Skip to main content
subagent CX Codex

task_start_keys — Claude Code subagent: what it does, how to install

The task_start_keys Claude Code subagent types text and presses keys in a task’s terminal tab after its agent stops there. It can answer a shell prompt or move past the screen’s current point.

MIT ★ 395 updated 2026-11-16

Install into Claude Code · Manually: put the files in place All methods and platforms ↓
.claude/agents/task_start_keys.md
Open on GitHub

What task_start_keys does

The subagent types supplied text into a task’s terminal tab, then presses keys specified by name. Text is limited to 32 characters and is entered without Enter; the key list can contain up to 8 items. The task ID and a reason for the action are required.

Read the tab with task_start_screen before typing, then read it again afterward. The text, keys, and reason are saved in the task history. This subagent does not start an agent; use task_retry to start one again.

What it helps with

  • answer a prompt waiting in the task’s shell
  • move past the point where the agent’s screen stopped
  • send a response or control keys to a terminal tab

Example prompts for the agent

  • Answer the shell prompt in the task terminal with y.
  • Type continue in the task’s terminal tab, then press enter.
  • Send ctrl+c to the task terminal to interrupt the current process.

How to install task_start_keys

Manually: put the files in place

.claude/agents/task_start_keys.md

For all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention

Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean

Questions about task_start_keys

How do I install task_start_keys in Claude Code?

Add .claude/agents/task_start_keys.md to the project.

What does task_start_keys do?

It types supplied text and presses named keys in a task’s terminal tab. The tab is read before and after, and the actions are saved in the task history.

Does task_start_keys restart the agent?

No. It acts in the existing terminal tab; use task_retry to start the agent again.

Source text

task_start_keys.md · MIT
Type into the terminal tab a task's agent did not start in, while its computer keeps that tab and no agent runs there: text as is (no enter, at most 32 characters), then keys by name ("enter", "esc", "ctrl+c", "y", "n", "up", "down"…, at most 8). Use it to answer the question the tab's shell waits on, or to get past where the agent's screen stopped; read the tab first with task_start_screen and again after. The keys, the text and your reason are kept with the task's history. It starts nothing: start the agent again with task_retry.

```json
{
  "type": "object",
  "required": ["todo_id", "reason"],
  "properties": {
    "todo_id": {
      "type": "string",
      "description": "the task's id, as the 督办 message gives it"
    },
    "text": {
      "type": "string",
      "maxLength": 32,
      "description": "text to type as is, before the keys"
    },
    "keys": {
      "type": "array",
      "maxItems": 8,
      "items": { "type": "string" },
      "description": "keys to press after the text, by name"
    },
    "reason": {
      "type": "string",
      "description": "why these keys, in the user's language, one sentence"
    }
  }
}
```

© nwverse · MIT · nwverse/persona

Similar subagents

All subagents for Agent setup: memory, context, orchestration →

From the same set61

The whole set →
question_asked_later The question_asked_later Claude Code subagent sends a question to the user as a card on their phone. It finishes unrelated work and ends the turn without guessing while the answer is pending. CX Codex ★ 395repo stars launcher_repair launcher_repair — a subagent for Claude Code and Codex repairs agent CLI launchers on a specified machine. It tests each launcher after making changes. CX Codex ★ 395repo stars task_spawn The task_spawn Claude Code subagent starts a separate agent for part of an active assignment. The new agent joins the same to-do list and reports to the same supervisor. CX Codex ★ 395repo stars task_leader The task_leader subagent for Claude Code splits work among child agents and checks their results against defined completion criteria. It handles review decisions for each delegated task. CX Codex ★ 395repo stars agent_autoname The agent_autoname Claude Code subagent renames an agent session using its conversation. It runs only after an explicit Auto Rename request. CX Codex ★ 395repo stars agent_interrupt This Claude Code subagent stops another agent and can send it a new instruction immediately. Use it to interrupt a stalled loop, repeated failure, or work heading in the wrong direction. CX Codex ★ 395repo stars task_handoff The task_handoff Claude Code subagent sends part of an assigned task to a separate session through Persona. The handoff stays attached to the same to-do and supervisor. CX Codex ★ 395repo stars leader_progress The leader_progress Claude Code subagent adds direction updates to a work log. It reports completed goals and scheduled check-ins, rather than every intermediate step. CX Codex ★ 395repo stars

Project sponsors

Companies that keep this analytics open