Summary
Side-by-side digest of two agent-CLI releases shipped on 2026-11-06: Claude Code v2.1.319 and Codex CLI v0.167.0. The listed items were selected because each has an analogue in the "yoyo" agent CLI, making them useful deltas for cross-tool feature comparison.
Claude Code's release is small but behaviorally significant for permission handling and unattended runs: a deny-instead-of-prompt mode, deny-before-ask precedence, backgrounded MCP tasks, a script-safe remote-control exit-code fix, and Bash which flag guarding. Codex's release concentrates on supervising background work (an Agent Command Center, resumable multi-backend sessions), richer approval review (full inline patches), and infrastructure plumbing (GCP Vertex onboarding, bounded Code Mode output, per-chat MCP leases, hook ordering before policy checks).
Key Points
- Deny instead of prompt. Claude Code adds
permissions.denyInsteadOfPrompting: a tool call that would block on a permission prompt is instead denied and the turn keeps going — a fail-closed, non-interactive policy that is explicitly not the same as--yes(auto-approve). It applies across terminal sessions, subagents, and teammates. - Deny beats ask. An exact session-level "ask" rule is refused when a broader deny rule already covers that command; precedence moves to the deny.
- Backgrounded MCP tasks. MCP tool calls run as tasks are backgrounded and their results reported on completion;
CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUNDrestores blocking behavior. - Silent-zero exit fixed.
remote-controlused to exit silently with status 0 in scripts when a first-run question received no answer; it now explains the problem and exits 1. - Bash
whichguarding. The Bash tool now asks for permission before runningwhichwith flags outside a recognized lookup set. - Post-interrupt hooks.
PostToolUsehooks now fire (capped at 10s) for tool calls that finish after an interrupt.