Datasets, Rebuilt on S3
Dataset content now stores in S3 as chunked JSONL, so one dataset holds thousands of rows without bloating the database. Files upload straight from your browser.
What You Can Do Now
Bigger datasets, without the database bloat
-
Upload CSVs through a redesigned drawer and dropzone.
-
Bulk upload multiple datasets at once.
-
Confirm column names and types on upload, and edit column types on existing S3 datasets.
-
Files upload directly from your browser to S3, with visible progress and clear errors when an upload fails.
-
The editor surfaces read truncation, so you always know when you are looking at a partial view of a very large dataset.
Explore traces
-
Round 5 of trace explorer improvements.
-
The trace ID copy toast is now dismissible.
Faster, and safer edits
-
The app loads faster: the drawer registry is now lazy-loaded, cutting that bundle from 4.3 MB to 232 KB.
-
Scenario stopped silently losing edits in the editor, with criteria saved on blur and a save-and-run flow.
Langy
- Langy loads its skills as native opencode skills straight from the canonical source.
Screenshots
Datasets now scale to thousands of rows, backed by S3.
Upload CSVs through a redesigned drawer and dropzone.
Bulk upload several datasets at once.
Notes
-
Security: closed a CRITICAL @vitest/browser alert and bumped starlette and the vite floor across workspaces.
-
Performance: a ClickHouse pass with index gaps closed, opt-in quantileTDigest, a conversation-id skip-index for thread-to-traces lookups, and pruned poll-window subqueries.