Agent-aware tabs

Four states, one glance

An agent tab is always in one of these states. Idle shows no pill.

• Working: a mint, pulsing badge.
• Needs you: an amber badge, pill and underline, for example when the agent waits for a permission or an answer.
• Done and Error: when a turn finishes or the agent stops with an error.

The sidebar sums it up per workspace ("1 agent needs you · 2 working"), and the window title gets a count such as (1) car-rental — SplitNest, so it shows in the taskbar.

Supported agents

Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Cursor Agent, Aider, OpenCode, Qwen Code and Amp get their logo and name on the tab. If the agent reports a session title, the tab shows it (for example Fix login redirect · Claude Code); a name you gave the tab still wins. Agent logos are trademarks of their owners.

How SplitNest knows

1. The agent reports itself through the Program Status Protocol (OSC 7501). SplitNest answers the protocol's feature query, which Claude Code waits for before reporting; Codex reports on its own. This was checked with Claude Code 2.1.316 and Codex 0.166.0.
2. Process detection in PowerShell tabs: a hidden helper looks for the agent's process under the tab's shell, every 3 seconds while something changes and otherwise every 12 seconds. Command lines never leave the main process and are not logged.
3. Fallbacks for an identified agent that does not report: a terminal bell means "Needs you?" and a spinner in the title or recent output means "Working?". Output, bells or titles alone never make a program count as an agent.

No agent settings or hooks are changed.

Reported or estimated: you can tell

A state the agent reported is shown solid. A guess is shown dashed with a "?" ("Needs you?", "Working?"), reads "may need you" in the sidebar, is not counted in the window title, and the tooltip says what it is based on. Once an agent has reported, nothing about it is guessed any more for that run.

Opening or typing into a tab clears its notification and a Done or Error you have now seen. It never clears what the agent is waiting for: a permission request stays until the agent itself reports that it continues.

When the agent exits

When the agent's process ends, the shell's prompt or title comes back, or the session ends, its logo, state and session title all go, so the next program in that terminal starts clean.

Limits

• WSL tabs cannot use process detection, because Linux processes are not visible to Windows. They rely on the agent's own reports or Claude Code's exact title.
• Claude Code's first-run questions (for example "Claude in Chrome extension detected") are not reported by Claude, so they show as idle.

Agents in the user guide →