[FEATURE] Interactive step-into mode for Task subagents

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When Claude Code spawns a Task (subagent), the user loses all control until it returns. You specify a task, wait, and hope the results match what you wanted. If they don't, you start over with a better prompt.

This fire-and-forget model fails for complex tasks because users can't:

• Course-correct mid-exploration ("not that, focus on X instead")
• Guide the agent's investigation in real-time
• Control what ends up in the summary returned to the parent

The current design conflates context isolation (a technical constraint) with loss of control (a UX problem). These are orthogonal—you can have isolated context while still letting users steer.

Proposed Solution

Add a "Step into" option when a Task is invoked:

[Approve]  [Reject]  [Step into]

• Approve = fire-and-forget (current behavior)
• Step into = enter the subagent conversation interactively

When stepping in:

• User enters an interactive session with the subagent
• Context remains isolated (separate thread/memory)
• Prompt shows depth: [general-purpose:1] >
• User can guide the exploration, ask follow-ups, redirect
• /return exits and sends a summary back to the parent context
• /summary lets user view/edit what gets returned
• /context shows position in the conversation stack

Alternative Solutions

Currently users work around this by:

• Avoiding subagents entirely (losing the context isolation benefit)
• Re-running subagents multiple times until they get lucky
• Writing extremely detailed prompts trying to anticipate every direction

We implemented this feature for deepagents-cli (LangChain's agent framework) as a proof of concept:
PR: https://github.com/langchain-ai/deepagents/pull/829**

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

1. I ask Claude to research the authentication system in a codebase
2. Claude spawns a Task subagent to explore
3. Instead of "Approve" (fire-and-forget), I select "Step into"
4. I'm now in an interactive session with the subagent
5. Subagent: "Found 3 auth approaches—JWT, sessions, OAuth. Which to focus on?"
6. Me: "JWT only, ignore the others"
7. Subagent explores JWT, I guide it to specific files
8. Me: "Add your findings to the summary"
9. Me: /return
10. Back in main context, parent agent receives the summary I helped craft

Additional Context

We built a working implementation (~450 lines) for deepagents-cli:

• Context stack tracks nested conversations
• Summary file at ~/.deepagents/{agent}/branches/{id}/summary.md
• Subagent knows about summary file, can edit it when asked
• Parent agent automatically processes summary on return

Context isolation is about memory. Control is about agency. They were never the same thing.