Changelog

1.7.4 — 2026-10-20

• A Capability Storyboard can supply the talk track and scene order; its mockups are replaced by real capture or labeled.

1.7.3 — 2026-10-20

• Voice narration: add the VoiceStudio desktop app as an unrated local trial candidate, with its AGPL-3.0 app license separated from model terms and the permission rule for cloning.

1.7.2 — 2026-10-18

• Route conceptual motion explainers to a motion-design lead; retain real capture, narration and delivery ownership for mixed films and ordinary walkthroughs.
• Name official HyperFrames creative and animation skills for composition and choreography under the existing optional route. Require a representative moving proof for substantial authored inserts, preserving reference direction and teaching pace rather than importing promo defaults.
• Link the public-source decision review; no renderer, paid service or installation is added by this guidance update.

1.7.1 — 2026-10-18

• HyperFrames route: when a requested mix includes SFX or music, earshot (MIT, CPU-only) can screen sound-to-cut offsets against the storyboard's cues and render code-built effects with measured sync points. Listening still accepts the mix; narration review, loudness measurement and the full listening reviewer are unchanged. Its optional cloud critique uploads audio and stays off without consent.
• Record the decision in the October 18 earshot review.

1.7.0 — 2026-10-11

• Gemini adapter supports the GA Gemini 3.8 Flash TTS / Flash-Lite TTS schema: verbatim transcript in text, instructions in speech_metadata.style, speechConfig.voiceConfig.voice (prebuilt, library or voice_... IDs), and a validated RIFF WAV response saved without re-wrapping. gemini-3.1-flash-tts-preview keeps the legacy path.
• ElevenLabs: document eleven_v4 as the current flagship and reject voice settings v4 does not accept (only stability and similarity_boost).
• Provider guidance: dated Oct 10 blind-preference snapshot as audition candidates, plus re-audition required when an accepted voice name moves to a new model.
• HyperFrames: install the official skills by name (hyperframes, hyperframes-core, hyperframes-keyframes, hyperframes-cli) under this route; note upstream v0.8.149 while keeping the proven 0.8.137 pin.
• Sources: October landscape check (video-shotcraft, onetake, agentic-product-demo, product-demo-director, official HyperFrames/Remotion/ElevenLabs skills).

1.6.0 — 2026-10-06

• Offer HyperFrames as an explicit choice for each new video, carrying that choice through revisions; keep ordinary capture/composition and paid-media authorization separate.
• Add a pinned HTML/GSAP project builder with measured footage, an audio-first storyboard, camera paths, speech-linked highlights, phrase captions, optional music/SFX and music ducking. Author portrait layouts separately and preserve portable source assets.
• Add optional ElevenLabs character timestamps and validated word sidecars; existing basic and polished assemblers consume measured captions while retaining their labeled fallback and explicit-cue precedence.
• Add full-decode, media-spec, black/freeze and loudness checks plus transition review frames. Record synthetic landscape/portrait render evidence separately from listening and viewer acceptance.

1.5.6 — 2026-09-20

• Add a conditional application-narration route covering whole-operation budgets, uncertain charges, reusable clip identity, transcript review and manifest delivery.
• Distinguish cached playback from local inference, and technical/transcript checks from heard voice quality. Preserve the ordinary single-voiceover route.
• State that this is destination-pipeline guidance; the bundled helper does not gain budget enforcement, caching, transcription or a runtime player in this release.

1.5.5 — 2026-09-20

• Add a lightweight route for quick screenshots or recordings, with capture access checked before promising video and a truthful fallback when either format is accepted.
• Verify fresh local-page loading and detected screenshot formats; preserve unsaved state, output dimensions and existing host authorization boundaries.
• Record empty-window-list evidence without inferring its cause; add manual policy cases and align the previously stale Claude marketplace version.

1.5.4 — 2026-09-14

• Add a route for capturing an installed desktop application on the user's own machine, with the executed mechanics: accessibility clicks report success without actuating embedded web content or moving the OS cursor, real HID events do both, and a compilable pointer program ships as an asset.
• Record the recorder contract: screencapture -v writes only when it ends on its own, so pass -V seconds; two takes were destroyed by signalling it. A capture script must be the tracked process, not a detached child.
• Require every pointer move to sit on a phrase that names what it points at: schedule the path from silence-detected narration timings against a monotonic clock, and treat stillness as a choice. A move with no verbal cue reads as nervous and pulls attention to the pointer.
• Do not pan or zoom a captured still. zoompan resolves its crop to whole pixels, so a gentle ramp pumps: measured 0.2362 mean frame-to-frame luma change against 0.0007 for the same shot held static.
• Record that activate is not a window selector: with two windows of one browser on different profiles, the app fronted the signed-out one and its page carried a notice the accepted footage did not. Verify the expected state in a still before recording.
• Add programmatic take review: diff a downscaled crop of every frame against a confirmed reference frame to locate foreign windows, and scan the assembled cut for unintended jumps. A two-and-a-half-second intrusion that eye review missed was found this way.
• Record the operator as the main hazard to a take, with the mitigations: activate the target immediately before the recorder starts, state the hands-off window, and clear permission prompts beforehand.
• Document fitting the picture to accepted narration with a setpts ratio, the roughly 1.3x slowdown ceiling for cursor-only motion, and the resulting capture headroom.
• Record Algenib as an accepted single-narrator direct-Gemini profile, with the pace and the constant delivery instruction that let re-synthesised lines splice against earlier takes.
• State plainly that every prebuilt voice on the model is selectable: the voice names in this package are dated listening results, not a supported subset, and none of them is a default.

1.5.3 — 2026-09-13

• Record Algieba + Sulafat as an accepted direct-Gemini, male-presenting two-host profile and Kore as an accepted female-presenting Southern-style product-demo voice through OpenRouter.
• Separate Gemini voice availability from provider-route capabilities: shared voice names do not prove matching prompt, multi-speaker, format, quota or catalog behavior.
• Add the observed long-form limitation: split two-speaker production at natural one-to-three-minute exchanges and verify speaker continuity.

1.5.2 — 2026-09-06

• Offer native project + source assets + MP4 before capture; reuse an established editable-project preference.
• Rank Cap Studio first for documented editing breadth and Recordly second, with an editing matrix directly in the README.
• Require native zoom/audio/layer preservation, packaged-project reopening and edit checks; distinguish this target from the limited existing trials.

1.5.1 — 2026-09-06

• Put the preference, quality evidence and pros/cons matrices directly in the README, including Mac narration status.
• Correct the Cap trial record: later frames showed an occluding Edge window, missed by earlier review.
• Document the fresh foreground-controlled take, matched narration in both native examples, and stronger content/audio verification.
• Make the Cap recommendation conditional on foreground control and keep framework showcases separate from the product-demo page.

1.5.0 — 2026-09-06

• Add a separate voice-narration skill for provider selection, auditions, audio-only output and revoicing, sharing the existing audio helper.
• Add MiniMax synchronous speech support with explicit model/voice, validated hex MP3 responses, prior paid-call authorization and no automatic retries.
• Distinguish MiniMax Desktop Design, coding, generated assets and speech API routes; retain real capture evidence and untested quality labels.
• Move provider guidance under the narration skill, preserve existing links, and document complementary OpenAI speech and natural-writing skills.

1.4.2 — 2026-09-06

• Add the preferred Mac workflow, qualitative fit/quality assessment, pros/cons and explicit trial boundaries.
• Separate established Kore narration from unintegrated, unauditioned Chatterbox/Kokoro candidates; exclude the NVIDIA route from the Mac default.
• Document shorter cursor/camera timing and the remaining viewer-acceptance and combined-pipeline checks.

1.4.1 — 2026-09-06

• Require a visible standard pointer for mouse-driven demos unless cursorless output is requested.
• Record the tested browser-pointer overlay repair for Cap/Recordly and distinguish baked footage from editable native cursor layers.

1.4.0 — 2026-09-06

• Record actual HyperFrames and Remotion composition trials, including working configurations, timed-out attempts, camera bounds and reuse limits.
• Document executed Cap CLI capture/config/export and Recordly bundled-engine recording, keeping supported interfaces distinct from experimental internals.
• Add automatic-work routing across browser navigation, native capture, composition, and refinement of existing videos; preserve editable sources and a normal cursor option.
• Require recorded-pixel checks for window occlusion, persistent native menus and real pointer telemetry; successful page assertions or a valid capture project alone are insufficient.

1.3.0 — 2026-09-05

• Separate cursor rendering from 25fps browser capture and animate cursor/camera at 60fps; fade idle pointers and retain real click timestamps.
• Reduce repeated zoom resets and competing motion in the existing-site example; add a conservative target-bounds camera fallback and loudness normalization.
• Assess screen-editor/presenter effects, automatic versus conditional defaults, local tools, local TTS, and HeyGen HyperFrames as distinct from hosted avatars.
• Preserve the source-cadence limitation and actual hover-at-arrival behavior; do not claim native 60fps app animation or product parity.

1.2.0 — 2026-09-05

• Add real 2× capture with eased cursor movement, actual click pulses and timestamped targets, plus a Pillow/FFmpeg compositor for explicit camera zooms, framing and captions.
• Add a reproducible demonstration using the existing Design Router Field Guide and user-auditioned Gemini Kore narration.
• Add OpenRouter speech support, including the live-confirmed Gemini PCM requirement, secure shared-credential discovery, partial-take voice metadata and response-format validation.
• Require playable examples for quality claims and keep manual camera planning/scene-level captions distinct from automatic editing and word alignment.

1.1.0 — 2026-09-05

• Add Google Gemini TTS/Kore support, natural-voice audition gates, silent technical tests and explicit-only local test speech.
• Add dated benchmark evidence and polished production routing for Clueso, Recordly and Remotion; distinguish basic capture from demonstrated visual quality.

1.0.0 — 2026-09-05

• Real UI capture by beat, screenshots, configurable voice quality/provider/model, secure environment credentials, captions and silent revoice output.
• Measured media timing, non-destructive takes, explicit supplied captions, and synthetic fixture for reproducible smoke testing.