Skip to content

feat(mulmoterminal): Agent Picker and Any Backend clips; JA narration on gemini-2.5-pro-preview-tts - #120

Merged
ystknsh merged 9 commits into
mainfrom
feat/mt-agents-0904
Sep 6, 2026
Merged

feat(mulmoterminal): Agent Picker and Any Backend clips; JA narration on gemini-2.5-pro-preview-tts#120
ystknsh merged 9 commits into
mainfrom
feat/mt-agents-0904

Conversation

@ystknsh

@ystknsh ystknsh commented Sep 6, 2026

Copy link
Copy Markdown
Contributor

Summary

MulmoTerminal の機能クリップ mt-agent-picker(台帳 #54)と mt-any-backend(台帳 #55 を英日で追加し、EN を投稿済み・予約済みの 9 本の JA ナレーションを gemini-2.5-pro-preview-tts に切り替えた。以後の規約(en / ja とも Pro)を CLAUDE.md に明記。

  • mt-agent-picker: ランチャーのエージェント行で Codex → acme-api チップ → ▶、+ → Antigravity → acme-mobile チップ → ▶、3 セル並走(俯瞰)。クリックはすべて「動くポインタ + 直前の枠パルス」
  • mt-any-backend: Model ピッカーで OpenRouter のプリセット(Kimi K2.7 Code · 3/3)→ ▶、設定ビート(ランチャーの help がアプリ自身の文面で providers の JSON を出す)、バナーの moonshotai/kimi-k2.7-code にアンダーライン → 質問 → 返事が返って緑
  • JA を Pro に: restart / right-click / shortcut-list / launch-form / quit / chat-agent-picker / status-colors / cockpit-roster / tmux-persist。Pro は約 1 割ゆっくり読むので、4 本の素材(restart clip2 / right-click / launch-form / quit)は最終フレームを 1.5〜2.0 秒伸ばした。EN の素材は共有だが伸ばした尾は使わない
  • 制作メモ: mt-agents-NOTES.md(新設)、mt-fresh-clips-NOTES.md(Pro 切り替えの節)

Items to Confirm / Review

  • Antigravity のバナーに出るアカウントのメールは、アドレスの文字列だけ boxblur で伏せてある(起動ビートと grid ビートの両方)。「(Google AI Pro)」の部分は残る
  • Codex のバナー directory: ~/mulmoclaude/github/…/mt-demo/acme-api と Claude Code のバナー ~/mulmoclaude/github/mt-demo-home/mt-demo/acme-web は、これまでの投稿済みクリップと同じ扱い(許容)
  • OpenRouter のセルに Claude Code 自身の「claude.ai connectors are disabled …」が 1 行出る。providers 機構に固有で消せないので、そのまま
  • mt-tmux-persist の dev サーバのログに code 404, message File not found(http.server の通常のログ・白字)が写る。作者判断で放置
  • 未投稿の EN(plain-shell 以降)とその JA は触っていない。レビュー後の修正時に両方 Pro へ

User Prompt

  • 台帳 feat: add weather-showcase demo deck #54 / feat: connected-collections and record-buttons feature demos #55 の 2 本を撮って作る。Antigravity のメールはアドレスだけマスク
  • レビュー: ビート 1 で Codex と言う/Any Backend の要点は「OpenRouter でいろいろなモデル」/17〜20 秒で Codex に寄らない/Kimi にアンダーライン(19 秒)/repo と + のクリックに矢印/フォルダパスが突然変わらないように(チップのクリックで撮り直し)/Any Backend の欄の矢印はいらない/ビート 4「返事は同じセルに…」を外す
  • TTS のモデルを確認 → 最近 X で投稿を始めた分(EN 投稿済み)の JA を Pro に変更して作り直す。次回から en / ja とも Pro

検証

  • 尺: render-two-pass.shcheck-beat-fit を全デッキで通過(余裕 0.5 秒以上)
  • 出荷判定: 別セッションの sub-agent が 0.5 秒刻み(ビート末尾は全フレーム)で、メール・実パス・リハーサルの写り込み・エラー表示・ビート末尾・ナレーションと絵の対応を確認。最終版は EN 2 本 + JA 11 本すべて PASS
  • 素材と rig: mt-demo-home/footage/2026-09-04/(agents-a〜g、切り出しの正本 cut-mt-agents.sh)、footage/2026-09-06/extend-tails-ja-pro.sh、rig は claude-config の skills/mulmoterminal-video/record-agents.mjs

関連

  • 台帳・X の記帳: receptron/mulmoterminal-marketing#39

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi

ystknsh and others added 9 commits September 5, 2026 08:38
…54 / #55)

Two feature clips, EN and JA each, from the 2026-09-04 capture (rig record-agents.mjs, footage agents-b/agents-c, cut recipe cut-mt-agents.sh in the demo home).

- mt-agent-picker: the launcher's agent row picks Codex, then Antigravity, then three cells (Claude Code / Codex / Antigravity) run side by side. Narration says "five agents, three shown here" because only those three are installed on the capture machine. The account e-mail in Antigravity's banner is blurred (string only) in both the launch beat and the grid beat; the launch framing stops above the launcher's resume list so an earlier take's prompt is not in frame.
- mt-any-backend: the Model picker chooses an OpenRouter preset (Kimi K2.7 Code · 3/3), a config beat from the launcher's own setup help (fixed-inset modal, so a still), launch with the model id on Claude Code's banner, a question typed on camera, the reply. Claude sessions only. Claude Code's "connectors are disabled" notice is left as is.

Durations are the pass-1 audio measurements + 0.9 (render-two-pass); every embedded clip clears check-beat-fit. Ship-checked by a separate reviewer on frames at 0.5 s / 0.25 s.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…at 1 says what it does, not "plan limit"

Author review 2026-09-05. Agent Picker beat 1 now says which CLI is being started ("this one starts Codex, one of five agents") and the "three of the five agents" caveat moves to beat 3. Any Backend beat 1 was an invented "why" (hitting a plan limit); the point is many models through OpenRouter, picked per session, so it is now a "what" sentence, and beat 4 says "the model you chose". The longer beat-1 narration needed a longer pick-model clip (9.4 s, banner held). Pictures unchanged; re-checked by a separate reviewer.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
… on the Codex cell

Author review 2026-09-05: the beat says "three of the five agents, one grid", so the picture stays the three-cell overview for the whole 10.5 s. The 1.8x push-in that read the Codex header's model name worked against that. The e-mail line in the Antigravity cell is still blurred for the seconds it is on screen.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
Author review 2026-09-05: the native select's open list is not captured by the screencast, so an amber underline under "Kimi K2.7 Code · 3/3 · 14s · 262k" points at the choice from the moment it is picked until the launch click. Picture otherwise unchanged.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…s model id

Author review 2026-09-05: the directory field and the toolbar "+" (New terminal) had no pointer, only the agent buttons and the play button did; every click now gets the moving pointer and the amber pulse (Antigravity clip starts 1.3 s earlier to include the "+"). The underline moves from the Model row in beat 1 to the Claude Code banner line `moonshotai/kimi-k2.7-code` in beat 3. Ship-checked on 0.5 s frames with the annotations enlarged.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…eply lands in beat 3, beat 4 dropped

Author review 2026-09-05. The launch take is reshot (agents-e); the directory field already holds acme-web, so nothing points at it. The fourth beat (a close-up of the reply) is removed: the point is that the same Claude Code runs on the chosen model, not what it answered. Beat 3 is ramped harder so the reply and the green done state land inside the narration; the clip is ~27 s. Ship-checked on 0.5 s frames.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…d, clean beat ends

Author review 2026-09-05. The launch takes are reshot so the working directory is filled by clicking its repo chip (pointer + amber pulse on the Codex button, the acme-api chip and the play button; on the toolbar "+", the Antigravity button, the acme-mobile chip and the play button), the Antigravity take holds until its banner is up (e-mail line blurred), and the grid take is reshot with a longer Claude task so no next-step suggestion sits in the input. The grid clip starts 1.2 s in so a two-frame terminal redraw does not fall on a beat end. Ship-checked on 0.5 s frames and frame by frame at the beat ends.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…rule goes into CLAUDE.md

Decided 2026-09-06: every speaker sets model gemini-2.5-pro-preview-tts, in both en and _ja decks (without it mulmocast falls back to the Flash TTS, which every deck so far used). Applied to the two new JA decks; Pro reads about a tenth slower, so the grid clip is held 0.6 s longer to keep the beat-fit margin.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
…2.5-pro-preview-tts

The nine clips whose EN is posted or reserved (restart, right-click, shortcut-list, launch-form, quit, chat-agent-picker, status-colors, cockpit-roster, tmux-persist) get `model: gemini-2.5-pro-preview-tts` on their JA speaker and are re-rendered. Pro reads about a tenth slower, so four assets (restart clip2, right-click, launch-form, quit) are extended by freezing their last frame 1.5–2.0 s (footage/2026-09-06/extend-tails-ja-pro.sh); the EN decks share those assets but never reach the added tail. All eleven JA clips (these nine plus the two new ones) checked frame by frame at the beat ends.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CKF9DD7JZ7jeHYDeuniHJi
@ystknsh ystknsh changed the title Feat/mt agents 0904 feat(mulmoterminal): Agent Picker and Any Backend clips; JA narration on gemini-2.5-pro-preview-tts Sep 6, 2026
@ystknsh
ystknsh merged commit ed7e83f into main Sep 6, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant