HyphenScreen: recording, editing and AI animation — version 1.9.5
HyphenTech · Field notes

HyphenScreen: recording, editing and AI animation — version 1.9.5

33 min read
Categories:
Tags:
Summary

Screen recording and editing with a Resolve-style timeline, automatic redaction and pre-export checks.

Current public release

Versions, changes and supported installers all come from the same public release records.

Version 1.9.5: macOS menu-bar Camera Video Effects restored, persistent recording switches and single-icon MCP bridges

  On macOS, opening the webcam while recording normally surfaces the green menu-bar Camera Video Effects item for Center Stage, Portrait background blur and Studio Light. In 1.9.4, splitting the desktop binary behind an execv launcher prevented macOS Control Center from attributing the active camera stream to com.hyphentech.hyphenscreen, hiding that menu-bar control. Version 1.9.5 consolidates the bundle executable into a single binary linked directly against Electron Framework: desktop launches invoke ElectronMain under the canonical bundle identifier so the menu-bar Video Effects icon returns immediately, while headless api, info, pack and help commands enter ElectronInitializeICUandStartNode directly with zero LaunchServices registration, a single Dock icon and zero idle GPU/network helpers. Microphone, system audio and camera switches—along with selected devices—now persist across app restarts so your recording setup stays ready as soon as the window opens.

  Adding a standalone intro or outro card to a 1080p or 4K screen recording used to drop Source exports down to 720p because the synthetic gap canvas was hardcoded to 1280×720 and counted toward the minimum timeline resolution. Excluding synthetic gap canvases from output sizing—tested on the installed Mac app with a 1080p recording followed by a standalone title card—keeps native Source export locked at 1920×1080.

  Building on the seven whole-video packaging layouts (opener, chapter, subhead, list, focus, metric and outro) and explicit 2.5D camera rigs introduced in 1.9.1, on-canvas selection bounds and drag handles now project forward and backward through active camera zoom, pan and parallax depth so dragging text mid-move stays locked under the pointer. Cross-cut coaxial seams avoid double-zoom overshoot, and switching across the ten master themes snapshots original layer colors so round-tripping between dark poster palettes and light paper palettes preserves card hierarchy and semantic red/green contrast.

  Everyday editing fixes round out the release: word-level transcription keeps natural sentence segments instead of creating one segment per word, bringing automatic repeated-sentence speech cleanup back online while extending sensitive-info redaction and auto-zoom across upper video tracks. Compact 1280×800 laptop windows keep all eight right-rail icons, the 0:00 timeline ruler and empty-state hints unclipped; the right-side animation library renders a clean, equal-width two-column grid across shots, packaging and graphics without horizontal scrolling, animation editor tabs fit on a single row, and inserting a fade-in card automatically advances the playhead to a visible frame. The Episode Workbench remains a signed-license feature not yet open for purchase, while whole-video packaging, AI animation creation and explicit camera rigs are open to everyone.

Download and updates: HyphenScreen releases

  The signed Mac 1.9.5 build was verified on physical hardware for Control Center camera attribution, persistent recording switches, single-icon MCP coexistence and native export; Windows and Linux currently continue to offer the verified 1.9.4 installers listed in the release download table.

Version 1.9.0 renders shots in parallel and locates export issues

On an Apple M5 Pro, rendering the same batch of four two-second shots took 103.81 seconds with one worker, 98.98 with two, and 78.41 with four. Four workers cut the measured time by about 24.5% in this shot-generation test. The workers can process separate shots at the same time and place the results in their original order. Each run produced 240 frames, with minor pixel differences in some frames from the parallel runs. Macs with at least 16 available logical cores and 32 GB of memory use up to four workers by default. Other configurations and platforms default to one.

In the Episode Workbench, open Finished Video Review and choose the exported file from the current project. The check reads the file for loudness, true peaks, black frames, silence and freezes. Choose Locate on a timed finding to jump to the editor timeline. The example below detected a black stretch at 5 seconds in the export and jumped to 7 seconds on the timeline because the project trims an earlier section. The Episode Workbench is a professional feature that requires a signed license and is not yet available to purchase.

Finished-video check and timeline navigation in the test workspace

The built-in model tool and the external editing API both natively exported the same ten-minute mixed project at 1080p and 30 frames per second. All 18,000 decoded video frames and the decoded audio matched across the two outputs. Both checks flagged low overall loudness, a high true peak and two near-black stretches traced to the source footage. The excerpt keeps the exported audio and shows an authorized screen recording, editable chapter labels, an upward push and a zoom transition.

The built-in tools and the external API can both find live footage, screen recordings and source images in a project, and filter shot templates for continuous video or screenshots. The benchmarks, sample and timeline navigation above were all tested in the Mac development environment. I then ran the signed Mac app from the disk image and exported the same project in 71.90 seconds. The output contained 18,000 frames and the finished-video check returned the same four findings.

Excerpt from the actual ten-minute mixed-project export ↗

View releases and downloads for each platform

The Windows and Linux packages passed cloud builds on their respective platforms, along with package resource checks and verification of the public installer files. Desktop interaction on both platforms still needs testing on physical machines.

1.8.9: export an episode from the command line or an external agent

The same episode project generated with DeepSeek can be exported from the Studio, the command line or an external agent. The command line reads the native timeline and the Studio’s saved export settings, including speech-aware music ducking. The external exportVideo tool can use Studio quality presets for animation-only projects.

On the installed Mac build, the Studio, command line and external agent exported the same six-second 1080p60 DeepSeek project. All 360 frames and the decoded audio matched. A separate 16-second project with music and narration also matched between the Studio and command line. The Windows and Linux installers passed cloud builds and package checks.

View the current public release and downloads

1.8.8: DeepSeek can recover from invalid animation tool arguments

DeepSeek could answer a chat message and then stall when it sent an unsupported argument to the animation catalog. Version 1.8.8 explains which argument failed and which parameters the tool accepts so the model can correct the call and continue. The app reports interrupted requests and repeated invalid calls.

For episode animation, the app checks the saved plan and license before writing. It verifies scenes, timeline clips and segment links at completion. The Episode Workbench requires a signed license and is not yet available to purchase.

On the installed Mac build, a test reproduced the original invalid arguments. DeepSeek then handled the subsequent requests over its official API, corrected the call and returned real animation candidates. The Windows and Linux packages passed their respective cloud builds and bundled-resource checks.

In a Mac test project with a signed license, the model created two animation clips and linked them to the episode plan. After reopening the saved project, the editor exported six seconds of 1080p60 video and all 360 frames decoded successfully.

1.8.7: previews now show the selected palette

Changing the colors of a packaging layout could leave its default cover on screen. The production renderer now supplies 112 previews, one for each layout and palette combination. Shot candidates can also play their matching rendered examples.

The Episode Workbench compares the current timeline with a proposed look, for a single segment or the whole episode. It is a paid feature and is not yet available to purchase. Existing signed licenses continue to work. Recording, editing and ordinary animation assets remain available.

Open the HyphenScreen menu in the top left and choose AI Settings to configure DeepSeek on its own page. Enter your own API key to load the official model list, choose a model, and test chat and tool calling. Your key is encrypted and stored locally.

1.8.3: write the episode contract, then inspect what actually changed

The new Episode production and acceptance entry sits directly beside Create animation. It preserves the original outline, model assumptions, required and optional requests, explicit prohibitions, narration for every segment, exact on-screen text and acceptance criteria. A plan with an outline but no itemized requirements is rejected. A delivery note that only says “make it polished” gives nobody a usable delivery address; the same applies to an episode that must be checked later.

Actual episode production and professional acceptance interface

Acceptance checks the real project for missing segments, shot links, media sources, timing, visibility and planned screen text. Model-based semantic review is stored separately. The first tab lists Problems and changes; the other tabs cover segment analysis, requirement tracing and human review. Repairs can target only the affected segments, followed by a comparison of new, persistent and resolved findings. The latest 20 reports remain in the project and can be exported as JSON or Markdown.

Professional execution is controlled by Ed25519-signed authorization. Public installers display the feature and its scope, but cannot generate reports, export them or run another review without authorization. Internal agents and external MCP calls use the same check; ordinary animation, recording and editing remain available. Authorization is enabled on my machine, and the private key is neither committed nor included in an installer.

A structural pass does not certify factual accuracy, visual taste, listening quality or audio-video synchronization. The useful safety net is narrower and more concrete: it connects what was requested, what the model attempted and what actually reached the project, instead of letting an unfinished attempt announce completion.

The final suite passed 4,857 tests with two environment-related skips. The installed Mac build passed both authorized and public-locked interface checks, plus external MCP rejection, authorized execution, saving and readback. Windows and Linux passed build, packaging and packaged-native-component checks; their desktop interfaces were not tested in this round.

1.8.2: explanation-card previews fixed

After 1.8.1 grouped the animation sidebar into shots, suites, explanation graphics and raw materials, the packaged build showed eight broken preview cards under explanation graphics while the dev build looked fine. The cause: those preview media ship inside the app’s Resources folder, but the cards resolved them with the dev server’s relative path, which does not exist in a packaged build. Thumbnails now resolve by where the resource actually lives, and both builds show the real film thumbnails; the same fix covers the episode-plan pane.

Mac is now on 1.8.2. The installer passed strict signature verification, and the installed version and full file-tree hashes read back identical; four new asset-path unit tests pass. Windows and Linux remain on 1.8.0 and will pick up this fix with the next cross-platform build.

1.8.0: verify completed animation work and make model statistics readable

A model could announce that it was about to write a scene, then end with only part of the project created. ChatGPT subscriptions often completed the animation but lacked usage and speed information. This update checks completion against the project and statistics against actual received data.

The AI panel preserves waiting, structure reception, tool execution and completion records. Live and completed turns share elapsed time, first output, visible text and reasoning, reception speed, tool count, latest input and cumulative usage. ChatGPT subscriptions now receive reported cache and reasoning tokens too. Unavailable fields are labeled. After completion, input is usable again; reopening chat retains the statistics.

Actual model-statistics interface

Reception speed measures actual character deltas, excluding tool execution and the initial wait for each call. Transport buffering affects it, so it is not hardware inference throughput. Latest input is one request’s payload; cumulative usage adds the requests across tool exchanges, like one parcel’s weight versus a day’s shipping total.

Animation creation verifies scenes, clips and explicitly requested time ranges. Missing work can be attempted twice within the original deadline. Parameter failures identify fields; repeated failures and stalls end explicitly while preserving completed edits. Rewriting the same scene at the same position avoids duplicate clips and protects manual trims. MiniMax Flash receives the selected reasoning effort, and original thinking, signatures and tool messages are retained.

Real ChatGPT and Poe runs completed three scenes and an appended ending. MiniMax preserved its work after a service error and completed verification on continuation. The local Sushi long task reached two scenes, with the third unfinished; structural errors and timeout records remain visible. Complex stories still require a suitable model.

These three complete 16-second native test outputs follow the same sequence: read sources, check information, generate animation and continue editing.

ChatGPT native animation test ↗

Poe native animation test ↗

MiniMax native animation test ↗

The full suite passed 4,777 tests with two environment-related skips. Mac was installed from the actual DMG; all 987 files, directories and links matched, with strict signature and image verification. Two real read-only ChatGPT turns in the installed workspace ended automatically in about 9.2 and 12 seconds; subsequent input, retained statistics, continuous 16-second playback and a middle seek passed. All three projects were re-exported by the installed 1.8.0 build, with 480 frames fully decoded and sampled per scene. These silent fixtures verify the workflow rather than complex art quality; they add no subjective sound review.

The Mac DMG is 426,742,409 bytes, about 427 MB and only 801 bytes larger than the previous release. Windows and Linux have independently published 1.8.0 too; downloads and checksums reflect their actual manifests. Mac verification does not establish desktop results on those platforms. Existing features, screenshots and demonstrations are retained.

Windows / Linux 1.8.0

Windows/Linux 1.8.0 unifies running and completed model metrics and checks actual scenes, clips and target timing before reporting animation delivery. Incomplete work gets at most two completion retries; corrections at the same scene position no longer duplicate clips. Windows passed 4,758 tests with 21 environment-related skips. Packaged resources and ten native MP4/GIF exports per platform passed verification and full decoding. Linux uses WSL software rendering. Apps were not installed; desktop capture and model quality were not retested.

Windows / Linux 1.7.8

Windows and Linux are now locally built as 1.7.8 with the production workbench, dynamic Poe catalog, local continuation compaction and waiting fixes. The Windows suite passed 4,723 tests with 21 environment-related skips; typechecks and production builds passed. Packaged resources and ten native MP4/GIF exports per platform passed verification and full decoding. Linux uses WSL software rendering. Apps were not installed, user projects were not modified, and desktop capture and model generation quality were not retested.

1.7.9: save completed edits immediately and make usage clear

A scene could already be written while chat kept checking, yet only appear on the timeline after Stop. Version 1.7.9 delivers each validated edit to the editor for display and saving while further checks continue. The same final snapshot is not applied twice, and manual-edit and project-switch conflict protections remain.

Actual paper-style animation exported by installed version 1.7.9

Scene catalogs now send an index rather than every existing scene’s complete structure. Successful writes use compact receipts, with exact layers available on demand. Full keyframes, curves, original records and output budgets are retained. In a real Sushi read-only task, the last request used 26,294 input tokens instead of 72,593, a 63.8% reduction. Cache state and tool counts differed, so the timings are not an animation-generation speed benchmark.

The header distinguishes the complete-request estimate from the latest actual input; cumulative turn usage stays separate. Automatically detected context windows refresh before a request while explicit settings are kept. Missing historical counts are not invented, and repeated edits to one scene are no longer counted as multiple scenes.

Three-second native animation export ↗

Play the native 1.7.9 animation sample

The same three-second, 45-layer scene produced 90 identical decoded frames in the old and new installed builds. The full suite passed 4,751 tests, followed by 161 targeted checks after the final catalog change. Installer files, signatures and image checks passed; I verified the installed UI and confirmed it works. Mac is now 1.7.9; Windows and Linux stay at the published 1.7.8. Existing features and historical images are retained.

1.7.8: inspect animation work and connect Poe directly

The production workbench shows the current stage first. Expand it for timestamped tool records, native scene structures, arguments and returned results, with copyable previews. Scene writes report their actual layer count and duration; writing a scene, exporting it and checking playback remain separate steps.

Poe is now a dedicated provider in AI settings. Use your official API key and choose from the live catalog filtered for text output and tool calling. Reading the catalog does not issue paid generation probes; test the selected model to verify account access. A real Poe account completed tool round trips and a three-second, two-layer native scene with readback. The small test model initially guessed invalid internal style IDs; explicitly reading the catalog first resolved that attempt. Complex work still depends on the model and the request.

After a local timeout or output limit, the next message or Continue action compacts earlier history as needed without a separate manual step. The transcript and authored scenes remain intact, and an ineffective summary leaves the original history in place. In an independent local test, estimated history decreased from 6,841 to 1,904 tokens and the conversation continued successfully.

Connection status is no longer counted as reasoning, and compatible local reasoning fields are read correctly. Average received characters per second, conversation estimates and provider-reported token usage are distinguished. Animation follow-ups retain their wait settings, and MiniMax reads its official live catalog. Source regression passed 4,742 tests with two skipped. The Mac app is installed: all 609 files and links match the build, with strict signature and branding checks passed. Its installed API read back a scene and exported a three-second 1080p30 animation with complete decoding verified. A keychain authorization wait followed by a system screen-capture error prevented final desktop click acceptance. Windows and Linux remain at the published 1.8.0.

Windows / Linux 1.7.7

Windows and Linux are now locally built as 1.7.7 with live metrics, context-window detection, persisted turns and output-truncation fixes. The Windows suite passed 4,703 tests with 21 environment-related skips; typechecks and production builds passed. Packaged resources and ten native MP4/GIF exports per platform passed verification and full decoding. Linux uses WSL software rendering. Apps were not installed, user projects were not modified, and desktop capture and model generation quality were not retested.

1.7.7: corrected metrics and a larger output budget

Two quantities in 1.7.6’s live readout were wrong, and this release fixes them: output characters now count the visible answer only, with reasoning shown separately; speed is measured over both, so a model that thinks first and calls tools no longer reads as a near-zero rate. When a provider reports only output tokens, the chip says “N out tokens” instead of inventing “0 in”.

The output budget moved too: 20,000 tokens per response for Anthropic-wire providers (MiniMax and friends) and 32,768 for local models, both overridable by environment variable. Every provider now stops explicitly when a response hits max_tokens/length and keeps the reasoning, tool results and text it already produced, instead of surfacing only “Empty response from model”. The animation prompt also asks the model to keep each scene small, split oversized scenes, and act after one short plan.

The features below arrived in 1.7.6 and remain in 1.7.7.

1.7.6: the chat keeps its trail, and unfinished turns can continue

The hard part of a model turn is rarely the waiting — it is not knowing what it is doing or where the time goes. The chat now keeps a live row above the composer: status, elapsed time, first-character latency, output characters and speed, tool calls, and context usage. Token counts appear only when the provider actually reports them.

The Create animation dialog after the layout pass: wait settings on one row, a larger close button

The production record is now time-stamped stages and tool calls — reading context, writing scene N, the scene read-back, every tool success or failure. That record and the model’s reasoning are saved with the message on this machine, so switching conversations, compacting, or restarting the app still leaves them readable.

An unfinished turn no longer leaves only an error. Its reasoning, streamed text and tool results are kept as an Unfinished · continue message; Continue picks up from there, and whatever already reached the project stays.

Context is no longer capped at 20 messages. When the local endpoint reports its window (方寸智匣, llama.cpp, vLLM, LM Studio) it is read automatically; otherwise set it in AI settings. Near 85% of the window, the older turns are compacted before sending. Compaction changes only the history sent to the model — it never deletes the conversation or the reasoning. Scrolling back no longer snaps to the newest delta, and switching to Media or Rec keeps the turn running; Edit brings it back.

The Create animation dialog was tidied as well: the wait settings collapsed into one quiet row, the close button grew from 28 to 36 px with a border, and continuous scroll now explains itself — one shared world and through-line objects with the camera following the action, while detail is still set by how you split the script, the duration and your custom requirements.

This round passed 4,722 tests with two skipped. The Mac package is 426,788,073 bytes (SHA-256 below). The installed build was checked for the entry, the dialog, the context settings and keeping the panel across tabs; the system keychain prompt was declined during verification, so no model generation ran in the installed app. Window detection and session reload are covered by unit tests, including the 方寸智匣, llama.cpp, vLLM and LM Studio response shapes. Windows and Linux remain at 1.8.0.

1.7.0: put your own signature on exports

For a course, demonstration or review copy, choose what appears in the exported corner. Export watermark defaults to 黑粉录屏, with a direct off switch. Expand customization to use a channel name, website, review note or local logo.

Actual shared watermark settings component, isolated browser verification

Use text, a local PNG/JPEG/WebP image, or both. Images retain their aspect ratio. Five positions cover the corners and center, with size and opacity controls. The position diagram is illustrative; exported text adapts to background brightness. Settings are saved on actual export. MP4 and GIF share the native overlay path, while ordinary editing previews remain unchanged. Project-authored text and logos are not removed by this switch.

Example Default content Use
Software default 黑粉录屏 Software credit
Channel credit 黑粉科技 · 原创 Replace with your channel
Website corner hyphentech.top Your homepage
Review copy 样片 · 仅供审阅 A faint centered review mark

Actual native outputs with default, disabled, custom text and image watermarks

These are frames from actual files. The channel test deliberately uses top left to check positioning; its UI preset remains bottom right. Landscape, portrait, light backgrounds, Chinese text and image combinations were checked. No payment, account or license entitlement is active; future commercial terms remain undecided.

The release candidate passed 4,660 tests with two skipped, plus native watermark checks and decoding/frame inspection of 12 outputs. macOS screen capture errors prevented desktop mouse-driven export, the system save dialog and reopening from being marked as passed. The existing GIF centisecond timing deviation was not fixed. The fixtures are silent, with no new subjective sound review. Mac advances to 1.7.0; Windows/Linux remain at 1.6.6 with their earlier watermark behavior.

The DMG is 426,722,449 bytes, approximately 427 MB, about 1.22 MB larger than 1.6.9. Image integrity, strict signing and comparison of 609 files/links passed; the local certificate is still not Apple notarization. The native module inside the package exported eight MP4s and two GIFs, all fully decoded. Default versus disabled changed 2,189 corner pixels, with no residual mark in the disabled corner; the contact sheet matches the development result byte for byte. This task did not replace the locally installed application.

1.6.9: watch the result before editing your explanation

Local composition now covers source evidence, processes, presenters, before/after comparisons, metrics, checklists, timelines and structures. Open Create animation → Local composition, play a real example, then edit this episode’s text, values, images and event timing. Generated native layers and keyframes remain editable. These fixed frameworks run locally without a model request.

Actual output from eight types; values are illustrative

Paper Signal and Dark Ink provide 16 actual 12-second examples. The following 96-second silent reel covers all eight Paper Signal types, demonstrating reusable structures and actions.

Complete silent reel of eight local composition types ↗

Open the complete silent reel separately

Presenter scenes use six complete illustrations of one character with consistent identity, clothing and proportions. A shared explanation image supplies the default; each stage can override it while retaining its original aspect ratio. Restore shared clears only that stage’s binding. The fixed image poses support looking, pointing and reacting, rather than arbitrary character acting.

Actual stage-image editing and save feedback

The inspector responds to its own width, normally placing stage cards in two columns. Image buttons no longer collapse and the Apply area does not cover settings. Native output now decodes WebP, avoiding images that appear in the browser but disappear from export.

All 16 native 1920×1080, 30 fps, 12-second outputs fully decoded; the six presenter stages were frame-checked. Desktop checks covered the system file chooser, applying a fourth-stage image, disk readback and a seek to about 6.3 seconds. Only the chosen stage changed, and undo restored its original composition. There were 4,650 passing tests and 2 skipped. Examples are silent; subjective listening, physical recording, every seek position and other systems were not retested. Windows and Linux builds are prepared separately; their current public packages remain 1.6.6.

Mac 1.6.9 passed image integrity, strict signature verification and a comparison of all 609 files and links. The native compositor in the installed application exported two new 1080p30 examples for stage images and metrics; both fully decoded. The installed studio launched, but system capture errors prevented the final interface-click and continuous-playback checks. The interface checks described above were performed in the independent development studio.

1.6.8: edit content and event timing locally

Mac adds Local composition inside Create animation. Choose source evidence, a process or presenter actions, enter content and timing, and create native animation without a model request. It suits recurring explanation segments whose layout remains consistent while screenshots, steps and actions change with each episode.

Actual local composition editor

Evidence retains the original local image, source name, URL and verification date, then introduces and focuses it at specified times. Processes support two to five steps with individually editable timing and existing project sound effects. Presenter scenes provide preset look, point, receive, place, react and rest actions. A local declarative compiler handles these tasks without executing model-generated code; the existing model creation workflow remains available for free-form design.

Actual content and timing controls

The Content tab edits ordinary text and local images, plus sources, steps, colors and event times in composed scenes. Actions and sounds share events. Animation can stay on the project ruler or follow one main recording clip that fully contains it, through moves, splits, trims and deletion. Consecutive compatible compositions reuse a track when they do not overlap, without crossing an intervening picture track.

Finalize content and timing before manual layer refinement. Recomposition rebuilds controlled layers, retains manually added layers and refuses to overwrite locked controlled layers. This is not a general character skeleton or arbitrary long-story generator. Independent narration and auxiliary-video source binding are outside this revision.

Desktop development checks covered generation, title and 4.25-second timing edits, saving, switching projects and reopening with edits retained. Two different 24-second projects and a 12-second presenter project passed seeking and continuous playback. All three native 1920×1080, 30 fps exports fully decoded. There were 4,637 passing tests and 2 skipped. Objective audio checks passed; subjective listening remains incomplete. Other systems and physical recording were not retested. Windows and Linux remain at 1.6.6.

1.6.7: compact packaging controls on Mac

Mac is updated to 1.6.7. A single segment list combines thumbnails, titles, layouts and timing; the current layout is changed in a dropdown. Fields sit side by side, and still and animated previews share one area. Typical four-segment settings fit at 1280×720 and 1024×768; smaller windows and longer lists can scroll.

1.6.7紧凑界面,实际组件截图

Ten Chinese display names now have four characters, including Red Poster. Existing theme IDs, project compatibility, palettes, animation and sound are retained. Windows and Linux remain at 1.6.6 without this interface update.

This revision passed 4,615 tests with 2 skipped, plus application and test type checks. All 570 installed files and links match the Mac installation image and strict signature checks passed. The installed app re-exported an independent four-segment project: 27.4 seconds, 1920×1080, 30 fps and 822 frames, with complete video and audio decoding. It reused existing media and is not a new generation run. App startup and opening the studio passed. A system screen-capture error blocked further desktop interaction; the author explicitly chose to test manually and continue release. Desktop generation and seeking are therefore not reported as agent-verified in this round.

1.6.6: whole-video packaging for Windows and Linux

Windows and Linux are locally built as 1.6.6 with ten packages and forty layouts. Generation now uses bundled OpenH264 instead of unavailable libx264, without an extra encoder installation. Mac was at 1.7.0 then.

Rendering checks on each platform generated four layouts from synthetic material: 16 seconds and 480 frames, with complete video and theme-audio decoding. All 331 packaging resources in the Windows package matched source bytes. The Ubuntu deb passed installation and file verification. This does not replace desktop capture or device testing.

1.6.5: one coordinated design for a whole episode

An opener, list and chart can each look good yet feel unrelated in one video. Version 1.6.5 adds coordinated video packages: choose a design, then fill several segments with this episode’s content while sharing typography, colors, motion and sound.

Whole-video packaging editor

Open a project and select Whole-video packaging in the AI panel. Ten packages each provide opener, list, focus and metric layouts, with real thumbnails and animated examples. One editor displays the fields needed for the current layout. Light and dark themes are readable; generation actions and progress stay in a fixed footer.

Each group supports 1–12 segments of 4–30 seconds. Enter content and timing, and select a frame from a project video for focus layouts. Generate locally and add the group together without a model call or uploading the recording. Effects default on, music defaults off, and existing narration remains. Saved configurations can regenerate and replace their original group; manually moved, trimmed or locked groups are protected.

A real four-segment, 27.4-second recording passed generation, insertion, replacement, seeking, continuous playback, saving and reopening in HyphenScreen, then exported at 1920×1080 and 30 fps. All ten default palettes and forty layouts were rendered and fully decoded. Objective audio checks found no clipped samples or overlap among the four theme tracks. The author listened to the real-recording edit and confirmed normal sound. Other palettes were not individually rendered; the other nine examples do not have separate subjective listening approval.

Generated clips can be edited on the timeline, but their internal layers are not individually editable. Change saved settings and regenerate to edit content. Picture and theme audio remain separate tracks, so select both when moving. Mac real-recording checks retain their 1.6.5 scope; see the 1.6.6 section for Windows/Linux packaging generation verification.

Earlier in 1.6.4: steadier camera movement

Version 1.6.4 adds steady and responsive camera styles to zoom regions. Select a region, then choose its focus mode in the Zoom inspector. The steady style suits document explanations; the responsive style reacts faster to movements between controls.

Camera follow settings

Think of the safe area as room for the pointer to move inside the shot: small movements hold the composition, while crossing the boundary starts a smooth pan. Zoom-out holds the last focus. Consecutive clicks are grouped by time and position to avoid repeatedly zooming during one menu operation. New automatic zooms use the steady style; existing projects retain legacy follow. Pointer-art smoothing is independent of the new camera path. Recorded cursor telemetry is required; external videos can use manual focus.

A 48-second comparison used the same real recording and cursor telemetry for all three modes. The isolated Mac workspace passed mode switching, saving, complete playback and seeking. Two returns to 19.94 seconds produced identical preview pixels. Native export produced 1,440 frames at 1920×1080 and 30 fps, with complete decoding and extracted-frame inspection. The initial preview canvas path was also corrected. This is a Mac camera-motion check; other systems and the full capture-device matrix were not re-tested.

The installed Mac app returned all three camera modes and exported the 48-second comparison with full decoding verified. Desktop checks covered continuous playback after seeking to 20 seconds and another seek to 35 seconds. The public installer was downloaded again and matched the local checksum; it adds only about 14 KB over 1.6.3. Windows and Linux 1.6.4 have now been built locally on the Windows workstation and WSL. The Windows package passed a synthetic three-mode native export and full decoding; the Ubuntu deb passed installation verification. Desktop capture workflows on those systems still need separate checks.

Earlier in 1.6.3: 230 purpose-based animation entries

The library consolidates the previous 316 discovery entries into 230 across nine categories: text and titles, data and charts, explanations, AI concepts, flows and architecture, intros and outros, social media, picture effects and hardware. Common is a shortcut filter rather than an additional category. Source assets, original IDs and existing projects remain compatible.

HyphenScreen 1.6.3 purpose library and real chapter preview

Shared entries include 22 title styles, 13 chapter styles, ten people profiles, six quote cards, 18 lower thirds, 11 column comparisons, ten line trends, four progress rings and six terminal demonstrations. New examples include statistical rings, terminal simulation, code typing, chat, ranking races and heatmaps. Default text and URLs use Chinese HyphenTech examples; necessary third-party licensing and provenance remain included. Blinds and clock-wipe transitions leave discovery while old-project resources remain available.

Consolidated title, chapter and profile entries

Choose a purpose, insert the material, then select a style in the animation panel. Independent animation and overlays retain their different editing paths. Some sampled-vector internals require regeneration; this is not a promise of arbitrary source-parameter recomputation. Model tools now correctly find reference variants inside mixed-source families when filtering by storage kind.

The development editor passed category browsing, Chinese search, chapter insertion and animation playback. Three cross-source switches passed real API save and fresh-profile reopen checks with clip start positions preserved. The complete suite passed 4,596 tests with two skipped. This does not mean every style received a manual switching review; full recording, physical camera, face restoration and online generation were not repeated.

Version 1.6.3 also includes the local 1.6.2 candidate’s multiselection reordering, track organization and package deduplication. Research sources, reference runtimes, sampling caches and test projects are excluded from the installer. Apple Silicon Mac, Windows and Linux now offer 1.9.0. Package size and checksums are listed below.

The signed 1.6.3 installer passed disk-image and resource checks; all 427 files and links in the installed app matched the image. The installed binary returned 230 entries, inserted and switched a native material, saved the project and exported a 1920×1080, 30 fps video that decoded fully. Public-download readback is recorded separately. Subjective audio review was not completed.

Earlier timeline and audio improvements

Mac 1.6.1 stops stale programme audio before preview returns to live mixing after timeline edits, and delayed playback requests recheck the current mode. Five music groups in the saved real project do not overlap; a regression reproduced the stale-audio path. Two-way handoff, delayed startup, isolated desktop playback, clip movement and saved-project readback passed. A real native export decoded fully and its music level showed no duplicate mix. Subjective listening review remains incomplete.

This release also fixes animation entrances, exits and source labels flashing over later recordings after earlier jump cuts. Animation placements now map from their actual editor-ruler position to source footage in both preview and export. The project format is unchanged; existing animations and manual edits do not need to be rebuilt. The fix is now available on all three platforms in 1.6.3.

Recordings lifted into separate picture rows also move horizontally in absolute time instead of repacking the entire main sequence to the start. In an isolated development desktop, a 10–20-second clip moved about three seconds later and back while the other clips stayed at 0–10 and 20–30 seconds. Vertical lifting and saved-project readback passed. The final full suite passed 4,535 tests, with two skipped; installation and public availability are tracked separately in the release record.

A fixed project snapshot reproduced the unwanted logo and webpage overlays in native preview and a short MP4 export; both disappear with the corrected mapping. Regression coverage spans main recordings and imported video. Both TypeScript checks passed. These short-segment checks do not replace full-film review, subjective audio review or desktop acceptance on every platform.

Version 1.5.9 makes jump cuts explicit: Add Jump Cut starts with the selected main clip’s full range and offers adjustable endpoints, reset and restore. Deleting a main recording shifts later independent visuals, audio and markers; source-anchored captions and effects keep their existing mapping.

Replay after a tail cut and active-source identity were corrected, and removed players release their resources. Targeted tests and development desktop checks passed, but the reported long-session player-creation failure was not reliably reproduced; complete resolution is not claimed. The editing API also consolidates compatible non-overlapping visuals. A private episode reduced 51 animation tracks to 3 while preserving all 51 clips and the order of 29 overlapping pairs. This is an episode-specific result, not a guaranteed ratio; the original edit was preserved.

Version 1.5.8 fixes preview in projects containing native animation, narration and music but no source video. Initialization could incorrectly send the first audio asset to the video decoder. Preview now selects the actual visual source from the compiled timeline, including the generated animation canvas, rather than an unrelated asset.

External scene-by-scene writing compatibility is also fixed. Before packaging, a private copy of a 146.83-second project with 11 animation clips and two audio tracks was checked in the real editor: start, middle and end seeks, plus continuous playback from beginning to end. The installed build passed those seeks and early playback across scenes; subsequent system capture errors prevented reading back its playback endpoint. The original project was preserved. This is a targeted preview check, not a claim that every device, model-generation task or subjective audio review has passed.

Animation generation: adjustable waiting and incremental feedback

Open a project, select your configured AI engine and choose animation creation. Describe the purpose, style, duration and actions; generated scenes can be edited on the timeline. For local generation, select LocalBrain in AI settings and load a model that supports tool calling in LocalBrain.

The pink clock button opens waiting settings. Total and idle timeouts can each be set from 1 to 360 minutes, with 15/45/90-minute shortcuts and settings remembered per engine. LocalBrain defaults to 45/10 minutes; online and subscription engines default to 15/5. Longer waiting allows a slow model more time; it does not make generation faster.

Animation waiting settings

Planning text appears as actual output arrives. Engines with streamed tool arguments show which scene structure is being written and the received character count. After each scene is written, the tool reads back its actual layer count and duration. Engines without argument streaming show real stages only. Intermediate feedback does not mean the project has been saved or the video exported. Automated tests, component checks, production build and package verification passed; complete story generation, playback and subjective sound review remain for user testing.

  • Apple Silicon Mac: 1.9.0, DMG 479,244,752 bytes, locally signed and not Apple-notarized.
  • Windows 10/11 x64: 1.9.0, installer 442,453,485 bytes, unsigned.
  • Linux x86_64: 1.9.0 in AppImage (642,596,147 bytes), deb (573,333,752 bytes) and pacman (530,525,736 bytes) formats, all unsigned.
  • Online model fees are charged by the selected provider; local models consume local hardware resources.

Download Mac 1.9.0 DMG · Download Windows 1.9.0 installer · Download Linux 1.9.0 AppImage · Download Ubuntu/Debian 1.9.0 deb · Download Arch 1.9.0 pacman · All platform releases

Use these SHA-256 values to verify the five downloaded installers:

Package SHA-256
Mac DMG 1d6423ea535f0ab8d62513328b9ffae8a649c5aac360ba67adeec22dd727c68c
Windows installer 5b548dda21efa241a13501538bdf271039788f74fe33931ad09239215790958e
Linux AppImage b5f8094d7b78460569bec265deb3fe1cf54c0da930b0d6343a6c92d9da2cfc4b
Ubuntu/Debian deb 5e7d6eea9ee339427daa75daf913145040afed8497452b7ceae6cd1fd2c83fa8
Arch pacman 705f7a40c2675cdbf6966d4c09410058f6a0f3d70f4fa04a282df3b1d232a8b6

The 1.8.9 release passed 4,899 tests with two environment-dependent skips. The installed Mac build was checked for the dedicated DeepSeek setup and the unlicensed Episode Workbench. Windows and Linux were built, packaged and checked for bundled resources; their desktop interfaces were not tested on physical machines in this release.

Earlier validation: The Windows package passed native component checks and byte checks for 331 packaging resources. Windows and Linux each generated all four layouts from synthetic material, with complete video and audio decoding. The deb installed and passed file verification on Ubuntu 22.04. Linux capture still needs testing in a desktop session with a ScreenCast portal.

Earlier feature tour and development history

The following screenshots and development records describe versions 1.0.1/1.0.2. Their layouts, counts and test results are historical, not the latest interface. Current platform versions and downloads are listed above.

Summary Recording is easy; the work starts after HyphenScreen 1.0.2 · official release · free download HyphenTech · first published 2026-09-16 · updated 2026-09-29 · also in 简体中文 and 繁體中文 (switch below the title)

Let me start with something that made my stomach drop. I built automatic redaction into my own screen recorder: it finds e-mail addresses, keys and the like in the picture and covers them with a mosaic. When it was done, every unit test passed. The boxes were in the right place, and on a still frame they fitted perfectly.

Then I exported a video and watched it from start to finish. During the few seconds the picture zoomed in, the mosaic fell behind, and an e-mail address and a key starting with sk- were plainly readable. It happened once more inside a vertical reframe. The tests checked whether the box coordinates were right; what I actually needed was that nobody could read anything sensitive in the finished video. Those are two different things, with zooms, tilts, crops and encoding in between.

Since that day the app has something most screen recorders do not: after export, it reads the finished video itself back, frame by frame with text recognition, to confirm that what should be covered is covered. This section was written for version 1.0.2; the current release is listed above.

HyphenScreen 1.0.1: the preview is zoomed in on a code block, the key and contact details are redacted automatically, captions below and a watermark in the corner; the timeline gives captions, annotations, zoom, video, voiceover and music each their own track

A word on where the money in screen recording comes from, because it decides where the effort goes. The usual moves: the free version adds a watermark or a time limit and turns export into a paywall; or the tool becomes a cloud service where you upload your screen and pay for storage and exports. Screen recordings are about the most private material there is — your e-mail, your paths, tokens, the chat window you forgot to close. The party that gains is the one holding your footage for storage and transcoding; the cost lands on whoever uploaded it. Desktop software has played this script many times: the one-off licence stops getting updates and features move into a subscription.

Why do I make HyphenScreen? Honestly: I publish videos every week, screen recordings are my main footage, and I do not want my screen on anyone’s server. There is no subscription, no cloud and no storage to sell, so there is no way for it to make money off you; the upkeep is mine. Since 1.0.2 there is a sponsor code for anyone who wants to chip in; sponsoring or not changes nothing in the app. Judge it accordingly: it will not get worse to earn from you, but it could stall if I run out of time.

Facts
Version at the time 1.0.2, official release (2026-09-29)
Built with Electron interface + a native Rust compositor; recognition, transcription and voice isolation all run on your computer
Installer Mac 265.7 MB (Apple silicon), signed with my own HyphenTech certificate, not Apple-notarized
Other platforms at the time Windows 10/11 x64 and Linux x64 received 1.0.1 (without 1.0.2’s sponsor code); unsigned
Download Free and public, with SHA-256 checksums; the source is private

Versions and installer size come from the project’s release notes and my local installer shelf, not estimates

Warning This is a version 1.0.2 record. Mac, Windows and Linux now offer 1.9.0. The Mac build is still not Apple-notarized, and the Windows installer is still unsigned. See the limits below for platform differences.

🧩 The full tour (screenshots from the app)

These are real screenshots, taken on a demo project made for the purpose: a recording of a scrolling web page, synthetic narration and background music. Every e-mail, phone number and key in it is fake — the pictures that introduce a screen recorder should not leak anything either. The two camera shots show my own camera. The recording, camera, multitrack, transition and AI chat shots show the Chinese interface; the rest show the English one.

🎥 Recording: make “what to record” simple

Recording: pick a screen, a window or a region; camera and microphone record in the same take; switch between saved setups

  • The whole screen, one window, or any region you drag — while dragging, click a window to take its exact bounds instead of lining up the edges yourself. Lock the region to 16:9, 9:16, 1:1 or 4:3 so vertical content needs no crop later. Region capture works on macOS and Windows.
  • Camera, microphone and system sound record together, along with the pointer, so the editor can suggest smooth zoom-ins where the pointer lingered. The camera records at the best quality the device offers.
  • Microphone, camera, capture area and camera look can be saved as named “recording setups” and switched in one click.

Screen recorders used to stop at recording: you opened another app to edit, exporting once and importing once, losing quality twice. This one was built as record + edit from day one — the take goes straight onto the timeline without an extra encode.

✂️ A timeline laid out like DaVinci Resolve

The timeline: captions, annotations, zoom, speed, cuts, video, original sound and voiceover each on their own track, every item labelled with what it is

  • Captions, annotations, zoom, speed, cuts, V1 video (add more video tracks above it), A1 original sound, then voiceover and music. Track headers show the name and item count and can lock, hide, mute and reorder.
  • Items say what they are: the annotation text, the template, the redaction category, the zoom factor, the length cut, the caption sentence — no clicking each one open.
  • Voiceover and music waveforms are drawn for the part you can see, at device-pixel detail, so zooming in makes them sharper.
  • Press M for a marker; markers copy straight into chapter timestamps for a video description. A/B for select and blade, Cmd+B to split, J/K/L to play, I/O for in and out — the keys editors already know.

🧱 Multitrack and on-canvas editing

Multitrack editing: three video layers, edge trimming, independent music and the clip effects panel

  • Import several videos and audio files at once. Video tracks have no fixed limit: the upper picture covers the lower ones and all sound is mixed. Drag clips and their edges to arrange and trim; Shift for fine control, arrow keys frame by frame.
  • While paused, select things right in the preview: move and scale the recording and the camera; drag and resize captions, text, images, arrows, blur boxes and templates.
  • Fullscreen preview (F in, Esc out). A heavy project can drop its preview to 1/2 or 1/4 resolution to stay smooth; exports always use full resolution.

🎭 Screen and camera

Camera effects: picture-in-picture, camera zoom, the camera making room when the screen zooms, camera only, side-by-side split, cut-out camera at the bottom

  • Picture-in-picture, side-by-side or stacked split, with one click to swap sides; the screen always keeps its aspect ratio.
  • The cut-out edge follows the real outline in the camera picture (macOS): fingers, glasses and shoulders stay crisp instead of a blurred halo. The background can be a blur, an image, a colour or a gradient.
  • In picture-in-picture the screen and camera make room for each other: when the screen zooms, the camera shrinks into its corner; when the camera zooms, the screen shrinks aside. A shrunken camera is sampled over each pixel’s footprint, so it neither flickers nor blurs.

Camera layout: layout, share, swap and corners — with a clip selected, only that clip changes

  • Every clip can have its own layout and look — side by side for one stretch, picture-in-picture for the next; whatever a clip does not set follows the whole video.

💬 Captions and Chinese line breaks

  • On-device transcription creates captions (built-in Whisper, or LocalBrain), keeping the recognizer’s real timestamps; silence and music markers never become captions.
  • Edit captions by hand, search them, restore the original wording, pick one of six styles. Chinese captions break between words, never inside a word or a model name, and preview and export break in the same place.

🔊 Sound

Sound: voice isolation, music ducking under speech and its three presets

  • Voice isolation (macOS): an on-device neural network removes keyboard, fan and other noise and keeps the voice, with low cut, EQ, compression and limiting.
  • Music ducking: the music dips while someone speaks and comes back after, with natural background, balanced voice and voice first presets, or your own lead-in, hold and release.
  • Split the original sound onto its own track, blade it and set each part’s level (−60 to +36 dB). Cut a video clip in two and the voiceover and music over it split with it and play straight on.

🎨 Visual templates and picture effects

Visual templates: pick by category, drop at the playhead, then adjust every setting

  • 35 templates: title cards, name strips, step badges, key caps, spotlight, result and comparison cards, an outro with a call to action, a corner watermark and more, with search across all of them.
  • 13 text styles and 11 picture effects recreated from my other app, HyphenCut (bold impact titles, handwritten quotes, warm documentary, teal-and-orange, vignette, RGB glitch…); nine fonts ship with the app.
  • 12 transitions (dissolve, dip to black or white, wipes and pushes in four directions, zoom) with the sound fading along, applicable to every cut at once.

Transitions: pick one at a cut between clips, set its length and audio fade

🛡 Automatic redaction that follows the picture

Automatic redaction: the key, e-mail and phone number are covered, and the boxes follow the text as the page scrolls

The hard part of redaction is not finding the text; it is following it. Finding the e-mail is only step one — then the picture zooms, tilts, gets cropped to vertical, and the page scrolls. The mosaic has to move with all of it; one frame late is one leak.

  • On-device recognition of e-mail addresses, phone numbers, ID numbers, bank card numbers, keys and your own keywords (automatic detection is macOS-only for now). ID and card numbers are checksum-verified, so there are few false alarms.
  • Redaction follows zooms, tilts, crops and vertical reframes. On a scrolling page the boxes follow the text stretch by stretch, covering only where it travels. Before 1.0.1 one box covered the whole path the text took — at worst half the frame. Fixed along the way.
  • If you turn the annotation or zoom track off for an export, the redaction track stays on and tells you so — that one is deliberately not switchable.

🩺 Finished-video check: trust the output, not the process

Back to the story at the top. The lesson: however green the checks along the way, that does not make the output right. So after export you can check the finished file itself, and every check is about the output:

  • Loudness and true peak, with a suggested output level instead of guessing by ear
  • Stretches of silence and black, each one marked
  • Frozen pictures: the output stops while the source keeps moving — the classic sign of a broken export
  • An output shorter than its timeline: a clip running past the end of its media, located at the moment the media runs out
  • Captions and titles that do not fit: anything cut off or outside the frame
  • Text recognition reads the picture back, frame by frame, to confirm nothing sensitive shows (macOS for now)

The numbers below are from a measured receipt on the installed 0.4.13 (exported at 1280×720/30fps on the acceptance project through the app’s own interface, then checked; the program wrote the numbers, I did not copy them by hand):

Check Result
Export time / check time 12.6 s / 34.8 s
Loudness · true peak · dynamic range −18.2 LUFS · −2.8 dBFS · 3.7 LU
Text read-back one frame per second, 150 frames analysed
Verdict on the normal output 0 problems, 0 warnings
Re-checked with the redaction deliberately removed e-mail and key recognised from 165.0 to 181.2 s; a red marker placed at 172.08 s: “Check: e-mail readable · key or password readable”

Receipt file docs/evidence/export-check-0.4.13.json

Tip The last row is the point. When a checker reports “no problems”, there are two possibilities: there are none, or the checker is not doing anything. If you cannot tell which, it is worthless. So every time I feed it an output with the redaction deliberately removed — it has to catch the e-mail and the key. Only when it does is its “0 problems” on the real output worth believing.

Findings can be marked on the timeline in one click; step through them with Shift+↓.

📱 Vertical highlights

Turn a stretch of a landscape recording into a 9:16 short: software is cropped around the pointer, web pages and documents are kept whole, the title sits above and captions below, and zooms stay inside the picture area.

🤖 AI: bring your own, nothing bundled

The built-in chat: use the Codex subscription you are already signed in to, or a local model; trim, caption and zoom with a sentence, plus one-click recipes

My choice for AI is not to bundle: no built-in model, no API fees collected on your behalf, no deciding for you. If you are already signed in to a Codex subscription, it is reused (no key to paste); for fully local, connect LocalBrain. 69 tools cover editing, sound, captions, transitions, redaction, zoom, markers, per-clip effects, visual templates, vertical reframes, export and the finished-video check — anything you can do by hand in the editor has a matching tool — plus 12 one-click recipes such as “shorten pauses”, “remove filler words”, “remove repeats and retakes” and “check before publishing”. External AI can use the same tools over MCP (73 with reading, saving, undo and redo).

The last time “AI editing” excited me was the one-click finished-video wave: stunning demos that never survived a real publishing routine, because they did everything for you and you could only take it all or start over. Here the AI works on your timeline: every step lands as an edit you can undo, and you can take over at any moment.

🗂 Projects, version history and storage

Version history: every version with its time, cause and changes since the one before — go back to any of them

  • Every edit is written to disk at once, and versions of the whole project are kept over time (every 5 minutes by default, only when something changed), each labelled with its cause and changes; go back to any of them, and undo the going back.
  • Save a project anywhere and open one from anywhere; the folder for new projects can live on an external drive.

Storage & cleanup: where projects, recordings and caches live, with each kind of cache cleared on its own

“Storage & cleanup”, new in 1.0.1, came from my own system disk running out of room:

  • Move the recordings folder and the existing recordings move with it, with every project and saved version pointed at the new place. Everything is copied and checked before an original is deleted. A take that has just stopped needs a few seconds to finish writing, and the app waits for it — it never moves half a recording. Moved files keep their dates.
  • The cache folder can move too: the old caches are simply deleted and rebuilt in the new place as needed. Each kind of cache shows its size and can be cleared on its own or all at once. Recordings and projects are never cleaned up.
  • An old annoyance fixed on the way: after quitting, voice isolation sometimes kept running in the background, holding a whole CPU core until a half-hour timeout, and fought the next launch for resources. Now background work ends when you quit.

☕ 1.0.2: a sponsor code, and where it shows up

The Support the Author dialog: the author's WeChat Pay code and a short note

The app costs nothing and shows no ads. 1.0.2 adds one place to support it: the app menu’s Support the Author, which opens the author’s WeChat Pay code. In a Chinese interface the same code also sits in the export-done panel, right below “Saved to” — the moment you have just got your video, not the moment you are about to start work.

  • In an English interface it never appears on its own; the menu item is the only way to it.
  • In a Chinese interface it waits until the app has been installed for three days (counted from when its data folder was created, so updates do not reset it), then shows after a finished export at most once per calendar day.
  • It never appears while a recording is running: anything that pops up then would end up in your video. After each export the app first asks whether a recording is running and, if it cannot tell, assumes one is.
  • It has no buttons and blocks nothing; sponsoring or not changes no feature. The code is an image inside the app: showing it needs no network, and nothing about it is counted.

⚠️ Current limits

  • No proper signing yet: the Mac version is signed with my own certificate and not Apple-notarized; Windows is not code-signed and SmartScreen will warn on first run; Linux is unsigned too.
  • Platform features differ: Mac, Windows and Linux offer 1.9.0. Voice isolation, automatic sensitive-text detection and text read-back of the finished video remain macOS-only. Clicking a window to select its bounds works on macOS and Windows. Windows and Linux desktop interaction still needs testing on the respective systems.
  • Linux has dependencies: all capture goes through xdg-desktop-portal, and it needs Vulkan drivers and your desktop’s portal backend.

🧭 My take

If you make tutorials, reviews or software demos, the real cost of screen recording was never the ten minutes of recording. It is what comes after: cutting pauses, redacting, reframing for vertical, checking nothing went wrong. Everything in this app is designed around “after recording”, including that slightly clumsy-sounding check — watching the finished video again after export.

The clumsy way has its merits: it trusts none of the intermediate steps I wrote, only the final file. After “all tests green but the e-mail showed”, I think that distrust is worth it. Taking the screenshots for this post, my own app taught me the lesson again: the screenshots revealed three problems — splitting a clip dropped the music after the cut, redaction boxes on a scrolling page were far too big, and silence got transcribed as the marker “[BLANK_AUDIO]” — none caught by tests, all obvious the moment you look at the picture. They were fixed before 1.0.1 went out. Adding the sponsor code in 1.0.2 taught me the same lesson again: all 14 of my real-app checks passed, yet the screenshot showed the strip stuck at the bottom edge of the export dialog — half a line visible, the buttons below the edge, the code opening out of sight. The checks asked whether it was there; the person using it needs to see it.

Summary: The design as described for 1.0.2 HyphenScreen is the screen recorder and editor I use myself. This historical account describes the Mac build at the time: local sensitive-text recognition, redaction through zooms and crops, and a read-back of the exported video. Current downloads, platform differences and verification limits are listed at the start of this article.


📚 Notes

Self-made software

Versions, changes and supported installers all come from the same public release records.

Share:

Comments

Loading comments…