ReaLackey v0.3.0

Added

• Reasoning display: models that expose their chain-of-thought — DeepSeek-R1,
  Qwen3, and Ollama "thinking" models over the OpenAI-compatible endpoint (via
  reasoning_content) — now stream it into a collapsible Reasoning block above
  the answer. It's shown separately from the reply and is not spoken as the final
  answer.
• Anthropic extended thinking, as a per-provider toggle ("Extended thinking
  (reasoning)" in the provider settings dialog, Anthropic only). When on, the
  request asks for adaptive thinking — the model reasons more on hard tasks, less
  on easy ones — and streams that reasoning into the same collapsible Reasoning
  block as above. The thinking blocks are stored and replayed verbatim (with their
  signatures) so multi-step tool-use conversations keep their reasoning continuity;
  toggling thinking back off cleanly drops them from the ongoing conversation.
  Requires a model that supports adaptive thinking (Claude Opus 4.8 / 4.7 / 4.6,
  Sonnet 4.6+); older models reject it.
• Lower per-request token usage (helps free-tier keys, which meter tokens/minute
  and re-charge the whole tool list + prompt every agentic turn):
  • Trimmed the tool descriptions and system prompt (deduped boilerplate now
    stated once) — ~12% off the static overhead, no behaviour change.
  • Media eviction: past screenshots/audio/video-clip frames were re-uploaded
    every later turn; now only the most recent captures stay live (older ones
    become a placeholder). Tunable via RAAI_MEDIA_KEEP (default 2).
  • Prompt caching on the Anthropic path (cache_control on the tools+system
    prefix): repeated turns re-read it as a rate-limit-free cache read.
    RAAI_PROMPT_CACHE=off to disable.
  • Progressive tool disclosure (opt-in RAAI_PROGRESSIVE_TOOLS=on): only a
    core set + a load_tools(query) loader are sent; the model pulls in the rest
    by capability on demand (session-persisted, with keyword pre-loading). Cuts the
    tool payload ~70–90% for the turn — aimed at free tiers like Gemini.
• Video-clip vision (capture_video_clip): instead of a single screenshot of
  REAPER's Video window, the assistant can grab several frames across a time
  range (stepping the edit cursor, playback stopped) and — for audio-capable
  models — the clip's audio, so it can reason about motion, cuts, transitions,
  on-screen text timing, and A/V sync. One consent covers the whole clip. Range
  defaults to the time selection; frame count defaults to 6 (2–12). The
  seek-to-frame settle delay is tunable via RAAI_VIDEO_SETTLE_MS (default 250 ms)
  if a heavy video-FX chain needs longer to re-render.
• Per-provider "Supports audio (listening)" checkbox in the provider settings
  dialog, next to "Supports images". Audio input was previously auto-detected from
  the model id only, so a locally-run multimodal model (e.g. Google Gemma
  3n/4 via Ollama or LM Studio) had no way to enable listening; now you can toggle
  it explicitly. Gemma is also recognized as vision-capable by default. (Whether a
  given local server accepts audio input is up to the server/model.)
• Prompt presets: save reusable prompts and drop them into the chat composer
  instead of retyping. Manage them (add / edit / delete) from Extensions →
  ReaLackey → "Prompt presets…"; in the chat window the Presets button (or
  Alt+P) opens a picker that inserts the chosen prompt into the composer so
  you can tweak it before sending. Stored globally in presets.json, next to
  providers.json.
• Chat message navigation + copy: Alt+1 … Alt+0 (and the key right of 0) jump
  to that message and move focus to it (the screen reader reads it); a quick second
  press of the same combo copies that message — your request or the model's
  response — to the clipboard.
• Action & keyboard-shortcut tools: search_actions (find actions by name),
  get_action_info (an action's name, toggle state, and bound shortcuts),
  run_action (run any action by id or named command — a catch-all for anything
  without a dedicated tool), delete_action_shortcut, and add_action_shortcut
  (opens REAPER's key-assignment dialog, since the API can't bind a key directly).
• Time-selection tools: get_time_selection, set_time_selection (with an
  optional seek to move the edit cursor to the start), and
  clear_time_selection; the time selection is also reported by get_transport.
  The assistant can now mark a range and then analyse, render, or measure it (the
  analysis/render tools default to the time selection).

Fixed

• The assistant no longer finishes silently when the model returns an empty
  response (no answer and no tool call). Previously the status just went to
  "Ready." with nothing shown — which reads as a crash — most visibly with local
  models (Ollama) that don't reliably do tool use. It now says the response was
  empty and hints at likely causes.
• macOS: the release .dylib is now a universal binary (Apple Silicon + Intel).
  It previously shipped arm64-only, so it silently failed to load in REAPER on
  Intel Macs or under Rosetta.
• macOS: the release is now Developer-ID-signed and Apple-notarized (when the
  repo's signing secrets are configured — see docs/macos-notarization.md), so
  it loads without the manual quarantine step. Without the secrets it falls back
  to an ad-hoc signature (users then run xattr -dr com.apple.quarantine).
• macOS: keyboard handling in the chat window. Keys typed in the composer could be
  swallowed by REAPER — arrow keys jumping focus to the arrange, and Cmd+C/V/X not
  reaching the text field (they hit REAPER's Edit menu). When the window is in
  front, its keystrokes are now handed to the webview's native editing
  (copy/paste/cut/select-all, arrow keys, typing) instead of REAPER's global
  shortcuts/menu.