Arm Overlord and launch in one command
At a normal shell prompt, choose the provider you want to run and add overlord:
codex overlord
claude overlord
grok overlord
PorkiCoder consumes the exact two-word meta-command, confirms that Overlord is armed, and then launches the allowlisted provider executable. Provider-supported flags can still be used through the separate activation path described below.
The forgotten-permissions use case: if Codex is already deep into a task and you forgot to enable broader permissions at startup, Overlord can supervise the current session. You do not need to kill the agent and lose its working context.
Turn it on during a running session
Activate the supervisor
Type the exact command /overlord, press the terminal shield, or use Overlord's control while a supported provider is in the foreground.
Let the agent continue
When a supported approval menu appears, local rules identify the current provider surface and choose a proven affirmative action. One-time or turn-scoped choices are preferred when the provider offers them.
Take over whenever you want
Any human keystroke invalidates an in-flight autonomous decision. Overlord also waits if the terminal foreground changes to a shell, editor, pager, or another process.
The exact /overlord line is recognized only after Enter. Paths, URLs, provider slash commands, and longer prompts containing the word “overlord” continue to the terminal normally.
What Overlord handles
Deterministic local rules recognize supported command, edit, network, permission, and MCP approval shapes from Claude Code, Codex, and Grok Build. If the same prompt remains after a write, Overlord can retry after a short acknowledgement window.
In version 2.17.74, a long Codex “don't ask again” option can fill the screen without stranding a complete approval menu. Overlord recognizes the exact command choices and confirmation footer, then selects Yes, proceed (y). Incomplete menus still wait for you.
Version 2.18.5 recognizes Codex MCP approval headings even when they wrap across lines and chooses the one-time Allow option on supported menus. Automatic cursor and focus reports no longer interrupt that decision; your actual keystrokes still take priority. Forms that need an answer still wait for you.
Agent questions are left for you. Stale surfaces, secrets, settings, browser completion, payment flows, and ordinary idle composers remain untouched.
Overlord reads xterm's logical live screen even while you inspect scrollback. Scrollback rows are never evaluated and cannot grant approval authority. If a second approval arrives while the first is settling, PorkiCoder immediately rescans the replacement screen.
Review the log and stop cleanly
Version 2.18.5 restores the classic compact OVERLORD overlay, with its portrait, provider, status and STOP control. The side log popup and goal, privacy and remote-access indicators remain available.
Each successful autonomous approval creates a compact speech bubble. Press LOG to open the bounded session history and see the recent actions Overlord submitted.
Disable Overlord with any of these controls:
- Type
/overlordagain. - Press the terminal shield.
- Press Overlord's STOP button.
- Exit the provider or close the terminal session.
The approval history is cleared when Overlord stops.
Understand the trust boundary
Overlord is an explicit auto-approval mode. After a screen is proven to be a provider authorization request, its content does not receive an additional risk veto. Broad permissions, destructive commands, production actions, network activity, and MCP tool calls can therefore receive approval.
Enable it for sessions you already trust. The provider sandbox and macOS are the containment boundary. Git can recover many tracked file changes, while external side effects can require separate recovery.
Approval snapshots are evaluated locally and never sent to Gemini or another model. Overlord has no question-answering path or free-text encoder: agent questions always wait for you. Snapshots stay out of chat history.