Claude rebuilds Projects as a live conversation while Codex and OpenAI's SDKs race ahead

Projects become a workspace, not a folder

The week's biggest structural change came from Anthropic, which rebuilt Projects as a single multi-threaded conversation: instead of nesting chats inside a folder, a project is now one thread where each request spawns work Claude runs in the cloud and reports back on, distinguishing what's ready for review from what's waiting on you. It's in beta for select Pro and Max users and rolling out across Claude Code, chat, and Cowork — existing projects are untouched for now, which softens the migration but also means two mental models coexist for a while.

That's part of a broader consolidation. Anthropic merged chat and Cowork into one unified Claude, so a quick question and a long-running task live in the same conversation, and launched Claude Slides, Design, and Docs in beta so a launch brief can turn into a deck, social images, and a one-pager without leaving the chat. Both push toward Claude as the place work happens rather than a place you ask questions. On the enterprise side, a Salesforce plugin with 37 pre-built sales skills extends that reach into CRM data.

Claude Code hardens its session machinery

Claude Code shipped a dense run of releases, mostly about making long sessions survive contact with reality. The most consequential is v2.1.275, which fixes a restored memory file's age note drifting between requests after compaction or resume — a subtle bug that silently caused prompt cache misses on every turn — and stops plugin and marketplace output from leaking passwords or tokens embedded in git, ssh, or marketplace URLs. Alongside it, 2.1.274 makes sessions that were stuck retrying "unexpected tool_use_id" errors self-heal corrupted transcripts or exit cleanly with a /rewind hint, and adds CLAUDE_CODE_MCP_STARTUP_WAIT_MS to bound how long the first non-interactive turn waits on MCP servers. v2.1.277 brings AGENTS.md fallback for projects without a CLAUDE.md and fixes claude -p hanging after internal errors — which now exit with code 1 — while 2.1.271 adds per-command allowed_domains to sandboxed auto mode so each Bash or PowerShell command reaches only the hosts it needs, plus fast mode in Remote sessions. Smaller follow-ups cleaned up after the rush: v2.1.276 patched a 400-error regression for proxy and gateway users, and v2.1.278 defaults auto mode to the server-side classifier for API, Bedrock, Vertex, and Foundry, avoiding classifier overhead charges.

Codex goes multimodal, OpenAI's SDKs worldwide

OpenAI's Codex CLI 0.155.0 added experimental /voice conversations with live transcripts and microphone controls, Touch ID verification for MCP requests in local TUI sessions on supported Macs, and a fix for a Windows-process escape from restricted WSL sandboxes — security work that pairs with its hardening of brokered shell snapshots. A quick 0.155.1 follow-up disabled reasoning summaries by default where providers reject them, and Codex for iOS gained folders in the file picker and writing blocks.

The SDKs coordinated on a shared API surface: Python 3.15.0 led with managed Responses WebSocket sessions, prompt-cache prewarming, and compaction progress events, with Node v7.18.0 and Go v3.62.0 following with the same managed WebSocket sessions. All three then added webhook endpoint management and deprecated MCP connector_id across Node and Go. Deeper in the week, Go and Node added vault credentials and safety webhooks, and a memory leak in Python's parse_response got fixed. If you maintain an agent against this API, the connector_id deprecation is the migration to plan for.