Changelog

1.0.0 (2026-10-22)

• BREAKING: Require Agent Plugins 1.0 with a root plugin.json. Marketplace metadata now lives only in marketplace.json; legacy manifests are removed.

0.2.0 (2026-10-19)

From cold runs, where a fresh agent made each video from the skill alone (#342051, #145202, TPI #342443 signed in, and five Explorer and Timeline bugs), then a review for size and a cold run of the smaller skill (#145202: 29 min, down from 39).

• Smaller, same features: the skill text is 4 files under 8 KB each (SKILL.md, video.md, scenario.md, slides.md; 43 KB → 24 KB), each rule said once; each example is under 8 KB (60 KB → 22 KB); voice_engine.py keeps only what the kit calls (52 KB → 11 KB); firstbuild.py is now item.py --commit <sha>.
• New in the docs: scope when an issue has no steps or its PR tests more, acceptance criteria, and when a video is partial; signed-in items on a small request budget; the Agents window, editors, dialogs, git history and the integrated browser; zoom timing and size, and fixing a zoom without recording again; recap.label and the verdict's word budget.
• New setup.mjs --selftest (a 45 s video proving every stage on a new Mac) and item.py <n> (an item's PRs, whether the installed Insiders has them, and the title's phrase).
• Runs live in one short folder under ~/.cache/test-videos/t/ (TEST_VIDEOS_TMP), VS Code's temp files included; each run prints it.
• Runner: dialogs are drawn in the window; no Chat side bar when signed out; new workspace.commits, keybindings and ctx.setSettings().
• Helpers: h.paste(); h.click({ button, modifiers }) for right-clicks and ⌘-clicks; h.command() runs only the row labeled exactly as asked; h.pickRow(), h.quickInput(); h.pickModel(), h.modelLabel(), h.chatRunning(), h.cancelChat(). Key badges put modifiers in the Mac's order (⇧⌘P). The drawn pointer follows only the kit's own moves.
• check.py checks sync against where speech starts, up to 0.6 s needs no action, and every line that does starts with CHECK. compose.py writes out/looks.json for flags, refuses a failed recording, renders ↵ and keeps punctuation after a code span. flags.py warns about pauses over 8 s; with no flags, cards.py uses out/ as it is. A new recording deletes a stale run/timeline.recorded.json. cards.py --preview estimates the verdict's length. The voice check accepts ordinals transcribed as digits.
• Both examples pass every step on Insiders 341e0fa934; the single-issue one serves its own localhost page.
• setup.mjs prints FFMPEG= and re-runs take 2 s; no __pycache__ in the plugin folder.

0.1.0 (2026-10-19)

• generate-test-video: a short narrated "how to test" or "how to verify" video for a VS Code test plan item (TPI), a fixed bug or a group of fixed bugs of one area. It's recorded in real VS Code Insiders in a throwaway profile, and every step is checked in the app. Flags pause the video where something is wrong; a verdict slide opens it (verified, issues, changed, partial or failed, with the build and what isn't covered) and a recap ends it.
• The kit from the 1.143.0 endgame, made portable: no paths to one machine, its own Python environment and playwright-core in ~/.cache/test-videos/, the Kokoro voice offline or MAI-Voice-2 with an Azure key, and videos delivered to ~/Downloads/test-videos/. macOS only.