OpenAI's Codex 0.151.0 adds an "exit checkpoint" that lets you inspect and rewrite an MCP tool's return values before the model reads them. Alongside 0.150's Interrupt hooks and the ongoing autonomy fixes, we trace how agent safety design is opening up to the operator's discretion.
OSS specialist Designet launched its "In-House AI Infrastructure Construction Service" on August 26, using a local LLM to process internal documents without sending them outside. Here we lay out the RAG configuration built on Ollama, Dify, and Fess, the pricing starting from ¥1.4 million, and the bl
Claude Code v2.1.251 is out. It ships PreModelSwitch/PostModelSwitch hooks that intercept model switches, a prompt-cache breakdown and spend-limit bar in /cost, and a fix for a symlink loophole in file operations — all at once, shoring up the plumbing for unattended operation.
Claude Code v2.1.248's `--restricted` is a launch option that removes command execution and WebFetch and confines file operations to the working directory. It hands least privilege to other people's repositories and unattended runs. It also includes quiet fixes like preventing credential exfiltratio
Ollama ships official Claude Desktop support. With a single toggle, you can run open models (Qwen/DeepSeek/Kimi/GLM) locally or in the cloud without leaving Claude's interface. This piece lays out the practical gains in cost and privacy, how to split work between the two, and the burdens — speed, qu
Claude Code v2.1.247 is out. It adds /claude-api cost-optimize, an interactive command that trims Claude API spending while measuring the effect of each change, and ships with Admin API support plus a SendFeedback tool that drafts bug reports.
On August 14, Auto mode became the default, moving approval from humans to an AI classifier. v2.1.246 adds a tab in /permissions to view and edit those rules, and warns about the over-broad permissions caused by wildcards placed before subcommands. A step toward reclaiming control over "what you del
On August 24, OpenAI made its GPT-5.6 models (Sol/Terra/Luna) available in AWS's development environment, Kiro. With spec-driven development that nails down requirements, design, and tasks first, the cost per completed task on Terminal-Bench 2.1 reportedly fell by about 82%. We also unpack how to pi
On August 25, OpenAI added browser updates to Codex and ChatGPT Work. The highlight is 'Site Tools (WebMCP),' which lets agents call the tools a site itself declares. We unpack the shift from guessing at the screen to sites handing features over to agents, along with the expanded browser support and
Claude Code v2.1.243 adds a per-loop cost breakdown to /usage, making it easier to pinpoint the "runaway tokens" of unattended operation. The Linux binary is compressed from 340MB to 75MB, and memory use is cut too. The same-day v2.1.245 fixes a startup crash on glibc 2.44 environments.
ByteDance is folding its Trae coding assistant and Coze agent-building tool into its Doubao chat app and launching an office-focused "Doubao Work," a move aimed at catching Tencent's WorkBuddy (over 13 million DAU in July). We break down the simultaneous consolidation of the "entry point" by China's
GitHub has released Copilot in public preview for both Slack and Microsoft Teams. Mention @GitHub and you can trigger Issue triage, fixes in a cloud sandbox, and PR creation right from chat. Here's a rundown of how it's designed to ride on your existing GitHub permissions and approval gates—plus cav
