# Snapshot Backups

> Use when backing up a Cognigy project before changing an existing AI Agent, rolling a project back to a previous state, undoing agent changes, or working with Cognigy Snapshots — create, restore, list, delete.

- Skill: `cognigy/snapshot-backups` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cognigy/snapshot-backups`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cognigy/snapshot-backups/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Cognigy (https://skillmd.com/u/cognigy)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/cognigy/snapshot-backups

---


# Snapshot Backups and Rollback

Use `manage_snapshots` to take a restorable backup of a Cognigy project and to roll
it back. A Snapshot is an immutable copy of a **whole project**.

## Read this before offering a backup

- A snapshot covers the **entire project** — every AI Agent, Flow, Connection, LLM,
  Lexicon, Extension, Function, Playbook, Goal, Snippet and Locale in it. Restoring
  reverts **all** of them, not just the agent you are working on.
- A snapshot does **NOT** contain:
  - Endpoints (and their API keys)
  - Knowledge AI — stores, sources, chunks, connectors
  - Intent Trainer learning sentences
  - Analytics data, contact profiles, logs
  - Other snapshots and packages
- So for an agent that uses knowledge, a snapshot is **not a complete backup**. Say
  this in one line when you offer it, rather than letting the user believe otherwise.
- Restoring is **irreversible** and deletes resources before recreating them.
- Snapshots cannot be renamed or edited after creation.

## Supported workflow

### Back up before changing an existing agent

The server enforces this **once per project**, so you do not have to remember it.
The **first** attempt to change an existing agent in a given project —
`update_ai_agent`, `create_tool`, `update_tool`, a mutating `manage_flow_nodes`, or
`delete_resource` — is **held**: it changes nothing and returns
`error: "backup_not_offered"`.

When you see that error:

1. Do **not** report the change as done. Nothing happened.
2. Ask the user, in one short line, whether they want a restorable backup first. If
   the agent uses knowledge, add that Knowledge AI is not covered.
3. If they accept:
   - `manage_snapshots { operation: "create", projectId: "<projectId>", label: "pre-persona-update" }`
   - Wait for `created: true`.
4. If they decline:
   - `manage_snapshots { operation: "decline", projectId: "<projectId>" }`
5. **Retry the held call.** It now goes through, and nothing else is held **for that
   project**.

Because the gate is per project, a later change to a **different** project can be
held again in the same session — that is correct behaviour, not a malfunction. An
answer given for one project says nothing about another, so ask again for the new
one. (When the server cannot tell which project a call targets, it falls back to
holding once for the session as a whole.)

The gate does not fire for an agent or project created in this same session — there
is no prior state to roll back to — and never for read-only operations.

If you do not have the `projectId`, read it from
`get_resource { resourceType: "agent", id: "<aiAgentId>" }`, which returns it.
If `create` returns `error: "snapshot_limit_reached"`, follow *At the snapshot limit*
below. If it returns `error: "task_status_unknown"`, the outcome is unknown: poll `read_task` before saying anything, and do NOT create a second backup. If it returns `pending: true`, the backup does **not** exist yet — poll
`read_task` until it is done before changing anything, and note that a pending
create does not satisfy the gate.

`label` is a short reason, not a name. The plugin builds the name itself as
`[AI Backup] v<N> <label> — <timestamp>` so the backup is always identifiable and
always uniquely named. `<N>` only counts up within a project, so you can refer to a
backup as "v3" rather than by timestamp; `list` returns it as `version`.
Numbers are never reused within a session, but two sessions that both start from
the same list can mint the same number. If `list` ever shows two backups with the
same `version`, identify the one you mean by its timestamp or id, not by "v<N>".

### Roll back to a backup

1. `manage_snapshots { operation: "list", projectId: "<projectId>" }`
2. Pick the snapshot. `isPluginBackup: true` marks the ones this plugin created.
3. If the target is **not** the backup this session just created — an older backup,
   or a human-made snapshot — the project's current state is about to be destroyed
   with nothing to come back to. Offer to snapshot the current state first, and
   create it if the user agrees and a slot is free. Skip this when the restore
   simply reverts this session's own changes to the backup it just took.
4. Get a preflight — this changes **nothing**:
   - `manage_snapshots { operation: "restore", projectId: "<projectId>", snapshotId: "<snapshotId>" }`
5. Show the user the returned `warnings` and `notRestored` lists, and the snapshot's
   name and age. Ask for explicit agreement.
6. Only after they agree:
   - `manage_snapshots { operation: "restore", projectId: "<projectId>", snapshotId: "<snapshotId>", confirm: true }`
7. Afterwards:
   - Every resource id in the project has changed. Re-list agents and flows; do not
     reuse any id from earlier in the conversation.
   - Tell the user to check Endpoints assigned to non-primary Locales in the Cognigy
     UI — those are marked with a red dot and need manual repair.
   - Knowledge AI was not restored.

### At the snapshot limit

A project holds a limited number of snapshots (10 by default, configurable per
installation). `create` pre-checks this and creates nothing when the project is full.

1. `create` returns `error: "snapshot_limit_reached"` with `count`, `assumedMax`,
   `deletableBackups`, and `oldestDeletable`.
2. If `oldestDeletable` is present, ask the user whether to delete that backup to make
   room, naming it. Only if they agree:
   - `manage_snapshots { operation: "create", projectId: "<projectId>", label: "<why>", confirmDeleteOldest: true }`
   - This deletes the **oldest plugin-created backup** and then creates the new one.
     The response reports it as `deletedToFreeSlot`.
3. If `deletableBackups` is empty, **stop**. The plugin never deletes a human-created
   snapshot. Tell the user to delete one themselves in the Cognigy UI under
   **Deploy > Snapshots**, then retry.

### Delete a backup the plugin created

1. `manage_snapshots { operation: "delete", projectId: "<projectId>", snapshotId: "<snapshotId>" }`
2. If it returns `error: "not_a_plugin_backup"`, the snapshot was created by a human.
   Do not try to work around it — tell the user to delete it in the Cognigy UI.
3. If it returns `inUseByEndpoint: true`, an Endpoint is serving that snapshot. The
   user has to point the Endpoint at a different snapshot first.

## Operations

### `list`

Lists the project's snapshots with `isPluginBackup` on each, plus the current count.

Required:

- `projectId`

Optional:

- `limit`
- `skip`

Returns `count`, `assumedMax`, `atLimit`, `snapshots`, `oldestDeletableBackup`.

### `create`

Creates a backup snapshot of the project and waits for the task to finish.

Required:

- `projectId`

Optional:

- `label` — short reason, used inside the generated name. Do not pass a full name.
- `confirmDeleteOldest` — only after the user agreed to free a slot
- `waitForCompletion`
- `timeoutMs`

### `restore`

Rolls the project back. Returns a preflight report and changes nothing unless
`confirm: true`.

Required:

- `projectId`
- `snapshotId`

Optional:

- `confirm` — must be `true` to actually restore. Set it only after the user has seen
  the preflight and agreed.
- `waitForCompletion`
- `timeoutMs`

### `delete`

Deletes a snapshot. Accepts **only** snapshots this plugin created.

Required:

- `projectId`
- `snapshotId`

Optional:

- `waitForCompletion`
- `timeoutMs`

### `decline`

Records that the user was asked for a backup and said no, which releases the backup
gate **for the given project**. Another project touched later in the same session is
still held once, and needs its own answer. Touches no API and creates nothing. Only
call this after actually asking the user, and pass the `projectId` the held call
targeted.

Required:

- `projectId`

### `read_task`

Reads a create/restore/delete task and returns normalized status and progress.

Required:

- `projectId`
- `taskId`

## Notes

- Create, restore and delete are asynchronous platform tasks. By default this tool
  waits for completion; a long-running task can outlive the wait, in which case the
  response carries `pending: true` and a `taskId` to poll with `read_task`.
- Snapshot names must be unique within a project. The generated name includes a
  second-resolution timestamp, so back-to-back backups never collide.
- Identification: a plugin backup is named `[AI Backup] …` **and** carries a marker in
  its description. Both must match, which is why a human snapshot that happens to
  share the name prefix is still protected from deletion here.
- Downloading, packaging and uploading snapshots are deliberately not supported. Use
  the Cognigy UI for those.
- A snapshot deployed to an Endpoint cannot be deleted at all — the platform refuses.