Skip to content

CHANGELOG.md

Latest commit

 

History

History
465 lines (366 loc) · 19.9 KB

CHANGELOG.md

File metadata and controls

465 lines (366 loc) · 19.9 KB
Raw

Changelog

All notable changes to Bokashi are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.13.0 - 2026-10-06

Japanese names are found without Apple Intelligence.

Changed

  • Japanese names can be found without Apple Intelligence. With the privmask CLI installed (brew install snaka/tap/privmask), Bokashi uses the Japanese name model that comes with it. The model is not bundled, so the app stays small; without the CLI nothing changes. The Homebrew cask now depends on the privmask formula, so brew install --cask snaka/tap/bokashi brings the model along. Apple Intelligence still adds to it when enabled. The debug overlay draws findings from the name model alone in purple.

Fixed

  • Fewer false positives from privmask 0.5.0: hostnames, job names and English product names are no longer masked as personal names, and later mentions of a name already found are masked too.

0.12.0 - 2026-09-28

Auto-masks can now be removed one at a time.

Added

  • Eraser tool. A new toolbar tool removes the annotation you click, including masks added by auto-mask, so a wrong detection no longer means undoing the whole batch. Masks and filled shapes are hit anywhere inside; outlines, lines and arrows only near their stroke, so something drawn inside an outline stays reachable. Where annotations overlap, the topmost one goes. Undo brings it back. The pointer turns into an outlined eraser while the tool is selected. (#45)

0.11.0 - 2026-09-28

Bokashi now finds Japanese personal names when it auto-masks a screenshot, and it can open images shared to it from any app's Share menu.

Added

  • Share menu integration. Bokashi appears in the macOS Share menu; an image shared from any app opens in the editor with the usual annotation and masking tools. Auto-mask follows the existing toggle, as capture and clipboard import do. (#42)
  • More kinds of sensitive text are detected: postal codes, Japanese My Number (verified by its check digit), and API keys, tokens and secrets — found either by a published prefix or by the name that introduces them, so AWS_SECRET_ACCESS_KEY=… masks the value and leaves the key name readable. (#43)

Changed

  • Text detection now uses privmask. Japanese personal names were effectively never found before, because Apple's NLTagger has no Japanese named-entity model. On Macs with Apple Intelligence, the on-device model is now asked for Japanese names only, and each candidate is checked against a family-name list before it is masked. The model is no longer asked for everything at once, which is where false positives such as port numbers read as addresses came from. Bare digit runs are only treated as phone numbers at Japanese lengths with a leading 0, so order numbers and timestamps stay readable. (#43)

Privacy

  • Images sent via the Share menu are written to Bokashi's App Group container until Bokashi next launches and imports them. This is the one exception to Bokashi never writing a screenshot to disk unless you ask. (#42)
  • All detection still runs on device.

Notes

  • macOS leaves third-party share extensions switched off. Enable Bokashi once under System Settings → General → Login Items & Extensions → Sharing. (#42)
  • Usernames / @handles and credit-card numbers are no longer detected automatically; the previous broad model prompt covered them loosely. They are tracked upstream in snaka/privmask#25 and #26. Mask them by hand or add them as custom terms meanwhile. (#43)
  • English-only text is no longer sent to the on-device model; English names are still found by NLTagger. (#43)

0.10.0 - 2026-09-12

Bokashi can now mask an image it did not capture itself. Anything on your clipboard — a screenshot from another tool, an image copied out of a browser or a chat app — opens in the editor with the same annotation and masking tools.

Added

  • New from Clipboard. A menubar item and a ⌃⌥⇧V hotkey open the editor on whatever image the clipboard holds. Reading the clipboard needs no Screen Recording permission, so this is the one path that works before you grant it. The image is decoded at its true pixel dimensions, so masks line up with the pixels you see. (#39)

Changed

  • Auto-mask now covers imported images too. The menubar toggle is relabelled from "Auto-mask sensitive info on capture" to "Auto-mask sensitive info", because it applies to anything that opens in the editor rather than only to fresh captures. Your existing setting carries over. (#39, #40)
  • Copying offers PNG alongside TIFF. Bokashi previously put TIFF only on the clipboard, which for a Retina screenshot is roughly 24 MB of uncompressed pixels that every app has to copy when you paste. PNG now comes first and TIFF is still offered, so anything that can only read TIFF is unaffected. Measured on a 640×400 image: 5,820 bytes of PNG against 1,027,430 of TIFF. (#40)

Notes

  • "New from Clipboard" needs the image itself on the clipboard. Copying a file in Finder puts a file reference there rather than image data, and Bokashi will report that the clipboard holds no image.

0.9.1 - 2026-09-01

A packaging fix. The app itself is unchanged from v0.9.0.

Fixed

  • Homebrew cask now states its macOS requirement correctly. The generated cask wrote depends_on macos: ">= :tahoe", a string Homebrew compares lexically instead of by version. It now uses the symbol form, depends_on macos: :tahoe. (#37)

0.9.0 - 2026-07-20

On-device sensitive-info detection moves to Apple's own AI. The optional Ollama vision-LLM detector is gone, replaced by Apple Intelligence for contextual text and a Vision-based face detector for avatars — both run entirely on your Mac with no setup, no network, and no model downloads.

Added

  • Apple Intelligence detection. Detect now routes OCR text through Apple's on-device Foundation Models to catch sensitive strings the pattern-based detectors miss: usernames and account handles, API keys and other secrets, credit card numbers, and names in context. The language model only decides which text is sensitive; the mask rectangles come from Vision's OCR bounding boxes, so coordinates stay exact. Toggle it in Settings → Detectors; a status line shows whether Apple Intelligence is ready, still downloading, needs enabling, or is unsupported on this Mac. (#35)
  • Face and avatar masking. A Vision face detector masks any detected face, including profile pictures — covering the case the old vision-LLM prompt handled without needing a language model. Runs on every supported Mac; toggle it in Settings → Detectors. (#35)

Changed

  • Requires macOS 26 (Tahoe). The deployment target moves from macOS 14 to 26 for the Foundation Models framework. The pattern-based detectors (email, phone, address, name, custom terms) continue to work on every supported machine; Apple Intelligence detection additionally needs Apple Silicon with Apple Intelligence enabled and degrades silently where it is unavailable. (#35)

Removed

  • Ollama vision-LLM detector. The optional detector added in v0.5.0, along with its endpoint/model settings and connection test, is removed in favor of the built-in Apple Intelligence detector. Its stored preferences are cleared automatically on first launch. (#35)

Privacy

  • All detection still runs on-device. With Ollama gone, the detection pipeline makes no network calls at all; no cloud-LLM integrations exist and none are planned.

0.8.1 - 2026-05-31

A small fix to the window picker.

Fixed

  • Window picker now targets the window you actually see. When windows overlap, hovering and clicking resolve to the front-most window under the cursor instead of one hidden behind it. Previously an occluded window could be selected and jump to the front, making the window you meant to capture hard to pick. SCShareableContent does not guarantee z-order, so candidates are now ordered front-to-back via the on-screen window list before hit-testing.

0.8.0 - 2026-05-30

A polish-and-papercut release. Capture-mode hotkeys realign to the macOS digit conventions, the menu bar finally shows what those combos are, the raw capture is on your clipboard the moment you press the shortcut, and Sparkle's automatic update check is now flippable from Settings.

Added

  • Clipboard-on-capture. The raw screenshot is copied to the clipboard the instant capture completes, before the editor opens. The existing Copy & Close flow still overwrites with the annotated image when you confirm, so "paste right away" and "edit, then paste" both work without extra clicks.
  • Window-capture hotkey. New default ⌃⌥⇧5 triggers the click-to-pick window flow, joining full-screen and region as a first-class shortcut. Configurable through UserDefaults / KeyboardShortcuts.
  • Hotkey hints in the menu bar. Each Capture item shows its bound combo on the right (e.g. ⌃⌥⇧3), so you no longer have to remember the mapping. Hints follow whatever combo is currently bound, ready for the future Settings rebinder.
  • "N items masked" toast. After the editor's Detect button finishes, a transient banner reports either the count of new masks or "No sensitive info detected". Auto-mask-on-capture stays silent on empty results so it doesn't toast every screenshot.
  • Automatic update-check toggle. New Settings → General tab with a Check for updates automatically switch bound to Sparkle's SPUUpdater.automaticallyChecksForUpdates. Default stays off so the first launch behavior is unchanged; flip it on and Sparkle starts polling the appcast in the background.

Changed

  • Capture hotkey digits aligned with macOS. Defaults are now ⌃⌥⇧3 (full screen), ⌃⌥⇧4 (region), ⌃⌥⇧5 (window) — mirroring the system's ⇧⌘3 / ⇧⌘4 / ⇧⌘5 mapping so the digit is recallable. Previous defaults were ⌃⌥⇧4 / ⌃⌥⇧6 / ⌃⌥⇧5. The ⌃⌥⇧ prefix is retained so we don't clash with the reserved system shortcuts. A one-shot migration runs on first launch of v0.8.0: if your persisted shortcut still matches the previous default, it is rewritten to the new default; explicitly-customized values are left alone. After migration, the schema is marked and never re-runs.
  • Editor window minimum size now respects the toolbar. The toolbar's natural width is measured at runtime and used as the window's contentMinSize, so resizing can no longer hide any toolbar control behind the chevron overflow.
  • Menu bar version string reads CFBundleShortVersionString from the app bundle instead of a hand-maintained constant. The v0.7.0 build shipped showing Bokashi v0.0.0 because that constant had never been bumped; with this change the menu follows MARKETING_VERSION automatically on every release.

Notes

  • Builds on macOS 14+. No bundle identifier or signing identity changes; the v0.7.0 → v0.8.0 update goes through Sparkle's normal signed appcast flow.

0.7.0 - 2026-05-12

First release that exercises Sparkle's full download / install flow: users on v0.6.0 should see "Check for Updates…" surface this version and walk through the prompt, signature verification, and relaunch.

Added

  • Five line-width presets in the annotation toolbar. Replaces the previous thin / medium / thick ladder with hairline (1pt) / thin (3pt) / medium (6pt) / thick (10pt) / heavy (16pt). medium keeps its 6pt width so existing default-styled annotations look identical. (#27)

0.6.0 - 2026-05-12

This release ships Sparkle 2 auto-update infrastructure and nothing else user-facing. Existing v0.5.0 installs predate Sparkle and cannot self-update — run brew upgrade --cask snaka/tap/bokashi (or download the v0.6.0 DMG manually) one last time. From v0.6.0 onwards, "Check for Updates…" in the menubar talks to the signed appcast at https://snaka.github.io/Bokashi/appcast.xml.

Added

  • Sparkle 2 auto-update plumbing. New Check for Updates… entry in the menubar invokes Sparkle, which fetches a signed appcast, verifies the new DMG's EdDSA signature, and offers to install. (#25)
    • Automatic background checks are intentionally off in this release (SUEnableAutomaticChecks = false); a Settings toggle to flip them on is on the roadmap.
    • The release pipeline now signs every DMG with Sparkle's sign_update and appends an entry to docs/appcast.xml so the public feed always lists the latest version.

Privacy

  • The Sparkle feed is hosted on GitHub Pages from this repo. The only network call on update check is GET against that public XML; no user identifier is sent.

0.5.0 - 2026-05-12

Skips v0.4.0 — that tag was already published earlier and is left untouched. This release bundles everything since v0.3.0 under a fresh minor version.

Added

  • User-defined custom mask terms. New Custom Terms tab in Settings (⌘,) where you register literal strings — your own name, company, product code, etc. They are masked automatically when found in a captured screenshot. Matching is case-insensitive and matches the substring exactly as entered, so registering Hoge masks just the Hoge portion of HogeFuga. Persisted as JSON at ~/Library/Application Support/Bokashi/custom-terms.json. (#17)
  • Bulk-register custom terms from a screen selection. Drag-select a region on screen and Vision OCR extracts every visible line; pick which ones become custom terms via a checkbox list with inline edit. (#17)
  • Multi-display support.
    • Region selection now covers every connected display; the overlay spans every screen and capture routes to the correct one. (#18)
    • Full-screen capture targets the display the cursor is currently on instead of always grabbing the main display. (#19)
  • Click-to-pick capture window with hover preview. Mission Control-style overlay replaces the menu-driven window list: hover any window to see a blue outline and the window's actual contents drawn inside the cutout. Buried / occluded windows still show their real content (snapshots are pre-captured upfront via SCContentFilter(desktopIndependentWindow:)). Click to capture, Esc to cancel. (#20)
  • Optional Ollama vision-LLM detector. New Detectors tab in Settings can route the captured image through an Ollama instance running on your machine for vision-LLM-based PII detection. Configure endpoint (default http://localhost:11434) and model name (e.g. llama3.2-vision, qwen2-vl, MiniCPM-V); the Test Connection button verifies the endpoint is reachable and the model is pulled. Detection failures (unreachable / model missing / malformed response) are silent — the OCR detector still completes. (#22)
  • Developer-mode mask source overlay. A Highlight mask source toggle in Settings outlines each auto-detected mosaic with a colored dashed border in the editor — blue for OCR-based detection and orange for Ollama. Useful baseline for tuning prompts and comparing detectors. (#22)

Changed

  • The OCR / regex / NLTagger detection path is now one implementation of a new SensitiveRegionDetector protocol. Future detectors (local CoreML model packs, etc.) slot into the same pipeline. No user-visible behavior change. (#21)

Privacy

  • All detectors run on-device. The Ollama detector talks only to the endpoint you configure (default localhost); no cloud-LLM integrations exist and none are planned.

0.3.0 - 2026-05-04

Added

  • One-click sensitive-info auto-mask. A new Detect button in the editor toolbar runs Vision OCR on the captured image and turns each detected match into a mosaic annotation as a single undo step. Detection covers:
    • Email addresses (regex)
    • Phone numbers (NSDataDetector, multi-format)
    • Postal addresses (NSDataDetector)
    • Personal names (NLTagger with .joinNames, supports Japanese)
  • Click-to-mask any text region. With the Mosaic tool selected, a click on any OCR-detected text masks just that text region. Drag still produces a freeform rectangle.
  • "Auto-mask sensitive info on capture" toggle in the menubar. When on, every capture opens with the auto-mask pass applied. Persisted via UserDefaults.
  • App icon and menubar icon. Placeholder rounded-square mark generated by Tools/generate_icon.swift (run it again to refresh).

Changed

  • The Detect button shows a spinner while the upfront OCR scan is in flight, not just while detection is running, so the editor never silently waits on Vision.

Privacy

  • All detection runs on-device via Apple's Vision and NaturalLanguage frameworks. No image or text leaves the machine.
  • Place names, organisation names, dates, and URLs are intentionally not in the default mask set — they are usually not private and masking them by default would be noisy.

0.2.0 - 2026-05-04

Added

  • Mosaic annotation tool — drag a rectangle with the new Mosaic tool to pixelate a region. The mosaic is treated like any other annotation: it participates in undo/redo and bakes into the rendered image at native resolution. Block size is fixed at 16 image-pixels for this release.

Changed

  • The editor window now opens at the captured image's native point size, clamped to 80 % of the visible frame in either axis. Region and window captures stay 1:1; full-screen captures get a more generous initial window than the previous 2/3 cap.

0.1.0 - 2026-05-04

First public release.

Added

  • Menubar app with global hotkeys (built on top of KeyboardShortcuts).
  • Three capture modes:
    • Full screen — default hotkey ⌃⌥⇧4.
    • Window — menu picker that lists currently shareable application windows and snapshots the chosen one at native resolution.
    • Region — drag-to-select transparent overlay, default hotkey ⌃⌥⇧6.
  • Editor window opens after every capture; nothing is written to disk unless the user explicitly saves.
    • Copy & Close (default action, ⌘W / red close button) copies the annotated image to the clipboard.
    • Save… (⌘S) opens an NSSavePanel; on confirm, writes a PNG and posts a banner notification with Reveal-in-Finder.
    • Discard (Esc) drops the capture.
  • Annotation tools:
    • Arrow, box (filled / outlined), ellipse (filled / outlined), line.
    • Eight color presets (red / orange / yellow / green / blue / purple / near-black / near-white).
    • Three stroke-width presets: thin / medium / thick.
  • Undo / redo via the window's UndoManager (⌘Z / ⌘⇧Z).

Notes

  • Distributed as ad-hoc signed zip from GitHub Releases. macOS Gatekeeper will warn on first launch; right-click → Open (or System Settings → Privacy & Security → Open Anyway) gets past it.