vitals

Performance improvements

• The log viewer serves log files as-is rather than parsing and re-serializing them, making opening a log effectively instant (previously several seconds for logs tens of MBs in size).
• vitals_view() now reads only the leading bytes of each log file when

generating the homepage; listing a directory of large logs is roughly 250x

faster.

• Eval log files are dramatically smaller (roughly 4x for multi-turn, tool-heavy evals). Repeated content in a sample's events and base64 images in its messages are now de-duplicated into the sample's attachments pool, mirroring Python Inspect's behavior.

Bug fixes and minor improvements

• $log() now records its fallback temporary directory on the task and reports the path it wrote to when no VITALS_LOG_DIR is set, so the automatic $view() succeeds.
• Task$new() now assigns a valid task name when given an inline (unnamed) dataset, so $log() succeeds.
• vitals_bundle() now tolerates logs whose results contain no scores.
• vitals_view() will now show correct metadata for each log. Previously, the log viewer could display one log's metadata (task name, model, score) in place of another's, both in the log listing and when clicking into a log (#208).
• detect_pattern() now supports case_sensitive = TRUE on R 4.5 and later.
• The log viewer will now appropriately display tool calls called in parallel.