Push a file or directory from the node to a write-enabled connection

Push a file or directory from a RUNNING node to your own storage through a stored org connection with the "write" verb (list_connections shows verbs): the platform exercises the access on the host - never in this chat, never inside the node. dest must sit inside the connection's scope. S3 family (kind "s3-keys"): dest is an object key (e.g. "s3://acme-ml/results/run7/model.tar") and lands as ONE object (a directory is tarred; extract with tar -xf). Your org's own storage (auto-minted storage/<bucket>, kind "gmn-storage"): dest is an object key like s3, OR a '/'-terminated PREFIX - a directory then lands as a mirrored TREE of objects, one per file, committed as they land (get_export shows files_done/files_total). The connection names the bucket, so a dest here may be bucket-relative or full: "corpus-v3/" = "s3://datasets/corpus-v3/". Tree re-runs RESUME: objects already landed with matching size are skipped, never re-sent. Landed bytes count against your storage quota and bill as stored bytes. Hugging Face (kind "huggingface"): dest is "owner/name[@revision][:path/in/repo]" (e.g. "acme/ckpts@main:runs/run7/"; "datasets/owner/name" for dataset repos) and a directory lands as individual files under the path in ONE commit - model checkpoints stay loadable, never a tar; the repo must already exist on the Hub. Returns {export_id} immediately; poll get_export(node, export_id). A stopped node is a refusal, never a wake. Export time is node time: you're billed while it streams, and the transfer holds the node out of grace. Size is bounded (the refusal names the numbers; s3 and platform-storage exports STREAM as multipart, so big checkpoints fit - the Hub arm bounds any single FILE at one PUT, so shard big checkpoints or push them to s3; tree members too are single-PUT bounded). Every call is recorded to your org's audit log (connection, node, source, dest - never any credential). For a download link instead of a push, use export_file.