[BUG] Cannot continue previously spawned subagent — SendMessage referenced but unavailable

Summary

The Agent tool description states:

To continue a previously spawned agent, use SendMessage with the agent's ID or name as the to field. The agent resumes with its full context preserved.

However, SendMessage is not available as a tool — it does not appear in the tool list, among deferred tools, or anywhere else. The Agent tool schema has no to parameter either. As a result, every Agent invocation creates a new subagent with a fresh context, making multi-turn interaction with a subagent impossible.

Steps to Reproduce

1. Configure a custom agent (e.g., git-committer) that requires user approval mid-flow
2. Launch it via the Agent tool — it returns a result with agentId: <id> and a note: use SendMessage with to: '<id>' to continue this agent
3. Attempt to continue the agent:

• SendMessage — not in the tool list
• Search deferred tools for SendMessage — not found
• Agent tool has no to parameter in its schema (description, prompt, subagent_type, isolation, run_in_background, model only)

1. The only option is to call Agent again, which spawns a new agent that has lost all prior context

Expected Behavior

Either:

• SendMessage should be available as a callable tool, or
• The Agent tool should accept a to parameter for continuing an existing subagent

Actual Behavior

Neither mechanism exists. The system prompt references a tool that is not provided. Each Agent call starts fresh, breaking workflows that require back-and-forth (approval flows, iterative refinement, etc.).

Environment

• Claude Code CLI v2.1.79
• Model: Claude Opus 4.6 (1M context)
• Platform: Linux (Mint aka Debian/Ubuntu)