Feature: Allow subagents to inherit parent conversation context

Problem

When spawning subagents via the Agent tool, each subagent starts with a completely blank context. The only way to give a subagent relevant context is to manually summarize/paste it into the prompt parameter.

Meanwhile, /fork (aka /branch) creates an interactive conversation branch that preserves the full parent history — but it's interactive only, not backgroundable.

There's no way to combine these: spawn a background subagent that inherits the parent conversation's context.

Use Case

When working on a complex investigation (e.g., analyzing what would break if moving a directory), the parent conversation accumulates significant context — file contents read, tool results, intermediate reasoning. If you want to spin off a parallel investigation that builds on that context, you currently have to either:

1. Use /fork and interact with it yourself (gets history, but blocks you)
2. Use the Agent tool and manually re-summarize everything into the prompt (backgroundable, but lossy and verbose)

Proposal

Allow the Agent tool to optionally inherit the parent conversation's transcript, similar to how /fork works but in a non-interactive, backgroundable way. Something like:

Agent(
  prompt="Now investigate X based on everything we've discussed",
  inherit_context=true,  // or fork_from="parent"
  run_in_background=true
)

This would effectively be a "background fork" — the subagent gets the full parent context as a starting point, then runs the new prompt autonomously.

Experiment Done

We verified the current limitations:

• Passing /branch as a prompt string to Agent → treated as plain text, no history inherited
• Passing a parent session ID via resume → fails with "No transcript found" (resume only works with subagent IDs)
• /fork works for interactive branching but can't be dispatched as a background task

Alternatives Considered

• Summarize into prompt: Works but lossy, especially for large contexts with many tool results
• Resume a subagent: Only resumes the subagent's own prior transcript, not the parent's

🤖 Generated with Claude Code