SendMessage for continuing spawned sub-agents is documented but not callable

The system prompt and the Agent tool result both tell the model to use SendMessage to follow up with a previously spawned sub-agent, but the tool is never actually exposed on the tool surface — so multi-turn sub-agent conversations are advertised but unusable.

Repro

1. Open a Claude Code session (Opus 4.7, 1M context build).
2. Have the model call the Agent tool, e.g.:

``
Agent({ subagent_type: "general-purpose", prompt: "Hello, how are you?" })
``

1. The tool result comes back with a trailer like:

``
agentId: af9219c00828602bd (use SendMessage with to: 'af9219c00828602bd' to continue this agent)
``

1. On the next turn, try to use SendMessage:

• ToolSearch with select:SendMessage → No matching deferred tools found
• ToolSearch with keyword queries (send message agent, continue spawned agent follow-up, +SendMessage) → returns Slack/QBO senders, RemoteTrigger, Monitor, etc., but no SendMessage.

1. Net effect: the only way to "follow up" with the prior agent is to spawn a brand-new one with Agent and re-pass all the context, which the Agent tool description itself warns about ("A new Agent call starts a fresh agent with no memory of prior runs").

Expected

Either:

• (a) SendMessage is actually wired up and ToolSearch surfaces it (ideally after the first Agent call so it doesn't clutter the default tool list), or
• (b) the system prompt and the post-spawn trailer stop mentioning SendMessage and instead document the real continuation mechanism (or its absence).

Actual

The model is told to use a tool that does not exist in the harness, wastes turns probing for it, and either has to give up or invisibly start a fresh agent (losing the prior conversation state without telling the user).

Why it matters

Multi-turn sub-agent conversations are useful for protecting the parent context window while iterating with a specialist — e.g., handing a research agent a follow-up question after it returns initial findings. Without SendMessage, every follow-up either pollutes the parent context or restarts cold.

Env

• Model: claude-opus-4-7 (1M context)
• Platform: darwin 24.6.0
• Claude Code CLI in terminal