Skip to content

v4.5.0 — Native media, diagnosis, and safer tools

Choose a tag to compare

View all tags
@github-actions github-actions released this 17 Oct 13:02
· 75 commits to main since this release
v4.5.0
5ae7941

Mods v4.5.0

Native media, symptom-first diagnosis, and target-based approvals. This release includes everything since v4.1.0.

Highlights

Native image, audio, and video input

  • Read images, WAV/MP3 audio, and H.264/AAC-LC MP4 video directly with --image, --audio, and --video — no OCR, ASR, ffmpeg, or nested process.
  • Discover and read media through fs_find_images / fs_read_images and fs_find_media / fs_read_media.
  • Media capability checks follow the provider protocol instead of hardcoded model-name lists, and unsupported media is rejected explicitly rather than silently dropped.

Diagnosis that investigates before changing

  • A new symptom-first diagnosis policy and the built-in diagnose-not-reproducing skill guide investigation of stale caches, hot reloads, copied installs, and wrong components.
  • Added a scoring command that evaluates saved session transcripts against a diagnosis rubric.

Review and approvals

  • Authorization is target-based: reads run without review, while non-temporary writes are approved by local directory subtree or normalized remote origin.
  • In auto mode, Mods infers the session's write targets from the request (with one optional batch of read-only discovery), so routine tasks need fewer prompts.
  • Command reviewability is advisory: at most two simplification nudges, then ordinary approval. Pre-written script files are exempt and run exactly as written.
  • request_user_input supports up to 10 options, longer labels, and multi-line questions.

Terminal experience

  • Unified TUI styling and activity indicators, with the todo plan shown in a fixed sidebar/panel.
  • Rewritten native configuration wizard, improved session browser, light-background themes, and optional Nerd Font glyphs.
  • Git HTTPS authentication can prompt in the terminal and continue the same command without exposing secrets to the model.

Providers and reliability

  • Added OpenCode Go with session headers; DeepSeek v4-pro now uses the Responses API on the official endpoint; Copilot setup discovery and styling were fixed.
  • Preserved completed tool rounds and session transcripts across retries, ignored SSE keep-alives, forced UTF-8 output in PowerShell hosts, and patched vulnerable dependencies.

Breaking changes and removals

  • Plan mode and the --plan / -p flags are removed, along with the prompts.plan override.
  • Web search is opt-in; Tavily is the default provider and DuckDuckGo support is removed.
  • Removed the max-input-chars, max-tokens, no-limit, and builtin-tools.shell-max-output settings and their truncation behavior.
  • Removed prompt-intent and write-scope authorization settings and classifier prompts; target inference replaces them.
  • Windows now offers only powershell_run for shell work; shell_run is no longer registered there.

Full changelog: v4.1.0...v4.5.0

Assets 9
Loading