Jot 0.3.3

Jot 0.3.3

Models and updates now uses responsive model cards with clearer roles, revision details, and update states. App updates have their own card, with layouts that adapt to narrow windows. Includes the Activity dashboard and quieter idle behavior from 0.3.2.

Jot 0.3.2

Jot 0.3.2

Activity now features live Mac-impact statistics, CPU and memory line charts, and dictation and ambient-use insights in a responsive layout. Removed the capture-event list. Paused housekeeping and diagnostic logging run about once a minute; Activity charts refresh once a minute and usage totals every five minutes while the tab is active. Fn dictation remains immediate. Agents can read aggregate usage through the activity_report MCP tool and jot activity CLI. This development-signed prerelease updates through Jot on Macs that trust the existing development certificate.

Jot 0.3.1

Jot 0.3.1

Protect JotLab private scratch cleanup on hangup and broken pipe, and refuse output paths that resolve inside Git repositories.

Jot 0.3.0

Jot 0.3.0

Accuracy groundwork.

• Words no longer drop at recognition chunk seams when one decode heard them and the next missed them (#216).
• jot lab runs a recording through the real pipeline once per settings variant and compares word error rate and speaker word accuracy against captions; jot settings import applies the chosen variant (#217, #231).
• keepTuningAudio (off by default) keeps each session's audio as a 16-bit WAV for tuning, deleted after 30 days or oldest-first past 10 GB, and with its session (#230).
• The longest recognition chunk can be set up to 15 s; the default stays 3 s (#227).
• The recognition engine is its own JotEngine module with package tests (#229).
• Settings tests no longer leave files in ~/Library/Preferences (#225).

Jot 0.2.13

Jot 0.2.13

Move large Accessibility field reads, insertion comparisons and Unicode event preparation off the main actor. Preserve target, focus and cancellation checks before delivery, and stop ambiguous writes instead of inserting twice. Build the app and core in Swift 6 with complete strict concurrency. Separate evaluation-only suggestion behavior from production and organize supporting types into named files. Production suggestions remain request-only; existing history, settings and socket compatibility are preserved.

Jot 0.2.12

Jot 0.2.12

Optional window images can now ground explicitly requested suggestions when Accessibility misses visible content. Enable Include window image in Suggestions settings to use it; it stays off by default. Screen Recording access is requested only when enabling the option. Images are captured once per request, kept in memory, and processed by the on-device model. No screenshot history is saved.

The image spans the actual field's horizontal column, from the window top through the field, clipped to the matched window. It never widens or shifts horizontally, and horizontal pixel rounding stays inside that column. This is a geometric boundary, not a guarantee that every application's layout separates conversations by columns.

Ambiguous window matches are rejected. Dismissal cancels capture, stale requests cannot supply newer requests, and late native values are discarded. Unsupported or rejected images fall back to text when existing selected sources or a rewrite seed ground the request. Source attribution and receipts reflect whether an image was actually used.

Development prerelease, signed with Monroe's Apple Development identity; not notarized. Requires a Mac that trusts this development certificate. Basic Jot support starts at macOS 14; Apple Intelligence text enhancements require compatible hardware and macOS 26 or later. Image input requires macOS 27 and a model supporting images; text behavior remains available without image support.

Validation: 374 Swift tests, 78 Python tests, 27 structural fixture scenarios, signed Debug/Release builds and isolated recovery checks passed. The actual real-model audio recovery path passed with generated speech. Three native synthetic-window captures excluded the neighboring sidebar and supplied the visible meeting time correctly to the on-device model. Capture plus the bounded question measured 1.38 seconds first and about 0.47 seconds warm. These are probe timings, not the full production suggestion coordinator.

Test when you sit down: enable the optional image setting and follow the Screen Recording permission flow; double-Fn a suggestion and accept with Tab in native and Electron fields; hold/release Fn for ordinary dictation; switch windows/fields during a request; try multiple displays. Compare suggestion usefulness, complete request latency and battery impact. Permission enable/revoke, physical shortcuts, full coordinator behavior, multi-display behavior and real-world energy use remain unverified.

Jot 0.2.11

Jot 0.2.11

Dictation now shows the field outline as soon as its target is found, and the processing wash loops until insertion finishes. Very short filler-only holds and ambient backchannels such as “Yeah” and “Mm-hmm” are filtered more carefully. Empty recognition results avoid unnecessary store work, and speaker processing runs at utility priority.

Setup, model-download, permission and recovery guidance remain reachable in narrow windows, including the compact icon rail. Speech and suggestion modules now receive their dependencies explicitly, and one-shot model completion handling is shared.

Adds an optional Claude Code jot-transcripts plugin for reading local Jot sessions. The plugin is installed separately; explicitly requested transcript excerpts can be sent to Claude's model. Optional window-image suggestions remain pending in PRs #166 and #185 and are not included.

Development prerelease, signed with Monroe's Apple Development identity; not notarized. Requires a Mac that trusts this development certificate. Basic app support starts at macOS 14; Apple Intelligence enhancements require compatible hardware and macOS 26 or later. Capture and stored history remain local; models download separately.

Test when you sit down: hold and release Fn in a native and an Electron text field and watch the outline and processing wash; try a short filler-only hold, double-Fn suggestions and Tab acceptance. Check setup and recovery controls in a narrow window. If using the new plugin, install it separately and ask Claude to list recent Jot sessions.

Validation: 348 Swift tests, 78 Python tests, 27 structural suggestion-fixture scenarios, signed Debug build and isolated recovery-controller checks passed on current main. Generated speech also passed the real-model audio recovery path. Release packaging runs the Swift tests again and verifies the exact signed Release product. Native window checks also passed for setup, permission and recovery guidance at 719 points wide and 520 × 420. Physical Fn and cross-app insertion remain end-to-end user checks.

Jot 0.2.10

Jot 0.2.10

Jot stays responsive while loading large histories and saving speech. Store queries, Live-feed grouping, recognition saves, cleanup, recovery and speaker edits now run off the main thread in submission order. Changing the selected session while a read is pending keeps the latest selection; delayed updates use the current saved text without duplicating rows.

The dictation shortcut runs on its own event-tap thread. Target lookup is asynchronous with short timeouts, and a late lookup is rejected after the app or field changes. A failed session deletion preserves speech still being recognized; a successful deletion prevents late results from recreating the session. Updates wait for pending storage and speaker work.

The underlying modules now take their sibling dependencies directly, with separate files for store data types, shortcut handling and screen-context reading.

Development prerelease, signed with Monroe's Apple Development identity; not notarized. Requires a Mac that trusts this development certificate. Basic app support starts at macOS 14; Apple Intelligence enhancements require compatible hardware and macOS 26 or later. Transcripts and audio stay on the Mac. Models download separately.

Validation: 346 Swift tests, 48 Python tests, signed Debug/Release builds and recovery-controller checks passed. Generated speech exercised the real recognition and cleanup models. A 7,000-row history load measured a 1.35 ms longest main-thread gap; a speaker pass and Regroup over 3,000 rows dropped no audio in the synthetic listening check.

Test when you sit down: hold and release Fn in native and Electron fields, try a rapid tap, double-Fn suggestion and Tab acceptance, and switch fields during target lookup. Switch Sessions and search while speech is saving; rename, regroup and export a saved session. Physical Fn and cross-app insertion remain user checks.

Jot 0.2.9

Jot 0.2.9

Single-word held dictation inserts without a trailing period and bypasses the prose model. This happens before explicit spoken symbols and personal vocabulary are applied, so a dictated "period" and preferred spellings keep their punctuation. Recognized source text and existing input-field text are preserved.

Apple Intelligence dictation cleanup now distinguishes phrase fragments from complete sentences: "the blue one" stays open, while "I want the blue one" gets a period. The single-word fix also works with cleanup disabled or unavailable.

Also includes the changes already on main since 0.2.8: a first-install setup page that resumes progress, automatic recovery from silent microphones, updates that wait for active listening to finish, and preserving the current tab when changing microphones.

Development prerelease, signed with Monroe's Apple Development identity; not notarized. Requires a Mac that trusts this development certificate. Basic app support starts at macOS 14; Apple Intelligence enhancements require compatible hardware and macOS 26 or later. Transcripts and audio stay on the Mac. Models download separately.

Validation: 364 Swift tests, 48 Python tests, signed app builds, recovery-controller checks, and the actual on-device model distinguishing fragments from short complete sentences. Synthetic audio exercises real recognition and dictation delivery.

Test when you sit down: hold Fn to insert one word in the middle of an existing sentence; try "the blue one" and "Go now" with dictation cleanup on; say "purple period" to request an explicit dot. Check that surrounding text stays intact. Physical Fn and cross-app Accessibility behavior remain user checks.

Jot 0.2.8

Jot 0.2.8

Separate suggestions from saved dictation. Double-Fn only requests suggestions; when suggestions are off, it never inserts recent room speech. Review saved dictation explicitly to preview and copy an undelivered hold, including while paused. The old Recovery window setting and automatic fallback are removed.

Apple Intelligence remains optional. Suggestions explain unavailable states, cleanup preserves recognized text when unavailable, and previously enabled unavailable features can be switched off.

Development prerelease, signed with the local Apple Development identity; not notarized. Requires a Mac that trusts this development certificate. Basic app support starts at macOS 14; Apple Intelligence enhancements require compatible hardware and macOS 26 or later. Transcripts and audio stay on the Mac. Models download separately.

Validation: 323 Swift tests, 48 Python tests, signed Release build, recovery-controller checks with real-model synthetic audio, and all simulated model-unavailability states. Older-OS hardware and physical Fn/Tab acceptance remain user checks.