Permission models, default models, and the week of the agent platform

A new default and a new permission posture at Anthropic

The biggest story of the week belongs to Claude Code, which landed two significant releases. First, Claude Code v2.1.197 made Claude Sonnet 5 the default model, bringing a native 1M-token context window and promotional pricing through August. For developers working with large codebases, this is a material change: you can now drop a whole monorepo into context without hitting the old ceiling, and the model itself is the default from the moment you update.

Then just days later, Claude Code v2.1.200 shifted the default permission mode to Manual across CLI, VS Code, and JetBrains. The AskUserQuestion dialog no longer auto-continues, and you opt into an idle timeout via /config if you want it. This is a meaningful trust-and-safety change: the agent now pauses and waits for you by default, rather than assuming it can proceed. The release also fixed a raft of background-session issues — sessions silently stopping after sleep, daemon crashes leaving stale locks, subagents cut off by rate limits — that make the background agent a more reliable companion for long-running tasks.

A smaller but notable fix in Claude Code v2.1.201 removed the mid-conversation system role for Claude Sonnet 5 sessions, and Claude Code v2.1.196 fixed a critical bug where waking a background job deleted the conversation and reran the original prompt. The same release also hardened security: claude mcp list and get no longer spawn repo-approved .mcp.json servers in untrusted workspaces, closing a vector for supply-chain-style attacks.

Grok Build: voice, media, and a dozen small refinements

xAI's Grok Build had an unusually active week, shipping eight releases between June 30 and July 6. The most user-facing change is probably voice dictation (Ctrl+Space or F8), which landed in v0.2.80 alongside a voice language picker in v0.2.86. If you're working heads-down, speaking a prompt instead of typing it is a real workflow improvement.

On the media side, Grok Build 0.2.86 removed per-session file limits for media generation and increased image and video byte budgets. The 0.2.84 release added a remote_fetch option under [features] in config.toml to disable all backend catalog fetches for air-gapped deployments — a signal that enterprise and regulated environments are on the roadmap. That same release dramatically lowered idle CPU and memory usage after long sessions, which is the kind of fix you notice when you leave a session open overnight.

The /effort command in v0.2.82 and the /docs command in v0.2.87 add useful slash-command shortcuts for changing reasoning effort on the fly and browsing documentation. Tool call grouping (collapsing consecutive read/search/list calls into single rows) is now enabled by default in v0.2.88, which makes the agent's output much cleaner for long planning sessions.

Codex on iOS and Cursor's team MCPs

OpenAI's Codex had a notable week on mobile: ChatGPT for iOS now lets you create, search, fork, and manage Codex tasks directly from conversations, with branch filters for staged, unstaged, and branch changes. SSH support using private keys (or no credentials) makes it practical to connect to remote dev environments from a phone. The inline photo and camera pickers, plus attachment previews, suggest a push toward mobile-first agent interaction.

Cursor expanded team marketplaces to support Team MCPs and organization groups. Admins can now configure MCP servers once and distribute them across cloud agents, the agents window, IDE, and CLI. Marketplace access can also be restricted to organization groups beyond SCIM directory groups, which gives larger teams finer-grained control over which tools are available to whom.

Devin's Slack sync and analytics

Cognition's Devin shipped bidirectional Slack thread sync, a read-only MCP toggle, and a weekly PR ratio analytics chart with CSV export. The redesigned environment page and cross-platform repository cloning are quality-of-life improvements, but the Slack sync is the standout: it means you can start a Devin session from a Slack thread and keep the conversation flowing without switching contexts.