AI Coding News: July 20, 2026 — GitHub Copilot CLI Cuts Subagent Depth From 6 to 4, Codex Fixes GPT-5.6 Context Windows
Two smaller releases land today, but they rhyme with a pattern that's been building all month: agentic coding tools are independently deciding that letting subagents spawn subagents needs a leash. GitHub Copilot CLI 1.0.71 cut its default sub-agent nesting depth from 6 to 4, while Codex CLI 0.144.6 fixed a context-window metadata bug affecting GPT-5.6 models.
GitHub Copilot CLI 1.0.71: Sub-Agent Nesting Cut From 6 to 4, Canvas Support Ships
Copilot CLI's July 16 release reduces the default depth a chain of sub-agents can nest to before the CLI stops spawning further ones, dropping it from 6 down to 4. It's a quiet line in the changelog, but it matters for anyone running Copilot CLI's autopilot or multi-turn subagent flows unattended: a shallower default ceiling means a runaway delegation chain burns through fewer turns and less budget before it's capped, at the cost of some workflows that genuinely needed deeper recursion now needing to raise the limit explicitly.
The release also adds Canvas support for extension-driven interactions, enables multi-turn subagents by default for follow-up messages, turns on tool search for Claude Haiku 4.5 and above, and ships new plugins marketplace subcommands alongside a /voice devices command for persisting a chosen microphone across sessions. On the fix side, slash commands now match regardless of case and invalid settings.json files surface a clear startup warning instead of failing silently.
Codex CLI 0.144.6: GPT-5.6 Context Windows Corrected to 272,000 Tokens
OpenAI's July 18 patch for Codex CLI refreshes the bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrects their advertised context windows to 272,000 tokens. It's a metadata fix rather than a new feature, but incorrect context-window reporting is the kind of bug that's easy to miss and expensive when it bites: an agent that thinks it has more (or less) room than it actually does will misjudge when to summarize history, trim tool output, or split a task, degrading output quality in ways that look like model weirdness rather than a bookkeeping error.
This follows the 0.144.5 patch two days earlier, which tightened dangerous-command detection to catch more forced-rm variants and gives clearer explanations when Codex declines to run a command — part of the same guardrail-hardening line this Codex CLI series has followed all month.
Why It Matters
Copilot CLI's nesting cut is now the third independent move this month to rein in subagent sprawl: Claude Code added a default 200-call cap on subagent spawns per session in v2.1.212 on July 17, OpenCode shipped a configurable subagent_depth limit in v1.18.2 on July 15, and now Copilot CLI has halved its own default nesting ceiling. Three separate teams reaching the same conclusion in the same two-week window is a signal worth taking seriously — recursive subagent delegation has quietly become a real cost and reliability risk across the agentic coding category, not a problem unique to any one vendor's architecture. If your workflows lean on deep subagent chains in any of these tools, check whether you're now hitting a lower default ceiling than you were last week.
Sources & Further Reading
Comments
Share your thoughts and join the conversation
Leave a Comment
Keep reading.
Daily Dev Stack Release Audit — July 28, 2026: Redis, MongoDB, and Nuxt Rush Out Critical Security Patches
AI Coding News: July 28, 2026 — Model Context Protocol Ships Its Final 2026-07-28 Spec
