cover voice conversations persisting across thread switches, draft recovery during history searches, tmux mouse handling and Terminal.app SSH scrollback, and proxy routing for realtime connections and web search.
GitHub Copilot for JetBrains 1.18.0 Adds AI-Assisted Approvals and Org-Wide Skills
GitHub shipped Copilot for JetBrains 1.18.0 on September 22, and its headline feature is AI-assisted tool approvals: low-risk tool calls during agent sessions are approved automatically, while higher-risk actions still prompt for a decision. The update also lets developers rewind a conversation — re-editing an earlier message rolls back both the chat and any file changes made after it, so you can course-correct without cluttering the thread.
Organizations can now push shared skills and custom instructions that apply across both local and agent-based sessions, and the Codex agent inside JetBrains gained plan review, letting you inspect a proposed strategy before it runs. MCP tool management also improved, with persistent per-tool controls and a toggle for the built-in GitHub MCP Server.
What It Means for Developers
None of today's releases is a flagship model launch, but each addresses friction that shows up in daily use. Claude Code's v2.1.282 fixes — especially the redacted_thinking crash and the unprompted recursive rm fix — matter most for anyone running long agentic sessions in auto mode. Codex's GPT-6 Sol and Luna on Bedrock open the door for enterprise teams standardizing on AWS infrastructure, and the background-server and fork-conversation features push Codex further toward a persistent, multi-session workflow rather than one-off terminal runs. Copilot for JetBrains' AI-assisted approvals and rewind feature are the kind of trust-building controls that make agentic coding sessions easier to run unsupervised — worth trying if you've held back on letting Copilot's agent operate with less oversight.
Resources & References
Comments
Share your thoughts and join the conversation