Two agents, one reader — Tsumugu's study layer lands

Yesterday Tsumugu grew a voice. Today the reader grew into an instrument you can study with — and it was built by two agents working the same repo at the same time.

The reader became an instrument

The synced reader became something you drive:

• Click a sentence and it plays — just that sentence, stopping at its end instead of running on. ↑/↓ (and ,/.) step line to line and auto-play; Space replays the line under the cursor; the ▶ button still plays the whole video through.
• An audio-source toggle (🎙️). The same gestures play either the original video audio or Serena’s clean TTS take, with the video parked on the sentence’s frame. One source at a time — a bug where both played at once traced to YouTube’s async seekTo resuming a paused player.
• An A/B loop strip (🆎) for looping a slice of the video. A true waveform of YouTube audio is impossible (the embed hands us no audio samples), so the strip is a sentence-snapping timeline: drag the ends and they lock to cue boundaries, so even a long video stays select-by-sentence without needing zoom.
• Hoverable Chinese summaries — the “now talking about…” line is now segmented into the same hover-to-grade words as the body text.
• Quiet by default. Hover stopped popping a card on every word; it’s Shift-hover now, Migaku-style. The catch was a persisted setting silently overriding the new default — a one-time migration cleared it.

Two agents, one repo

The part worth remembering: a second agent was building per-speaker dual-voice (assign different voices to different speakers in a dialogue) in the same files at the same time. It became a small lesson in agent coordination:

• We split lanes — one owned the reader/transcript core, the other took CSS, index.html, scripts/, and the libs it authored — so commits never collided.
• A scroll-back bug fix landed in shared territory. Rather than entangle two uncommitted trees, we left it live but uncommitted and let the other agent’s commit carry it. It survived intact, verified after the fact.
• The “player keeps refreshing back to the top” mystery turned out to be the dev server full-reloading on every file save while two agents typed at once — two hands on one keyboard.

Sweating the details

• 我 ran away. A bare single word would make the TTS hallucinate a six-second clip; the fix re-rolls any take longer than a length-based cap. 32 hallucinated word clips re-rendered clean.
• “(segmentation artifact)” was jargon. About a fifth of one transcript’s glossary entries were mis-segmented non-words tagged with that phrase. Reworded to the plain meaning plus a Chinese note explaining the bad split, and the generation prompt now knows to stop emitting the jargon.

Made it pretty — and truly offline

• A brand lockup: the 紡 glyph as the logo and favicon, set in the brush reading face with the badge’s blue→purple→pink gradient. The abstract swoosh is parked as a future loading spinner.
• A polish pass over buttons, the scrubber, print, dark mode, and a mobile layout.
• The one that matters: the reading font is self-hosted now. LXGW WenKai was the last thing the reader fetched from a CDN, so an offline reader fell back to a system serif. Vendored under OFL (~4.4 MB, lazy per unicode-range), the reading face works with zero network. Tsumugu is finally, fully zero-cost and offline end to end.

Now / next

• Karaoke word-sync scoped, not built: feasible via forced alignment, but it needs per-word timing data and a video-vs-Serena timing split. Tiered plan parked — Serena-aligned first (clean audio, reliable), the native video later.
• Deferred while the parallel build settles: a loading spinner (the swoosh earns its keep) and Read↔Review transitions.
• Open: confirm the dual-voice work as it lands; decide the karaoke tier; Vietnamese still needs its own TTS pass — Qwen3 doesn’t speak it.