AI Coding News — August 30, 2026: GitHub Copilot Adds a Thinking Effort Dial
Yesterday's edition closed on Claude Code's new model-switch hooks, GitHub Copilot's confirmed list of models retiring on August 31, and Cursor's cloud agents dropping their GitHub requirement. Today the headline shifts to Visual Studio, where GitHub Copilot picked up a thinking effort dial and org-wide custom agents, OpenAI's Codex CLI handed extensions the ability to rewrite MCP tool results before they reach the model, and OpenCode quietly shipped a free, 1-million-token stealth model.
GitHub Copilot gets a thinking effort dial and org-wide custom agents
GitHub's GitHub Copilot in Visual Studio — August update rolled out on August 28 and gives developers more control over how Copilot reasons, which models they use, and how teams share specialized agents.
- Thinking effort controls: supported models now expose Low, Medium, and High settings. Low favors quick, low-token responses for straightforward suggestions; Medium balances depth and speed for everyday coding; High spends more tokens on deeper reasoning for tricky algorithms, architecture decisions, and hard-to-debug issues.
- Organization-level custom agents: GitHub organization and enterprise owners can now publish custom agents for use across every repository in the org. Visual Studio automatically detects them and shows both the agent's description and its organization source in the agent picker.
- Git-integrated code review: you can ask the Git agent to review uncommitted changes or specific commits before opening a pull request, with findings shown inline in the editor and collected in a navigable list in Git Changes.
- Usage visibility: opening the context window from the Copilot prompt box now offers a "View all Copilot usage" option covering full plan details, useful for anyone tracking spend against a seat-based plan.
The update lands as GitHub's global model policy reaches general availability, with enforcement for Business and Enterprise plans rolling out gradually between August 26 and September 1. Grok 4.6, xAI's latest reasoning model, is also rolling into Copilot for agentic coding and complex multi-step workflows during the same window.
OpenAI's Codex CLI lets extensions rewrite MCP tool results
Codex CLI v0.151.0, published August 29, focuses on the plumbing that agentic MCP workflows depend on rather than user-facing features.
- A configurable grace period now governs how long Codex CLI waits to discover tools from optional MCP servers, cutting down on flaky startup when a server is slow to register its tools.
- Extensions can now inspect or replace MCP tool results before they reach the model, opening the door to redaction, caching, or reformatting layers sitting between a tool and the reasoning loop.
- Plugin catalogs now combine per-repository configuration and report invalid project marketplaces without hiding the valid plugins alongside them.
- Restored permission profiles are now preserved across TUI turns, and the /cd command can no longer be used to quietly weaken sandbox restrictions mid-session.
OpenCode ships a free, 1M-context stealth model
Per OpenCode's August release tracker, the SST-maintained fork launched Ox Alpha on August 20 as a free stealth model offering a 1-million-token context window and multimodal input — a way to trial frontier-scale context without a paid provider key. The rest of August's releases were maintenance: fixed Cloudflare AI Gateway routing for third-party providers, corrected handling of Anthropic's dashed model slugs, and resolved session-header and GitHub-auth bugs.
Anthropic's own Claude Code held at v2.1.251 (Aug 28) with nothing new published since — a rare quiet stretch after two releases in the same week, per the official Claude Code changelog.
What it means for developers
If you're on Visual Studio, try the High thinking effort setting on your next hard debugging session, and ask your org admin whether org-level custom agents are worth standardizing on. Codex CLI's MCP result hooks matter most if you're building or maintaining MCP servers — you can now filter or reshape what reaches the model without forking the server itself. Ox Alpha is worth a spin if you want to test a 1M-context window for free, but treat it like the stealth/alpha label suggests: not for production workloads yet. Claude Code and Cursor users have nothing new to act on today beyond yesterday's items.
Sources & further reading
Comments
Share your thoughts and join the conversation
