开发者用 Claude Code mods API 做了个德语闪卡插件，把智能体等待时间变成复习时间

I'm learning German, and I spend a lot of my day watching Claude Code work. So I built a small companion plugin that uses that wait.

**What it does**

- When a turn runs longer than 12 seconds and one of my German mistakes is due for review, a card opens above the prompt with a new situation, e.g. "Your train is late. Text your friend that you are waiting for her at the station."

- I answer in the card, or with `/nochmal Ich warte am Bahnhof auf dich.` while Claude keeps working (the slash command is registered as `immediate`).

- The answer is checked and saved as a review, and the back of the card shows my first mistake next to today's sentence: `Am 07.06.: „Ich warte dich.“ · Heute: „Ich warte am Bahnhof auf dich.“ ✓`

- The status line shows what is due: `FehlerDNA · fällig: 3 Muster, 5 Wörter · 17 Tage in Folge`

- `/fehlerdna` opens a pane with every mistake pattern on its 1-3-7-14-30-60-day review ladder; choose one to see its history.

- During a roleplay, a scene pane shows role, goal, turns and time, with a Hilfe (help) button.

**Who it's for:** people learning German who already use Claude Code. The tutor itself is a skill that also runs in Codex; the companion needs Claude Code's mods API.

**Cost:** free and MIT. No extra API key: writing a card and checking an answer are one Sonnet call each through your own Claude Code login. The status line and the panes make no model calls.

**My relation:** I wrote it.

**What I learned building on the mods API (early access; tested on 2.1.286)**

1. The mod never edits the tutor's memory. Every card is saved through the tutor's local Python engine, which refuses early reviews, reused prompts, copied answers, and a pass after a hint.

2. A `tool.call` hook can keep the agent honest: after you press Hilfe, the mod adds `--support hint` to the tutor's next `roleplay-turn` call if the tutor left it out, so help is never counted as unaided.

3. My first card generator wrote a hotel-room situation for the pattern "Pizza ist feminin", because I had told it to avoid the earlier mistake's words. The grader then saved a fail for an answer that never contained "Pizza". Now situations keep the pattern's own word, an answer that skips the target saves nothing, and the engine's `undo` took the bad record out.

4. Raw hex colours were unreadable in the desktop app's light theme; theme keys (`suggestion`, `success`, `error`, …) follow light and dark themes.

5. `claude plugin test` runs hook tests with stand-ins for the engine, the CLI and the models: 15 tests draw the band and both panes on the terminal and desktop surfaces without Python or a model call.

**Install**

claude plugin marketplace add ahmtsahin/deutsch-loop

claude plugin install deutsch-loop@deutsch-loop

claude plugin install nochmal@deutsch-loop

Repo: https://github.com/ahmtsahin/deutsch-loop

Feedback I'd value most: does a card during long turns feel useful or distracting? What would you put on it besides German?