AI Coding Roundup — September 6, 2026: Claude Code 2.1.263 Fixes Session Bugs as Copilot CLI Adds Custom Agents

Four AI coding tools quietly tightened their agentic tooling in the last day. Claude Code shipped a bug-fix release aimed squarely at long-running background and multi-agent sessions, GitHub Copilot's CLI grew a taskbar presence and support for custom agents, OpenAI's Codex CLI rolled out a remote plugin marketplace and vim-style undo, and OpenCode made subagent failures resumable instead of silently empty. None of it is as flashy as a new frontier model, but for anyone running these agents daily, these are exactly the fixes that matter.

Claude Code 2.1.263: closing the gaps in background and multi-agent sessions

Anthropic's latest Claude Code release is a pure reliability pass, and a useful one for teams leaning on background sessions and dynamic workflows. It fixes a command interrupted mid-run showing as a completed "Ran 1 shell command" with no sign it was cut short, stops pressing ← or running /background during a dynamic workflow from silently restarting already-finished subagents (it now asks first and states how many would restart), and resolves a race where opening a just-started session in claude agents while its worker was still booting killed it with a "stopped while the respawn was in flight" error. It also fixes claude agents double-listing a backgrounded named session, stops the background retention sweep from deleting git worktrees under .claude/worktrees/ that you created yourself, and scales the auto-mode safety-check deadline with prompt size so tool calls stop getting denied as "temporarily unavailable" on very large sessions. Full notes are in the Claude Code changelog.

GitHub Copilot CLI v1.0.83: a taskbar presence and custom agents

GitHub's Copilot CLI moved further toward feeling like a persistent desktop tool rather than a one-off terminal command. Version 1.0.83 now shows active sessions in the Windows taskbar with live status updates, so a long-running agent task no longer disappears into a background terminal tab, and it adds support for custom agents that developers can define and reuse across sessions. See the Copilot CLI releases for the changelog.

Codex CLI: plugin marketplaces and vim undo arrive in the TUI

OpenAI's Codex CLI picked up several workflow niceties this week that go beyond the GPT-6 Astra rollout covered here yesterday. Vim mode now supports undo and redo (u and Ctrl+R), preserving complete drafts including pasted content and attachments across grouped edit steps. The plugin CLI can now list, install, and remove plugins from remote marketplaces, with catalog entries showing source, version, install policy, and authentication policy. A new tui.auto_recap = false setting disables automatic recaps while keeping the manual /recap command available, and Plus and Team plans now get an earlier usage warning once less than half of their roughly five-hour allowance remains. Details are in the Codex changelog.

OpenCode: resumable subagent failures and full session export

SST's provider-agnostic terminal agent focused on making failures recoverable rather than silent. Failed subagent tool calls now surface with a resumable task_id instead of returning an empty result, so a dropped subagent can be picked back up rather than restarted from scratch. Provider timeout handling was also fixed, with default header and streamed-chunk timeouts both set to five minutes to accommodate slow model startups, and responses ending in a network_error are now retried automatically. On top of that, users can export full session transcripts as JSON straight from the UI. Full details in the OpenCode changelog.

Why it matters

With GPT-6 Astra now available across nearly every major assistant, raw model access is no longer a differentiator — every tool ships the same frontier models within days of each other. What's left to compete on is exactly what shipped today: can a background session survive an interrupted command, can a failed subagent be resumed instead of lost, can a long session be exported and audited, and does the tool integrate cleanly into how developers actually work (a taskbar icon, a reusable custom agent, a plugin marketplace). That's the layer where Claude Code, Copilot, Codex, and OpenCode are now visibly racing each other.

Resources & references

• Claude Code changelog
• GitHub Copilot CLI releases
• OpenAI Codex CLI changelog
• OpenCode changelog