Save a session.
Find your way back.

1. Check the tools are available

In your connected agent, ask: “Use AI Babysitter’s discover tool to find the session upload, status, search, and download capabilities.” Check the returned schemas before executing. If the tools are missing or access is denied, stop and contact your preview operator; reconnecting or buying a plan is not a substitute for preview access.

The blocks below are argument objects for the MCP execute tool, not terminal commands. Ask your agent to run them through AI Babysitter. Don’t paste keys or passwords into the transcript.

2. Save a synthetic transcript

Choose a fresh native_session_id for each new sample. This example uploads one generic JSONL message and marks the transcript complete. Keep the returned session_id; replace ses_FROM_UPLOAD in later steps with that value.

{
  "capability": "session-transcript-append",
  "arguments": {
    "harness": "generic",
    "native_session_id": "my-vault-sample-001",
    "byte_offset": 0,
    "content": "{\"type\":\"message\",\"role\":\"user\",\"content\":\"We chose a simpler signup page for the foxglove demo.\"}\n",
    "final": true
  }
}

For real uploads, confirm your export format first. Send the original UTF-8 text unchanged, not an agent summary, and split chunks at newline boundaries. For multiple chunks, use the server’s returned bytes_received as the next byte offset and set final: true only on the last chunk. Offsets are UTF-8 bytes, not character counts. Check status before retrying an uncertain upload; identical chunks can be retried, conflicting offsets are rejected. Keep chunks under 2 MiB.

3. Wait for the searchable view

{
  "capability": "session-transcript-status",
  "arguments": {
    "session_id": "ses_FROM_UPLOAD"
  }
}

Confirm the accepted byte count and finalization, then inspect derivation status. Search and readable exports need the derived view: wait while it is pending or running, and continue when it is ready. If it reports failed or skipped, do not assume the transcript was empty; save the status result for your preview operator.

4. Find the useful moment

{
  "capability": "session-search",
  "arguments": {
    "query": "foxglove",
    "session_id": "ses_FROM_UPLOAD",
    "limit": 10
  }
}

Check the result matches the sample sentence. To search your other saved sessions, omit the session filter. You can also ask your agent to discover session-list to browse session metadata.

5. Retrieve and check the content

{
  "capability": "session-download",
  "arguments": {
    "session_id": "ses_FROM_UPLOAD",
    "format": "raw",
    "offset": 0,
    "max_bytes": 262144
  }
}

Check the returned content against the sample. For larger transcripts, follow returned pagination using next_offset until the content is complete; one response may not include the full transcript. “Raw” means the stored transcript after secret-pattern redaction, not a guaranteed byte-identical backup.

{
  "capability": "session-download",
  "arguments": {
    "session_id": "ses_FROM_UPLOAD",
    "format": "slim",
    "max_bytes": 262144
  }
}

Use slim for derived Markdown, facts for structured metadata, or events for normalized events. Inspect truncation and pagination fields before treating any export as complete; discover the current schema for format-specific options.

6. Continue deliberately

Review the downloaded context, open the correct project, and give the relevant excerpt to your agent: “Here is the earlier decision. Check the current code before proposing the next change.” Retrieve on another machine if needed, but bring your repository, dependencies, and credentials separately.

Optional session notes can record decisions and next steps; discover session-annotate for its schema. Checkpoint/resume is a separate progress-snapshot workflow, not native restoration of an uploaded transcript.