Enable agent teams (SendMessage) by default

[pufit]

What

Adds agent.agent_teams (default true), which sets CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 for the CLI subprocess.

Why

The CLI gates the SendMessage tool behind that flag, but the Agent tool advertises it regardless — its own description says "Use SendMessage with the agent's ID or name to continue a previously spawned agent with its context intact", and every spawn result repeats it (use SendMessage with to: '<id>'). With the flag off, the model is told to resume sub-agents with a tool that was never registered: it tries, and it fails. Upstream report: anthropics/claude-code#35240

Nerve loads no settings files (setting_sources=[]), so ~/.claude/settings.json can't carry the flag the way it does for CLI/extension users — the env dict built in the claude backend is the only route in.

Verified

Probed with a standalone SDK client using Nerve's exact option shape (setting_sources=[], explicit env, stream-json transport), CLI 2.1.207:

	SendMessage in init tools	model's attempt
flag unset (today)	❌	replies TOOL_MISSING
flag set	✅	loads the schema via ToolSearch and calls it

The raw claude -p path lists SendMessage in the init handshake either way, so the SDK option shape had to be reproduced to see the real gate.

Notes

• Teammates stay opt-in per turn — the model still has to spawn them — and each costs a full context window. The flag only makes the tool exist.
• Known upstream limitation: the CLI cannot restore in-process teammates on resume. A session whose client is recycled (idle timeout, restart, crash retry) comes back without them, and the lead may message teammates that no longer exist. Documented next to the option.
• Hot-reloadable: applies to sessions started after POST /api/config/reload, same seam as the rest of agent.*.
• Backend-only; no frontend change. 3029 passed.

Generated by Nerve