meinKontextDocs
Navigation & search

Plugins

Conversations and sessions

Import transcripts, continue across clients and share through buckets.

Keep an agent conversation together even when the work continues in another client. A conversation has a title and context links. Each original client run is a session. Its entries preserve the original messages, tool events and branches. Summaries, handoffs, outcomes and insights are separate derived records with cited sources. Reported speakers do not prove who imported the content.

Import and continue

Open Conversations, create a conversation and select its export file. Choose Codex JSONL, Claude Code JSONL or Sessions JSON v1, then inspect the preview before importing. Files are limited to 4 MB; choose a stable source ID if the export has none. Unknown event types remain as raw line content and generate a warning. Invalid JSON, mixed session identities and changed duplicate IDs refuse the import. Raw line contents are preserved; original newline separators and blank lines are not represented by entries. Attach the original file separately when byte-for-byte file preservation is required (up to 525 KB per attachment).

Imports save in batches. If interrupted, choose the same file and source ID again: already saved entries are retained and identical entries are not duplicated. A conflicting entry is never overwritten. Do not change the source ID to hide a conflict. The normal host secret gate applies to transcripts and attachments.

Select a session and start a new session in the intended client to record an explicit continuation. This links the runs; it does not start another application or execute an agent. Save a handoff against the selected source session. Original text is untrusted data, including instructions that appear in imported messages.

MCP workflow

  1. sessions_query discovers conversations, sessions, entries and derivations. Filter by conversation, session, client, external source ID (externalId), creation-time bounds or full text. Follow offset/limit; an empty result is not evidence of an import failure.
  2. sessions_command accepts conversation.create/edit/link/unlink, session.start/finish/attach, entries.append and derivation.create. Supply current expectedRevisions when changing existing records and a durable idempotencyKey. Retry uncertain writes with exactly the same key and arguments.
  3. sessions_import_preview parses a selected export without saving. Follow nextOffset; append every returned batch to the resolved session.
  4. sessions_read returns a record and its revision. sessions_source requires that session revision and returns at most 50 entries with IDs and revisions. On STALE_SOURCE, read again and recompute the derivation. A derived text is a client's assertion, not a platform-verified fact or an automatic publication.
  5. sessions_export exports a conversation's currently authorized records and binary attachments. Public v1 record contracts support explicit integrations with other plugins. Context links grant no access to their targets.

Native session interchange is JSON with schemaVersion: 1, client, externalId and entries. Each entry has externalId, role, text and optional nullable occurredAt, parentExternalId, toolCallId, raw. The generic conversation archive is a different format, retaining record identity and binary files.

Buckets and lifecycle

Share a conversation through an existing bucket. Members can read its current and future sessions, entries and derivations; sibling conversations remain private. Use plugin_link_bucket with container_id, bucket_id and the reader role from plugin_access; role_id: null removes the link. The host implements this for any declared plugin container. Bucket membership never grants custody, onward sharing or execution rights. Removing a member revokes source and derived access.

Archive a conversation to hide it and its descendants. Use the archive filter to find and restore it. Export before archiving. Archival does not erase stored history; permanent account erasure follows the platform's existing deletion path. No automatic transcript collection, retention job, external sending, memory promotion or Wiki writing runs in the background.

In the browser, “Download session as JSON v1” downloads the selected session in the importable format (up to 4 MB). The browser also supports title/description edits, context links, all four derivation types and immutable original-file attachments up to 500 KB.

Exact sources, search and deletion

Use kind: "entry" to search original text, client for sessions, and participant (an exact reported label) or contextObjectId for conversations. after and before refer to recording time. derivationKind filters summaries, handoffs, outcomes or insights. Search results open the exact record. Participant labels are editable metadata; they never authenticate anyone or replace the host's recordedBy actor. The existing Tasks work-summary contract links projects without copying them.

Select original entries to cite precise passages, or select a session to cite its whole current revision. sessions_read accepts sourceId with a derivation id to open the cited revision; offset pages its entries. The UI marks changed sources and opens the earlier cited content. Later entries do not become part of an earlier citation. Current permissions apply to historical sources too. Individual attached files can be downloaded from the session.

Repeated imports verify the entire saved prefix before extending it. Conflicting or reordered originals are rejected. Resume an interrupted import using the same source identity and unchanged saved entries. Secret rejection is distinct from a source conflict: remove credentials from the selected export copy before retrying. Nontext blocks and multiple tool calls remain in raw, with explicit preservation warnings; importing never executes or sends their contents.

conversation.archive and conversation.restore accept id and require expectedRevisions[id]. Restore before reading or exporting an archived conversation's children. Permanent deletion requires an archived conversation and its human custodian using the Sessions browser UI. MCP deletion is refused. It removes all its sessions, entries, derivations and version history atomically, collecting only binary bytes no other content references. Siblings and buckets remain. Managed agents cannot delete. Deletion cannot be undone. After an uncertain lifecycle operation reread current state. The ordinary account erasure path remains available. Original capture commands retain their receipt-based exact retry protocol.

Browser workflow

The library separates conversations, originals, results, sessions and the archive. Use Originals to search source text, and Filters to narrow the library. Within a conversation, choose a session on the left and read its transcript on the right. Results contains cited summaries, handoffs and insights; Context links existing records from other plugins. Import, editing, continuation and sharing each open a focused dialog. Session details contains attachments and the session export. More actions contains conversation export and archival. Permanent deletion becomes available there only after archival and requires an explicit confirmation.