Claude Code

Control a Claude Code session from Grove

Grove starts claude -p with JSON streaming on stdin and stdout, then holds that connection open between turns. You send messages, interrupt work and answer structured questions without attaching to the terminal UI.

• A message is confirmed, not assumed. Every message you send carries an ID, and Grove waits for Claude Code to replay that ID before marking it delivered.
• Interrupt stops the current turn. The running tool call ends and its result is rejected, so a long command does not have to finish before you redirect the work.
• A model change is validated first. Grove asks the provider to confirm the model before reporting success, and the new model applies from the next turn.
• Structured questions come to you. Grove answers AskUserQuestion requests from the web dashboard, the TUI or MCP, so a waiting agent does not sit idle until you open its terminal.

What needs the terminal twin instead

Ordinary tool permissions and plan approvals cannot be answered through the native adapter. Pick claude-terminal when your workflow depends on those dialogs.

• Plan approval is a positional dialog. Approving a plan means choosing one of its rows, which is a decision the interactive UI owns.
• Permission prompts stay in the agent. Grove adds no permission gate of its own, so the agent's approval policy still governs every tool call.
• Pause is unavailable on a native session. The worker holds the conversation, so use Respawn to recover rather than Pause and Resume.

Exact status with hooks

Grove derives state from the transcript by default, which is accurate and slightly behind. Managed hooks report each lifecycle change as it happens, so a permission prompt reads as blocked rather than quiet.

{ "hooks": { "enabled": true } }

Grove launches claude_code agents with an extra --settings file and leaves your own .claude/settings.json untouched.

Configuration

{
  "agents": [
    { "name": "code-agent", "command": "claude --model claude-opus-4-5", "kind": "claude_code" }
  ]
}

A custom named Claude agent must declare kind: "claude_code" itself. An omitted kind silently defaults to generic, which leaves the dashboard with no transcript to read.

See also

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