Skip to content

Releases: highercomve/GhostPen

v0.3.4

Choose a tag to compare

@github-actions github-actions released this 16 Oct 13:49

GhostPen 0.3.4 improves resource cleanup and reduces playback and dictation animation work.

Fixed

  • Reject malformed WAV chunk sizes, unsupported channel counts and invalid channel selections without overflow or stalled reads.
  • Clean up partially allocated meeting jobs and roll back failed worker starts. Interrupted saves show a retryable error; Transcribe again also completes track merging. Questions and reprocessing report queue failures.
  • Release the local model helper's saved-plan parsing memory, threadpool, control-text list, token cache and command-line strings.
  • Release transcription server bind addresses on startup failure and stop audio monitor selection after the first match.

Changed

  • Meeting playback updates the interface when its displayed position, clock, line or word changes, and keeps a stable keyboard listener.
  • The dictation waveform stops requesting frames once flat and resumes when the phase changes.

Validation: 73 Zig unit tests passed; backend compilation and the frontend production build passed. Running-app visual verification was not performed.

v0.3.3

Choose a tag to compare

@github-actions github-actions released this 16 Oct 03:46

GhostPen 0.3.3

  • Meeting question suggestions now cover action items, decisions, key takeaways, unresolved questions, and follow-up messages, without speaker-specific examples.
  • Deleting a saved meeting answer now removes it from the meeting.
  • The local installer opens Settings after restarting, accepts window arguments such as --meetings, and verifies the launched process stays running.

Validation: frontend production build, Zig tests, and local CUDA/Vulkan AppImage installation and restart.

v0.3.2

Choose a tag to compare

@github-actions github-actions released this 16 Oct 01:25

GhostPen 0.3.2: meeting summaries and questions

v0.3.1

Choose a tag to compare

@github-actions github-actions released this 15 Oct 20:30

GhostPen 0.3.1: meetings, with who said what

v0.2.24

Choose a tag to compare

@github-actions github-actions released this 13 Oct 12:04

GhostPen 0.2.24

v0.2.23

Choose a tag to compare

@github-actions github-actions released this 11 Oct 21:27

GhostPen 0.2.23

v0.2.22

Choose a tag to compare

@github-actions github-actions released this 10 Oct 01:04

GhostPen 0.2.22

v0.2.21

Choose a tag to compare

@github-actions github-actions released this 06 Oct 15:16

Summary, made for reading

  • A redesigned summary reader with warm light and dark themes, clearer typography, and more space for the article.
  • Adjustable font size (14–24px), serif or sans-serif text, and relaxed or compact line spacing. Reading preferences are remembered across sessions.
  • The link form collapses once the summary is ready. Improved empty, loading, and error states, plus a Stop button while summarizing.
  • Drag the menu, Settings, Playground, and Summary headers to move their windows.
  • Wayland overlays release keyboard focus when you click another app, so you can keep working with GhostPen open.
  • Restored Windows package uploads and the in-app updater payload in CI.

v0.2.20

Choose a tag to compare

@github-actions github-actions released this 06 Oct 00:50

GhostPen v0.2.20

  • The link summarizer: paste a URL in the summary window and GhostPen reads the page and streams the summary as Markdown with the active AI profile.

  • Three summary depths: Brief (a TL;DR paragraph plus a few key points), Standard, and Detailed (fuller sections, more quotes, the document's numbers kept).

  • The model service's /metrics: requests, tokens and speed of the built-in model's prompt and generation, the last request and the one running now. Chat completions honor stream_options.include_usage.

  • The built-in runner remembers the split that loaded and reuses it across restarts.

  • Built-in models: a successful load no longer crashes the runner while saving its remembered plan (flash attention "auto" is the enum's -1, which @intcast to an unsigned type refuses). Since the plans feature, every GPU load loaded the model and then died as "could not be loaded".

  • Settings reads the built-in models' trained context from the GGUF header directly, without loading a model.

  • The summary's text column is centered in the window; the menu window is transparent.

  • Oriel v0.9.2.

v0.2.17

Choose a tag to compare

@github-actions github-actions released this 29 Sep 22:35

Full Changelog: v0.2.16...v0.2.17