Workflows Snapshots

Workflow snapshots & version history

Snapshots are immutable, point-in-time captures of the entire workflow graph — nodes, edges, prompts, scripts, tools, and positions.

Key concepts

• Automatic creation: after every graph-mutating node edit (clay workflows nodes create|update|delete; Sculptor: edit_node) and when a run starts. data[0] is newest-by-createdAt, not necessarily the current graph. Sharing/binding a trigger onto an existing node does not snapshot — undo that create by deleting the trigger, not restore. Settings-only trigger update/delete also do not snapshot — revert or recreate the settings; do not delete the trigger and do not restore. graph format takes two snapshots when validation has no errors (pre-layout, then post-layout); details under Undo.
• Restore does not reshuffle the list: snapshots restore rewrites the draft but does not create or reorder snapshots. Duplicate hashes reuse the existing row without updating createdAt. After a restore, data[0] can be a later history entry that is not the current graph, so the next edit's data[1] is often that stale entry rather than the pre-edit graph. Compare the latest snapshot to the current graph (or find the list row that matches) before treating data[1] as the undo target.
• Content-addressed: Each snapshot has a SHA-256 hash of its contents. Identical workflow states deduplicate to the same hash
• Immutable: Once created, a snapshot's graph content never changes
• Run isolation: Runs are pegged to a specific snapshot. Editing the workflow doesn't affect in-flight runs

Draft-history vs published versions

The same snapshot table holds two roles:

Publishing marks a snapshot of the current draft as a numbered release; it does not create a separate store. See the workflows entry-point skill's publishing.md for draft vs live.

This skill is for undo/history (list / get / restore). Publishing is a separate command (clay workflows publish); there is no CLI to list only published versions.

clay workflows snapshots list projects only id, hash, createdAt, nodeCount, and edgeCount. It does not show version / name / which snapshot is live. Treat the list as newest-first history, not a release picker — after a restore it is not a linear undo stack.

CLI reference

If clay isn't on PATH or clay whoami fails on auth, run the setup skill.

List recent snapshots

clay workflows snapshots list <workflowId>

Returns { data: [...] }, newest first, each with id, hash, createdAt, and nodeCount/edgeCount. So data[0] is the most recent snapshot by createdAt, which may not be the current graph (see restore above).

Show a snapshot (whole graph, or one node)

clay workflows snapshots get <workflowId> <snapshotId>
clay workflows snapshots get <workflowId> <snapshotId> --node-id <nodeId>

Returns the full captured graph: nodes (with types, prompts, code, tools) and edges. --node-id narrows nodes to a single node (edges are left intact).

Diff two snapshots

There is no built-in diff. Fetch both and compare with jq — but this raw diff is the underlying mechanism, not the thing you show the user:

clay workflows snapshots get <workflowId> <oldSnapshotId> | jq '.nodes' > old.json
clay workflows snapshots get <workflowId> <newSnapshotId> | jq '.nodes' > new.json
diff <(jq -S . old.json) <(jq -S . new.json)

Translate the diff into a plain-language change summary rather than pasting the raw diff output — e.g. "Node Find Email's prompt changed; the edge Research → Draft was removed; a new Score Lead agent node was added." When the change alters the graph's structure (nodes or edges added/removed/rewired), render before/after Mermaid diagrams so the user can see it, not just read it (see the workflows entry-point skill's presenting.md).

The only built-in diagram command always renders the workflow's current graph:

clay workflows diagram <workflowId> | jq -r '.diagram'   # always the current graph

There's no diagram command for an arbitrary snapshot id, so hand-write a small node/edge list from that snapshot's nodes/edges for its side. Label the diagrams by the direction of the change: "before" is the state you're changing from, "after" is the state you're ending up with — and use the command above for whichever side is the current graph. When previewing a restore (see "Restore to a snapshot" below), the restore overwrites the current graph with the snapshot, so the current graph is the "before" and the snapshot you're restoring to is the "after".

Restore to a snapshot

clay workflows snapshots restore <workflowId> <snapshotId>

Restores the workflow to the exact state captured in the snapshot. This replaces all current nodes, edges, prompts, scripts, and tools. Restore is destructive and does NOT snapshot the current graph first, and does not move the restored snapshot to the top of list — data[0] can still be a later history entry. The pre-restore state is recoverable only if a later edit or run already triggered a fresh snapshot capturing it (snapshots are taken automatically after each graph-mutating edit and at run start). If the current graph has unsnapshotted changes you might want back, run snapshots list first and note the snapshot that matches the current graph, not blindly data[0].

Restore ≠ publish. Restore rewrites the draft only. It does not change which snapshot is live, and undoing an edit does not roll back live automation. If the user wants live triggers to match the restored draft, they must Publish again in the editor (see the workflows entry-point skill's publishing.md).

Before restoring, show the user what will change. Summarize the difference between the current graph and the target snapshot in plain language, and — when the structure differs — show before/after diagrams (see "Diff two snapshots" above), so the user can confirm the revert before it destructively overwrites the current state. After restoring, render the restored graph so they can see the result.

Common workflows

Undo the last edit

Every graph-mutating write schedules a snapshot after it completes — fire-and-forget, so the new row may lag. Do not assume data[0] is the current graph or that data[1] is the pre-edit graph.

Trigger create only snapshots when it allocates a new canvas node. Sharing an existing workflowNodeId is not in the snapshot hash — undo that create with clay workflows triggers delete (Sculptor: surface delete), not restore. Settings-only trigger update/delete also are not snapshotted: revert or recreate the settings; do not delete the trigger and do not restore.

graph format is two steps when validation has no errors: a pre-layout snapshot is awaited, then a post-layout snapshot is scheduled. Restore the pre-layout id (same nodes/edges, different positions). If validation had errors, the pre-layout snapshot is skipped — .data[1] may be an older unrelated snapshot; do not restore it. If the graph had no prior snapshot, wait until the post-layout row has landed before treating .data[1] as the undo target.

Compare snapshots get vs the current graph (clay workflows graph get --mode full; Sculptor: MCP read with mode: "full") to tell current from newest-by-createdAt. If no row matches, the post-edit snapshot may still be in flight — poll or restore .data[0] (step 4); do not search for a snapshot of the unsnapshotted current graph. Only use Bash(clay *) / Bash(jq *) — poll by repeating list, not a shell for/seq/sleep loop (those are outside this skill's allowed-tools).

clay workflows snapshots list <workflowId> | jq -r '.data[0].id'   # newest — not always current
clay workflows snapshots list <workflowId> | jq -r '.data[1].id'   # next older — not always pre-edit
clay workflows snapshots restore <workflowId> <snapshotId>

1. Find before_id before the edit: if .data[0] matches the current graph, before_id is .data[0].id. If not, walk the list until a row matches — that id is before_id. If none match, there is no restorable pre-edit snapshot; do not pass a null id to restore.
2. With before_id (node/trigger edits): after the write, restore before_id. Do not restore .data[1] as a proxy — after a restore that id can be a later history entry.
3. With before_id (after graph format): restore before_id only if that id matched the current graph before format. If you had no matching snapshot before format, call list until .data[0] matches the current (post-layout) graph, then restore .data[1].id only if it is the pre-layout graph (same nodes/edges, different positions). If validation had errors, that pre-layout row was never created — .data[1] can be an older unrelated snapshot; if so (or if .data[1] is missing), report no restorable pre-layout snapshot rather than rolling back more than the layout.
4. Without before_id (late undo): compare .data[0] to the current graph. If they match, restore .data[1].id only if it exists — and only when you know .data[0] was already the current snapshot before the last edit (usual linear history). If .data[1] is missing, report no restorable pre-edit state. If they do not match, walk the list for a row that matches the current graph. If an older row matches (post-restore), restore that id — not .data[1], which can roll the workflow forward. If no row matches, the post-edit snapshot has not landed (fire-and-forget lag): poll list a few times. If .data[0] then matches, follow the match case above. If still none, restore .data[0] (still the pre-edit graph in linear history). Do not look for a snapshot of the current unsnapshotted graph — it cannot exist until that async row lands. Do not pass a null snapshot id to restore.

To undo multiple edits, pick an older snapshot id from the list instead.

Remember: this only rolls back the draft. Live automation (if already published) keeps running the previously published version until the user publishes again.

Compare current state to a previous version

List snapshots, then diff two of them with the jq recipe above.

Review what a run executed

Since runs are pegged to snapshots, inspect the exact workflow state a run used by looking up the run's snapshot id and passing it to snapshots get.