Codex

Control a Codex thread from Grove

Grove starts codex app-server --stdio, initializes its JSON RPC connection and creates a thread. The app server owns every turn inside that thread, and Grove drives it from the web dashboard, the TUI or MCP.

• A message either starts a turn or steers one. Sending while Codex is idle starts a turn. Sending while it is busy steers the turn already running.
• Steering names the turn it expects. If the turn moved on, Codex rejects the request rather than applying your input to work you were not looking at.
• Interrupt targets the current turn. With no turn running there is nothing to cancel, so Grove tells you instead of sending a request that means nothing.
• Approvals come to you. Grove answers supported command and file change approvals with accept or decline.

What Codex does not offer

• Delivery is queued, not confirmed. Codex acknowledges submission, which is weaker evidence than Claude Code's replay. Treat a sent message as handed over rather than read.
• A model change is recorded, not validated. The app server accepts the setting before checking it, so an unknown model surfaces on the next turn rather than at the moment you switch.
• Approval breadth is narrow. Session wide grants, cancellation and separate permission grant requests are unsupported. There is no general plan approval control.
• Questions need plan mode. request_user_input is only available in Codex plan collaboration mode.

Set the approval policy deliberately

Codex app server policy comes from Codex configuration, not from a terminal shortcut. A flag such as --yolo on the command line does not configure thread approvals.

Set approval_policy and sandbox_mode in Codex's own configuration, keeping the permissions your task actually requires.

Configuration

{
  "agents": [
    { "name": "codex-fast", "command": "codex --model gpt-5.3-codex-spark", "kind": "codex" }
  ]
}

Codex mints its own thread identifier, so Grove discovers the session rather than naming it at launch.

See also

• Agents: choosing between native, headed and headless.
• Claude Code and OpenCode: the other native adapters.
• Agent activity and sessions: what the transcript unlocks.