What is verified here, and what is not

Written 12 August 2026. Kept honest so a later reader knows which claims to lean on.

Confirmed

The agent-to-agent primitives exist and behave as documented. Read directly from the live SendMessage and ListAgents tool schemas in a running session, not from release notes. The contracts in cross-session-messaging.md — addressing by name, automatic delivery, <cross-session-message from="..."> wrapping, the enqueue-and-drain model, the [ref] disambiguation rule, the per-session permission boundary — come from those schemas.

The release facts. 2.1.224 on 7 August 2026, macOS and Linux, local socket for same-machine peers, text-only messages, --name / /rename with folder-name fallback, the 2.1.225 fix to false "Message sent" reporting. Corroborated across the Claude Code docs and several independent write-ups.

The prior art. sideclaude, spawn-planning-repo and new-claude-at were read from Claude-Hopper's published README on 12 August 2026.

interrupt-claude is unpublished. gh repo view danielrosehill/interrupt-claude-plugin returns "Could not resolve to a Repository" — it exists locally with a single scaffold commit and no remote. Its own plugin.json already carries a repository URL for that address, which will resolve only once it is pushed.

Not verified

No breakout has been run end to end. The procedure is a design. The specific failure modes it claims to prevent are reasoned from the primitives' constraints, not observed.

The skills have never dispatched. Their trigger descriptions are untested against a real session, and they overlap with Hopper's sideclaude — which is exactly the condition under which model dispatch gets unpredictable.

The socket permissions claim (srw-------) is second-hand from a single report and was not checked directly.

Whether the pattern survives contact with more than two concurrent tracks. interrupt's routing notes already observe that beyond about two parallel delegates the bottleneck stops being the machine and becomes the user's attention. Breakout spawns full sessions, which are heavier than subagents, so that ceiling is probably lower here. Untested.

First things to check when it is used for real

1. Does the spawned session actually appear in ListAgents promptly, or is there a registration lag that makes the handoff race the spawn?
2. Does the folder-name default produce collisions in practice across sibling repos in the same topic group?
3. Does the receiving agent, following breakout-adopt, start work without asking an opening question?