Skip to content

5.4.1 fixes from the 5.4.0 retest - #144

Merged
iLevyTate merged 9 commits into
mainfrom
fix/541-retest-bugs
Nov 18, 2026
Merged

iLevyTate merged 9 commits into
mainfrom
fix/541-retest-bugs

Conversation

@iLevyTate

@iLevyTate iLevyTate commented Nov 17, 2026 •

Copy link
Copy Markdown
Owner

Stacked on #143 (base feat/retest-quick-wins); retarget to main once #143 merges.

Fixes

  • Spreadsheets with prefixed XML are read. Visio and .NET (Open XML SDK) exports write <x:row>, <x:c>; both XLSX readers matched unprefixed tags only, so these workbooks read as empty. Element prefixes are removed before the shared-strings and worksheet XML is parsed (inline and worker paths). New test builds a prefixed workbook with adm-zip and fails without the fix.

  • Cloud analysis is no longer held to one file at a time. A batch took the concurrency the local hardware recommends (1 on a busy card) even when the document model ran on a cloud provider. When the router sends document analysis to a cloud provider, the documents section skips the hardware cap; queue pressure on local models still applies, and images and media keep the hardware cap.

  • Search and Chat name files as they are on disk. Analysis stores the suggested name in organization.newName before any rename; the keyword index and two index-rebuild paths used it as the display name, so chat cited unorganized files by names that did not exist. They now use the basename of the current path (a move already sets newName to the real name).

  • The index stats Refresh button stays still during analysis. Batch progress refreshed stats every 1.5 s and each refresh set the loading state. Background refreshes (batch progress, file moves) are now quiet; opening the panel, the button and rebuilds still show loading.

  • The embedding degradation warning names its reasons. A steady 25% embedding error rate was logged with no cause (reasons were debug-only). The warning now lists the five most frequent reasons. A local run did not reproduce the errors (0 of 181), so the cause will come from the next large run.

  • Cloud analysis no longer reloads the local text model after images. After each vision batch, _preloadTextAfterVisionRelease loaded the local text model even when documents go to a cloud provider; during a Claude reanalysis a 7B model loaded every ~5 minutes, evicting the embedding model, which reloaded for the next file. The preload is skipped when document analysis is remote (shared check in providerStamp.isDocumentAnalysisRemote, also used by the batch concurrency fix).

  • Folder provider notice is accurate. "Smart folder AI override ... Using the local model" was wrong: the router falls through to the default provider. The notice now names the provider that runs and the reason, throttled to once a minute per folder.

  • Manual e2e: reanalyze-embedding-errors.spec.js (local, cloud-key and prepared-profile modes) records the embedding pipeline's outcomes. Local, Claude, and a copy of a 1,000-file library all recorded 0 embedding errors; a real Claude reanalysis of 853 files after these fixes recorded 0 too.

Not changed

  • Vision model answering in prose: one occurrence (one 15 s timeout) in a 1,000+ file reanalysis; left as is.

Testing

  • npm run ci equivalent run locally on Windows: all 45 steps passed (format, lint, both typechecks, audits, compliance, coverage with thresholds, every verify step, knip, build).
  • Each fix has a test that fails without it.

Visio and .NET (Open XML SDK) exports bind the spreadsheet namespace to a
prefix (<x:row>, <x:c>), and the XLSX readers matched unprefixed tags
only, so these workbooks read as empty. Element prefixes are removed
before the shared-strings and worksheet XML is parsed, on both the
inline and the worker path.
The warning reported a steady 25% embedding error rate with no cause;
the reasons were kept in the recent window but logged only at debug
level. It now lists the five most frequent non-primary reasons. The
manual local routing review also records the embedding pipeline stats
(a local run had 0 errors in 181 embeddings, so the rate seen in a large
cloud reanalysis needs the reason from a real run).
A reanalysis through a cloud model ran one file at a time: the batch
took the concurrency the local hardware recommends (1 on a busy card),
though the document model ran remotely. When the router sends document
analysis to a cloud provider, the documents section skips the hardware
cap; queue pressure on the local models still applies, and images and
media keep the hardware cap.
Batch progress refreshed the index stats every 1.5 s, and each refresh
set the loading state: the Refresh button spun and disabled and the
counts faded for the whole run. Background refreshes (batch progress and
file moves) now update the counts quietly; opening the panel, the
Refresh button and rebuilds still show loading.
Analysis records the suggested name in organization.newName before any
rename, and the keyword index and two index-rebuild paths used it as the
display name, so search and chat cited unorganized files by names that
did not exist. They now take the basename of the file's current path; a
move already sets newName to the real name, so organized files read the
same as before.
…sons

Runs Reanalyze all on a Smart Folder of 3D models and notes with the
local models and records the embedding pipeline's outcomes. A local run
recorded 0 errors in 27 embeddings, so the 25% seen in a cloud-model
reanalysis is specific to that setup.
STRATOSORT_E2E_CLOUD_PROFILE copies a profile's AI settings and
encrypted keys into a fresh test profile so analysis runs on its cloud
provider while embeddings stay local; STRATOSORT_E2E_USE_PROFILE runs on
a prepared profile as-is. The watcher is stopped before the app closes,
since Reanalyze all keeps working in the background. Neither a cloud run
(27 embeddings) nor a run on a copy of a 1,000-file library (70) recorded
an embedding error; the test app runs CPU-only.
During a reanalysis through Claude, every image batch ended by loading
the local 7B text model, which nothing used; on a busy card it evicted
the embedding model, which then reloaded for the next file (the log
showed this every ~5 minutes). The preload after a vision batch is now
skipped when document analysis routes to a cloud provider; a local call
that needs the text model still loads it on demand. The remote check
moves to providerStamp and is shared with BatchAnalysisService.

The notice for a Smart Folder whose provider is not ready said the local
model would be used; the router falls through to the default provider.
It now names the provider that runs and the reason, once a minute per
folder rather than per file.
Base automatically changed from feat/retest-quick-wins to main November 18, 2026 05:02
@iLevyTate iLevyTate closed this Nov 18, 2026
@iLevyTate iLevyTate reopened this Nov 18, 2026
@iLevyTate
iLevyTate merged commit 78cba1e into main Nov 18, 2026
8 checks passed
@iLevyTate
iLevyTate deleted the fix/541-retest-bugs branch November 18, 2026 12:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant