AK× (@akxwire) header
AK× (@akxwire) profile picture

Same-day release notes for Claude Code, Codex, Grok Build, Cursor and open models: what changes for you, with the source link. By @KHASSINX

Joined September 2009
Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)Photo by AK× (@akxwire)
Pinned Tweet
AK× is for people shipping with Claude Code, Codex and Grok. → What changed in each release, and what breaks → Setup steps you can paste, checked against the official docs → A source link on every post No rumors, no hype. Follow if you build with AI coding agents.
52
Sandbox every session `claude remote-control` starts for your phone. Since Claude Code 2.1.321 the server takes its own --settings: claude remote-control --settings rc.json rc.json: { "sandbox": { "enabled": true } } It applies to every session the server starts.
2
29
Gotchas from the docs + changelog: → --settings goes after remote-control. Before the verb, it refuses to start. → --sandbox was removed in v2.1.300; passing it is an error. → Edited rc.json? No new session starts until you restart the server. code.claude.com/docs/en/remo…
9
Keep unattended Claude Code runs from hanging on a permission prompt. In ~/.claude/settings.json: "permissions": { "denyInsteadOfPrompting": true } A call that would prompt is denied and the turn keeps going. Allow rules and your mode still approve calls. Needs v2.1.319+.
1
28
Gotchas from the docs: → Only user or managed settings (or --settings). A repo can't turn it on for you. → It also drops AskUserQuestion and the plan-mode tools, so plan mode can't exit on its own. → Just one -p run? Pass --permission-prompts none. code.claude.com/docs/en/sett…
4
Keep your .env out of Claude Code. In ~/.claude/settings.json: "permissions": { "deny": ["Read(./.env)", "Read(./.env.*)"] } Rules run deny → ask → allow, first match wins, so a narrower allow can't override a deny. Run /status to confirm the file loaded.
3
1
117
Two gotchas from the docs: → .claudeignore has no effect. Move its entries into Read deny rules. → Deny rules cover Claude's file tools and cat/head/sed, not a script that opens files itself. For OS-level blocking, turn on the sandbox. code.claude.com/docs/en/perm…
14
Codex CLI 0.167 is out. → codex agents --no-daemon: command center with unread markers → Switch backends in the picker; drafts stay → Image preview while composing + GCP Vertex setup github.com/openai/codex/rele…
42
Claude Code or Codex guessing your iOS build from files? Give it Xcode. 1) Xcode › Settings › Intelligence: allow external agents 2) claude mcp add --transport stdio xcode -- xcrun mcpbridge or: codex mcp add xcode -- xcrun mcpbridge 3) Open the project in Xcode, then prompt
2
2
73
Check it: claude mcp list or codex mcp list. Xcode alerts you when the agent connects and while it's active. To turn it off, set the same menu to Never. Apple's doc: developer.apple.com/document…
19
Claude Fable 5.2 is out, 20% cheaper and up to 30% faster than 5.1. → $8/$40 per Mtok, 1M context → Fast mode: speed "fast", up to 2.5× → Default Fable in Claude Code 2.1.321 Note: edit the prompt prefix and resend 5.2 thinking blocks = 400 error. anthropic.com/claude-fable-a…
48
Claude Code 2.1.319 is out. → MCP tool-as-task calls go to the background right away → permissions.denyInsteadOfPrompting: deny and keep going → /claude-api latency-optimize finds where your API waits github.com/anthropics/claude…
13
Kimi K3.1 Pro weights are now public on Hugging Face. MoE, 2.8T total / 104B active, 1M-token context. vLLM and SGLang serving recipes are linked in the model card. Caveat: model-as-a-service firms over $20M/yr revenue need a separate license deal. huggingface.co/moonshotai/Ki…
5
Claude Code 2.1.318 now tells you when your stored claude.ai sign-in expires. → claude auth status shows loginExpiresAt → claude -p warns 10 days before it ends → Mods: side jobs can run at low effort → Cloud sessions can start with no repo github.com/anthropics/claude…
16
What you get from @akxwire: → Same-day notes on Claude Code, Codex, Grok Build and Cursor → Open-model drops, with the source link → What changes for you, and how to turn it on → A recap every Sunday
14
Claude agents no longer resend the whole conversation every turn. Messages API threads (public beta): → Send history once, then only new messages → ~35% lower median TTFT at 500K–1M tokens → Same token cost, 5-min expiry claude.com/resources/article…
10
DeepSeek open-sourced a coding agent for a whole GitHub repo, not one person. dsh-superagent (MIT, research prototype): → Teammates ask it for work on GitHub → It replies in the thread, e.g. with a PR → Acts only when asked or a team rule allows github.com/deepseek-ai/dsh-s…
6
Claude Code 2.1.317 now asks before a web or shell tool call sends your account email over the network. → Agent SDK hooks: onFailure: "block" stops the action if your callback fails → Ctrl+C to clear the prompt no longer kills background agents github.com/anthropics/claude…
1
5
The best AI model solves just 12% of real expert work. SemiAnalysis GDPx: 206 tasks, 11 domains, up to 8 hours each. → Claude Opus 5.5 12.0 → Gemini 4 Carbon 10.9 → GPT-6 Astra 9.6 → Grok 4.7 3.9 → Finance and legal: nearly every model scores 0 gdpx.semianalysis.com
8
Codex profiles switch your whole setup with one flag: codex --profile deep-review → Loads ~/.codex/config.toml, then ~/.codex/deep-review.config.toml on top → Only put the values that differ → Since 0.134, [profiles.x] tables are ignored developers.openai.com/codex/…
9