Feature: Mid-task message injection for real-time redirection

Summary

Allow users to send messages to Claude Code while a task is actively running, so Claude can read and incorporate the redirect mid-execution — without requiring the user to force-stop via Escape.

Current behavior

• Messages typed during task execution are queued and only processed after the current turn completes.
• The only interrupt mechanism is Escape (hard stop), which cancels the current tool call entirely.
• No soft-redirect path exists: user cannot steer Claude mid-action without fully stopping it.

Desired behavior

User types a message + Enter while Claude is executing. Claude reads it at the next natural checkpoint (e.g., between tool calls) and self-redirects without a full stop — similar to how OpenAI Codex handles in-flight message injection.

Why this matters

• Long-running tasks often go in the wrong direction early; catching it mid-task saves significant time vs. waiting for completion or hard-stopping and restarting.
• Escape-then-retype workflow loses context and disrupts Claude's chain of thought.
• Competitors (Codex) already support this interaction model.

Proposed behavior

1. User sends message while task runs.
2. Claude finishes current atomic action (single tool call).
3. Claude reads injected message and adjusts plan before next tool call.
4. No full stop required; execution continues with updated intent.

References

• OpenAI Codex supports mid-task message injection as a comparable implementation reference.