v2.2.2 — what carries where a transcript cannot go

switchboard session brief <id>

What a session was for, small enough to paste into another one: the human's instructions, the files touched, the tool profile, where it ran — read off the transcript rather than retold by a model.

switchboard session brief 9c91c9ee-…

On the first real handoff that is 960 bytes out of 4,537,249.

It exists for the place a capsule cannot reach. A handoff lands a transcript on a filesystem, which covers anywhere a session can run a shell — your laptop, another machine, a cloud container. It does not reach a session running in Claude's own cloud, which accepts no transcript through any endpoint. A brief does, because it fits in a message.

The filtering is the substance. Claude Code writes tool results, hook output, notifications and slash-command echoes as user records too, so what looks like 285 user turns contains 14 things a human actually typed. Repeated asks collapse — the same instruction re-sent after an interruption is not a second thing wanted — and an interruption marker is not an instruction. --limit keeps the first ask and the most recent ones, because a long session's shape is what it was for and where it got to.

The limits, written down

docs/claude-code.md now states plainly what a handoff can and cannot reach, checked against Claude Code 2.1.259: --cloud <session_id> attaches to an existing cloud session, --environment creates a new one on a self-hosted pool, neither takes a transcript; /move-to-cloud and /teleport are interactive-only and the teleport path seeds a summary rather than the conversation. It also documents the two things that do work in that direction — consulting the transcript on disk from a cloud container, and sending a brief.

Additive. Nothing that worked in 2.2.1 changed.