Codex 0.158 asks before elevated commands, adds MCP OAuth secrets
OpenAI's terminal agent now gates risky shell commands and reaches OAuth-locked MCP servers — plus Git 2.56's safer merges and a $0.15 open Qwen.

Copy markdown
Elevated commands now stop and ask
Codex 0.158.0 (Sept 28) turns on terminal-input approval by default for any command running with elevated permissions, so a privileged command pauses for your OK instead of running silently — and runtime-only grants no longer trigger needless reviews. The release shipped 141 commits from 53+ contributors.
Codex reaches OAuth-locked MCP servers
The same build connects to MCP servers that require pre-registered OAuth client secrets via `codex mcp add --oauth-client-secret`, and secures direct exec-server WebSocket connections with bearer tokens. Smaller wins: configurable copy-on-select and right-click paste in the fullscreen TUI that keep Markdown, and image edits that can request transparent backgrounds.
Git 2.56 makes finishing a merge safer
`git add --resolved` stages only files whose conflicts you've settled and warns on leftover markers; `git branch --delete-merged` bulk-cleans and new `git history drop` removes commits and replays their descendants. Merge-base got much faster (one monorepo 0.68s to 0.01s) and path-walk repack shrank a sample pack 558.5MB to 164.4MB — about 71% smaller.
Copilot CLI now reads your .claude/rules
Recent GitHub Copilot CLI builds (1.0.89-4 through -6) now read Claude Code rule files from `.claude/rules`, so shared repo conventions carry across both agents — plus PR-template checklists that preserve your checkboxes, unread-turn alerts, and auto routing-tier suggestions.
A 256K-context open Qwen for $0.15
Qwen3.8-Flash-Next — a 125B-total / 6B-active hybrid Gated-DeltaNet + sparse-attention MoE with open weights — is now served serverless on FlexAI at $0.15 input / $0.47 output per million tokens (cached input $0.0225), with a 256K context. 300+ quantizations including GGUF for Ollama and LM Studio are already up.