import

This skill uploads work history to the configured Mubit instance. It runs one Node process from the plugin directory, reads ~/.claude/projects — and, when asked, Codex's ~/.codex/sessions — and ingests what it finds. Nothing is sent unless --send is on the command line.

disable-model-invocation: true is not a formality here, and it is stricter than the reason activity and dashboard carry the same line. Those read. This one takes months of somebody's transcripts — every prompt, every command, every file they touched — and puts them on a server. Nothing in a conversation should decide on its own to do that. A person types it, or it does not happen.

Always dry-run first, and show them the number

node "${CLAUDE_PLUGIN_ROOT}/bin/import.mjs"

That is the whole command, and it sends nothing: a dry run is what happens when nobody says otherwise. It prints the scope it resolved and how many items it would send.

Relay that number and the scope before offering to send. "This would send 2,300 items from 41 transcripts across 3 runs" is a sentence somebody can answer. "Shall I import your history?" is not.

Then, only if they say so

node "${CLAUDE_PLUGIN_ROOT}/bin/import.mjs" --send

When it says ingest failed

The counts line is followed by the reason — ingest failed (<state>): <message>, the same sentence the plugin logs. unreachable with "no network access" means the shell the command ran in has no network: Codex's sandbox does this to an unapproved command, and the fix there is to run it with escalated permissions, never to change the endpoint. auth_failed is the key, which is /mubit-memory:auth. server_error is the instance. Relay the sentence, not the count: "failed 1" is not an answer to "why".

Scope

The default is the project you are in, plus every git worktree linked to it. A worktree is a different directory and therefore a different transcript folder to the host, but it is the same work, so importing one without the others would give a partial history that looks complete.