Codex 0.105.0 Ships Voice Input, Sleep Prevention, and a Complete Subagent Overhaul

Codex 0.105.0 dropped yesterday and it's the largest single update OpenAI has shipped for its CLI coding agent. The changelog reads like a laundry list of quality-of-life fixes, but buried in there is a complete rebuild of the multi-agent system that changes how Codex handles complex, multi-step tasks.

TL;DR

Voice input - hold spacebar to dictate prompts directly in the terminal
Theme picker - /theme command with live preview and syntax-highlighted diffs
Sleep prevention - Codex can now prevent idle-sleep on Linux, macOS, and Windows
Plan Mode - easier reasoning selection flow for planning agent tasks
Subagent overhaul - nicknames, visual cleanup, configurable spawn depth, custom role definitions
CSV batch spawning - fan out agent work from a CSV with progress tracking and ETA

Voice Dictation

Hold spacebar, talk, release. Your speech gets transcribed directly into the TUI prompt. This uses the Wispr voice transcription engine and works on macOS and Windows today.

To enable it, add this to your Codex config:

[features]
voice_transcription = true

Linux support isn't available yet. The feature is opt-in, which is the right call - accidental spacebar holds in a terminal tool would be annoying without an explicit toggle.

TUI Improvements

Theme Picker

The new /theme command opens a live preview picker. Codex now syntax-highlights fenced code blocks and diffs inline, and the diff colors adapt to whether you're running a light or dark terminal. This is one of those changes that sounds minor but makes a real difference when you're staring at diffs for hours.

New Commands

/copy - grabs the latest assistant response to your clipboard
/clear and Ctrl-L - clear the screen without losing conversation context
Long wrapped links are now clickable - a fix that was long overdue

Plan Mode

The reasoning selection flow in Plan Mode got a rework. You can now customize which reasoning strategy Codex uses when planning multi-step tasks. The UI for selecting between reasoning approaches is cleaner and more intuitive than the previous version.

The Subagent Overhaul

This is the headline feature for anyone running multi-agent workflows. The entire subagent system has been rebuilt.

Readable Names and Visual Cleanup

Sub-agents now get nicknames instead of opaque thread IDs. The agent picker interface is cleaner, and child-thread approval prompts are visible inline rather than buried in a separate view. When you run /agents, you see agent names, their roles, and "dead agents" - threads that completed or failed - for full auditability of what happened during a session.

Configurable Spawn Depth

Previously, subagents were limited to a single layer. 0.105.0 introduces a max_depth parameter that allows multi-layered subagent hierarchies. A top-level agent can spawn sub-agents that themselves spawn sub-agents, enabling genuinely complex delegation patterns.

Custom Role Definitions

You can now define custom subagent roles in your configuration. Instead of every spawned agent sharing the same generic prompt, you can assign specialized roles - a "reviewer" agent, a "test writer" agent, a "refactoring" agent - each with its own system prompt and constraints.

CSV Batch Spawning

The new spawn_agents_on_csv capability lets you fan out work from a CSV file with built-in progress tracking and ETA. Feed it a spreadsheet of tasks, and Codex spins up agents for each row, tracking completion across the batch. This is aimed squarely at teams processing large volumes of similar tasks - migration scripts, bulk refactors, test generation across modules.

Sleep Prevention

Codex can now prevent your machine from entering idle sleep while a session is active. This works on Linux, macOS, and Windows.

It's a small thing, but anyone who has lost a long-running agent session to a laptop lid close or a power management timeout knows the pain. The feature is automatic when enabled.

Approval and Permissions

Two new approval controls landed:

Extra sandbox permissions - commands can now request additional sandbox capabilities beyond the default set
Granular rejection - you can auto-reject specific approval prompt types without disabling the approval system completely

This gives you finer control over what agents can do without the all-or-nothing choice of previous versions.

Bug Fixes

The fix list is long. The highlights:

Issue	Fix
Long links broken when wrapped	Now clickable across line breaks
Enter pressed during streaming freezes follow-up	Fixed
`@` triggers file picker in commands like `npx -y @scope/pkg`	Parsing improved
WebSocket disconnects kill thread listeners	Listeners now survive disconnects
Linux sandbox missing `/dev` nodes	Minimal `/dev` now provided
`js_repl` kernel failures	Clear error reporting, clean recovery

Where It Falls Short

Voice transcription being macOS/Windows-only at launch is a notable gap for a tool whose core audience skews heavily Linux. The config-based opt-in also means most users will not discover the feature unless they read the changelog.

The subagent depth control is powerful but has no built-in safeguards against runaway spawning. If you set max_depth high and your agents are trigger-happy spawners, costs can escalate quickly. There is no budget cap or token limit tied to the depth parameter yet.

The CSV batch spawning is useful but doesn't support error recovery mid-batch. If a spawned agent fails on row 47 of 200, you start the whole batch over. Resumable batches would make this production-ready; right now it's best suited for tasks where individual failures are tolerable.

Finally, the theme picker and syntax highlighting are welcome, but Codex still doesn't match the visual polish of Claude Code's terminal interface or Cursor's IDE integration. The gap is narrowing, but it is still there.

Sources:

Codex 0.105.0 Release Notes - GitHub
Codex Changelog - OpenAI Developers
Codex CLI Features - OpenAI Developers
Codex by OpenAI Release Notes - Releasebot