AI Coding News: August 8, 2026 — Claude Code Ships Spend Limits and Agent Workspace Trust

Three tools developers reach for every day updated within the same 24 hours: Claude Code shipped two point releases with gateway spend limits and a new workspace-trust prompt for agents, OpenCode fixed a bug that could scramble session timelines, and GitHub Copilot CLI closed out a run of timeline patches. Cursor also rolled out a smarter model router. Here's what changed and why it matters for day-to-day use.

Claude Code 2.1.225 and 2.1.226: spend limits, agent trust, and a long fix list

Anthropic pushed two back-to-back releases of Claude Code on August 8. Version 2.1.225 is the meatier of the pair — full details are in the official changelog, but the highlights are:

• Gateway spend-limit support — the usage warning that fires when a team hits its cap now names the limit, its reset time, and any custom message the operator configured, which matters for teams running Claude Code behind an internal gateway.
• Workspace trust for agents — "claude agents" now shows the same untrusted-directory trust prompt that "claude" itself already had, closing a gap where launching an agent in a new folder skipped that check.
• OAuth reliability fixes — a transient 401 no longer overwrites a long-lived CLAUDE_CODE_OAUTH_TOKEN with a short-lived one (which used to break headless sessions until restart), and MCP OAuth servers on macOS no longer fail in bursts after a slow keychain read.
• Session and Remote Control fixes — cross-session messages no longer sit parked without a notice in headless sessions, conversation history survives session resume after very large conversations get compacted, and SendMessage can now open a conversation with a Remote Control session on another machine by name instead of waiting for it to message first.
• Auto mode no longer counts a safety-filter refusal of its own permission check toward the consecutive-block limit — the action still gets denied, but the model is told to move on instead of retrying into a wall.

2.1.226, shipped right behind it, is billed simply as "bug fixes and reliability improvements" — a quiet follow-up to a release with a genuinely long list of changes.

OpenCode v1.18.15 fixes a session-timeline ordering bug

OpenCode's v1.18.15 release makes chronological message ordering stay correct even when imported or legacy message IDs are out of order, so revert and fork actions now follow actual message chronology instead of ID order — a real fix for anyone importing sessions or working in long-lived conversations where the two could drift apart. The release also fixes blob-based attachments failing to load in the web UI, cleans up stale truncated files more reliably, and lets the desktop app export a full session transcript as JSON. Community contributors also shipped tmux clipboard fixes and configurable cursor styles.

GitHub Copilot CLI closes out its v1.0.79 patch series

After landing timeline duration headers and worktree support last week, Copilot CLI spent this week stabilizing that work. v1.0.79-6 fixes a rare internal delay that was printing a diagnostic warning on top of the interactive UI, retries and reports session-history load failures instead of leaving the timeline blank, and stops the scrollbar from jumping when a long session resumes in the background. On the model side, Kimi K3 went generally available in GitHub Copilot earlier this week — an open-weight model hosted on Fireworks AI at $3/1M input and $15/1M output tokens, rolling out to Pro, Pro+, Max, Business, and Enterprise plans (off by default for Business and Enterprise until an admin enables it).

Cursor ships a smarter model router

Cursor's latest changelog details Cursor Router's two auto-routing modes: Auto Intelligence, now delivering above-Fable-level satisfaction at 68% lower cost (an 18-point improvement since launch), and Auto Balance, which now beats Opus 4.8 quality at 41% lower cost while nudging satisfaction up another 3%. Both modes adapt their routing from live production traffic rather than a fixed model list, and Opus 5 has joined the pool they route across.

Also on the radar

• Gemini CLI patched through v0.54.4 this week, classifying capacity exhaustion as a terminal error so it stops retrying into a hang, with v0.55.0 nightlies already testing ahead of the next stable — see the full release history.
• OpenAI's Codex CLI shipped installable Agent Plugins, an --approve-for-me flag for auto-reviewed approvals, and Cursor-skill imports in rust-v0.147.0 earlier this week, and a rust-v0.148.0-alpha.1 build is already in testing.

Nothing here breaks existing workflows, but a few items are worth acting on today: teams on a Claude Code gateway should confirm they're on 2.1.225+ to get readable spend-limit messages, and anyone running long-lived or imported OpenCode sessions should update to 1.18.15 before relying on revert/fork.

References and further reading

• Claude Code changelog
• OpenCode releases on GitHub
• GitHub Copilot CLI releases
• Kimi K3 GA announcement — GitHub Changelog
• Cursor changelog
• Gemini CLI releases on GitHub
• OpenAI Codex CLI releases on GitHub