TranscriptImport

The one reader of the Claude Code transcript format: it turns a <session-id>.jsonl file into the part's events, once, on adopt (trove-parts §TPN-7, §TPN-8).


Group	libllmsession

In the namespace octomy::llmsession. Declared in src/libs/libllmsession/llmsession/TranscriptImport.hpp.

Detailed description

Claude Code writes one JSON object per line into ~/.claude/projects/<project>/<session-id>.jsonl. The official documentation calls the format internal and says that it changes between versions (§TPR-7). So the reader is lenient, and nothing else in the tree parses the format. Its rules:

An entry is one JSON object per line. An unknown type, an unknown content block and an unknown field are skipped, never an error. A line that is not JSON is skipped. When it is the last line, Claude Code was still writing it, and the result says truncated.
The transcript is a tree: each entry has uuid and parentUuid. The reader follows the main chain back from the last entry, through logicalParentUuid where a compaction cut the chain. It skips isSidechain entries and the entries of an abandoned branch. A tool result whose call is on the main chain stays, because parallel tool calls hang their results beside the chain.
One assistant reply spans several lines with one message.id. The reader joins them into one event per content block: a block that comes again, with the same apiBlockIndex or the same content, replaces the first one.
A user entry's content is a string or an array. Both are read. An image block becomes a placeholder block that names its media type.
isMeta entries, compact summaries, interrupt notes, local commands and system entries become system events.
message.usage becomes one usage event per reply, with the tokens in the form that the stream transport gives.

Each prompt starts a turn, and each turn ends with a turn_end of reason imported, so that a resume after the import counts its turns right. Every block carries "imported": true.