Trajectory storage¶

The trajectory processing model is StorylineDocument. ATIF, ACTF, OpenAI Messages, Storyline JSON, and AgenticMD are supported encodings; Codex and Claude Code session JSONL are decode-only inputs. Their timestamps continue to be parsed and preserved by the codecs.

StorylineLanceStore stores documents directly as runs.lance, steps.lance, and tool_calls.lance, with a content layer for large fields. CURRENT pins a published generation and exact Lance versions. Reconstructing a document uses that pinned generation and validates referenced content.

Import supports create, append, and replacement. Local whole-Dataset replacement stages output before swapping it into place; object-store whole-Dataset replacement has different failure semantics, documented in the import guide. Store-level publication uses writer leases and generation checks. Maintenance and deletion must preserve the same publication and content-integrity guarantees.

Compact JSONL is a separate record storage path preserving original JSON bytes. It does not infer Storyline Runs, Steps, or ToolCalls.

See Storyline Lance, Snapshot, formats, and import/export.