What shipped, and when

Guided folder setup, and searches that stop when you do

• Added

  First-run guidance for choosing which folders to index. Pointing the index at Documents and Downloads rather than the whole home directory is the single biggest factor in how long the first pass takes, and the app now says so instead of leaving you to work it out.

• Fixed

  A search that had already been superseded by newer keystrokes kept running to completion. Queries you have moved on from are now abandoned, which cuts the work done while you are still typing.

Homebrew install, and a release pipeline that fails loudly

• Added

  Installation via Homebrew: brew tap dhito-app/tap, then brew install --cask dhito. The cask pins the SHA-256 of the notarised DMG, so the checksum is verified before anything is mounted.

• Changed

  Release metadata now states the Apple Silicon requirement explicitly, so Homebrew stops with a clear message on unsupported hardware rather than installing something that cannot run.

• Fixed

  The release build now fails if anything unexpected has been bundled into the app, or if the configured version has drifted from the tag. Both had previously been possible to ship without noticing.

Trial email prompt, and a privacy policy that matches the app

• Added

  Free-trial users are asked for an email address from day 10. It is a prompt, not a wall — the trial keeps working whether or not you answer it.

• Fixed

  The privacy policy described a blocking email requirement the app does not have. Corrected to match the behaviour that actually ships, along with the support contact address.

Large text libraries, and the update flow

• Fixed

  Indexing a folder holding very large numbers of plain-text files could stall. Text extraction is now batched so the pass completes on libraries in the tens of thousands of files.

• Fixed

  Several version-upgrade paths (0.4.15 through 0.4.18) that could leave the app restarting into the wrong state, or reporting a stale subscription status after returning from the billing portal.

Memory released during indexing

• Fixed

  Model memory was not being freed between indexing batches, so a long first pass grew its footprint until the machine started swapping. Buffers are now released as each batch completes.

Model loading on a clean install

• Fixed

  On some machines the bundled inference libraries could not locate their own resources after installation, leaving indexing unable to start. The bundle layout was corrected.

Long videos, large files, and chat across every supported type

• Fixed

  Chatting with a long video or a very large document could hang. Transcript and document context are now chunked before they reach the local model.

• Fixed

  Indexing long video files could stall part-way through and leave the file permanently unsearchable.

• Fixed

  Files the indexer could not read were retried on every pass instead of being recorded as skipped, so each rescan repeated work that had already failed.

• Fixed

  A code-signing fault in the build pipeline that produced installs macOS reported as damaged (0.4.9 through 0.4.11).

Vision model resilience and faster first captions

• Fixed

  The vision model could fail to initialise in half precision on some GPUs. It now falls back rather than leaving images unindexed.

• Changed

  Detailed captioning and object detection are primed on the GPU at startup, so the first images to be indexed no longer pay the warm-up cost.

Diagnostics you can send, and an update that relaunches

• Added

  Optional diagnostic log submission, so a reproducible failure can be reported without reconstructing it by hand. It is opt-in and carries no file contents.

• Fixed

  Auto-update could install a new version but fail to relaunch into it.

• Fixed

  A local LLM chat that hung with no error, and a keychain timeout that could block licence checks on a slow first unlock.