v0.47.0

Added

• start_recording backend: "ddagrab": GPU screen capture (Desktop Duplication) for one display, region, or window, matched to its DXGI output by geometry. With NVENC, frames go straight from capture to the encoder: a 1080p60 recording took 0.5% CPU against 12.7% for gdigrab + x264, with no dropped frames (#199).
• start_recording encoder: "auto" | "x264" | "nvenc" | "qsv" | "amf": hardware H.264 encoding with draft/standard/high mapped to each encoder's constant quality. auto (the default with ddagrab) runs a one-time trial encode and falls back to libx264 on machines without a GPU encoder. doctor lists ddagrab and the encoders that work (#199).
• pause_recording / resume_recording: leave setup (logging in, loading fixtures) out of a recording. Paused stretches are absent from the final file, the event log, and markers; stop_recording joins the segments with a stream copy. Sessions gain a paused status, and shutdown finalizes paused sessions (#200).
• list_video_devices and start_recording { webcam: { device, corner, size, shape } }: records a camera beside the screen into <output>.webcam.mp4, aligned frame for frame with the recording (across pauses too); stop_recording returns a ready overlay call, and overlay gains shape: "circle" (#200).
• cursor (#170): record with start_recording cursor: "events" (the OS cursor left out, the event log on), then draw a cursor from the log, smoothed by a critically damped spring timed to arrive on time, scaled, hidden after hideAfterSec idle, with a ripple at each click and an optional spotlight. Positions drive the overlays through a sendcmd script, so a long recording costs no more per frame than a short one. Also a pipeline step that follows earlier trims, speed changes, cuts, crops, and scales.
• narrate (#175): offline voiceover. Each of lines ({ t | marker, text }) is spoken by the built-in Windows voices (no download or API key), placed at its second or at a marker, and mixed over the original audio, which is ducked under the voice by default. A matching .srt is written beside the output, and lines that run into the next one or past the end are reported in warnings. The text never leaves the machine; doctor lists the installed voices.
• A generated tool reference (#202): npm run docs builds the server, reads its live tool list and schemas, and writes a page per tool under docs/tools/, the README and home-page tool tables, mcp-tools.json, docs/llms.txt and docs/llms-full.txt, and the sitemap. The docs test regenerates them and fails on a stale copy; npm run docs:check checks without writing.
• A Claude Code plugin (#177): /plugin install screencast --marketplace TMHSDigital/screencast-mcp installs the server plus bug-repro, readme-demo, and verify-ui skills. A test keeps every tool and recipe the skills name in step with the server.
• A headless CLI and GitHub Action (#179): screencast-mcp record -- <command> records the screen while a test command runs, exits with its code, and on failure writes a GIF of the last 30 s, a contact sheet, and evidence.json; screencast-mcp call <tool> '<json>' runs any tool. No arguments still starts the MCP server. The composite action in action.yml runs it on windows-latest and uploads the evidence.

Changed

• mcp-tools.json is generated and carries each tool's full input schema as the server reports it; the README's per-tool descriptions are each tool's own description, linked to its reference page (#202).
• The tools/list size budget test is raised to 99 KB for this release's new tools and options (about 97.6 KB by default).