Providers
Back to spooling.aiKiro
How Spool ingests Kiro (AWS) sessions.
Kiro is AWS's AI IDE. It's a VS Code fork, but unlike Cursor and Windsurf, it does not store chat data in state.vscdb: that file's chat keys are stubs ({"version":1,"entries":{}}).
The real session data lives as standalone JSON files.
Where sessions live
~/Library/Application Support/Kiro/User/globalStorage/kiro.kiroagent/
workspace-sessions/<base64-workspace-path>/<uuid>.json
workspace-sessions/<base64-workspace-path>/sessions.json # indexLinux: ~/.config/Kiro/User/....
The base64-encoded workspace dirname uses urlsafe encoding with __ for == padding. Spool decodes it back to the original filesystem path so cwd reads correctly.
What we parse
- Each
<uuid>.jsonis one session, top-level keys includehistory,sessionId,title,workspaceDirectory,selectedModel. - Per-session
dateCreatedfrom the siblingsessions.jsonindex becomesstarted_at. - User messages are typed content parts; assistant messages are bare strings. Both shapes are normalized.
Gotchas
- No per-message timestamps. Kiro doesn't emit them, only the session-level
dateCreated. We use that for the first message's timestamp and leave subsequent messages timestamp-less. Falls back to file mtime ifsessions.jsonis missing. selectedModelshape varies. Older sessions have it as a string ("agent"), newer ones as an object withtitle/model/contextLength. Parser handles both.- No tool-call structure. Kiro stores
executionIdfor assistant turns that ran tools, but the actual tool-call inputs and outputs aren't in the session JSON. We capture the message body and skip the trace tree.