You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Even after fixing the header fallback (#1), the Inspect View UI may still hide files due to deduplication by task_id values parsed from filenames. The log_file_info() function populates EvalLogInfo.task and EvalLogInfo.task_id from the filename, and the client-side logs IndexedDB store uses these garbled values for grouping/deduplication — separate from the correct values in log_previews.
Context
This was observed during the investigation for #1. After renaming eval files to g2a_simlex999_{model}.eval, 13 of 17 files appeared (up from 1), but 4 remained hidden because:
voyage_3_large → task_id="3"
openai_3_small → task_id="3"
openai_3_large → task_id="3"
All three collide on task_id="3", so the UI deduplicates and shows only one.
Root Cause
The logs store in IndexedDB is populated from the /api/log-files response, which uses log_file_info() filename parsing. The log_previews store is populated from actual header reads. The UI uses the logs store for list rendering and applies deduplication/filtering on task_id — which contains the garbled filename-parsed values, not the correct header values.
Issue #1 fixes the log_file_info() return values by falling back to header reads. Once #1 is implemented, log_file_info() will return correct task/task_id from the header, which means the logs store will also have correct values. This issue may be fully resolved by #1.
However, if the client-side code separately caches or derives task_id from the filename (outside of what log_file_info returns), this issue would persist.
Summary
Even after fixing the header fallback (#1), the Inspect View UI may still hide files due to deduplication by
task_idvalues parsed from filenames. Thelog_file_info()function populatesEvalLogInfo.taskandEvalLogInfo.task_idfrom the filename, and the client-sidelogsIndexedDB store uses these garbled values for grouping/deduplication — separate from the correct values inlog_previews.Context
This was observed during the investigation for #1. After renaming eval files to
g2a_simlex999_{model}.eval, 13 of 17 files appeared (up from 1), but 4 remained hidden because:voyage_3_large→task_id="3"openai_3_small→task_id="3"openai_3_large→task_id="3"All three collide on
task_id="3", so the UI deduplicates and shows only one.Root Cause
The
logsstore in IndexedDB is populated from the/api/log-filesresponse, which useslog_file_info()filename parsing. Thelog_previewsstore is populated from actual header reads. The UI uses thelogsstore for list rendering and applies deduplication/filtering ontask_id— which contains the garbled filename-parsed values, not the correct header values.Relationship to #1
Issue #1 fixes the
log_file_info()return values by falling back to header reads. Once #1 is implemented,log_file_info()will return correcttask/task_idfrom the header, which means thelogsstore will also have correct values. This issue may be fully resolved by #1.However, if the client-side code separately caches or derives task_id from the filename (outside of what
log_file_inforeturns), this issue would persist.Verification Plan
After #1 is merged:
/api/log-filesreturns correcttask/task_idfor all fileslogsIndexedDB store has correct valuesAcceptance Criteria
task_idcollisions from filename parsing