TODO

TODO.md

@@ -105,3 +105,15 @@ def cmd_usage(client, bid, model):
 ```
 
 **GUI tab (later):** same operations as buttons on a saved local history table — list, cancel, usage, redownload, resume consume.
+
+## Local translation cache (game retranslate)
+
+- Persist every successful translation locally so a full retranslate from scratch can reuse prior results instead of paying again.
+- **Cache key:** hash the translation **payload** (source Japanese text sent to the model) plus target language — same approach as the existing global cache in `util/translation.py` (`get_cache_key`, `log/translation_cache.json`).
+  Payload keys naturally dedupe identical lines across files and survive re-extract / re-import as long as the source string is unchanged.
+- **Scope:** consider a per-game cache file under the game project (e.g. `<game_root>/translation_cache.json` or alongside `wolf_json/`) in addition to (or merged with) the tool-wide `log/translation_cache.json`, so caches travel with the game folder and are easy to back up.
+- **Write path:** on every successful translate (live and batch consume), store `{cache_key: translation}` before or as results land in `translated/`.
+- **Read path:** before calling the API, check cache; on hit, skip the request and apply the cached translation (respecting file-specific re-expansion where the payload is Japanese-only).
+- **Retranslate workflow:** when the user re-imports fresh JSON and runs Translate again, cache hits should fill `translated/` automatically for unchanged source lines; only new or edited source text should incur API cost.
+- **UI (later):** cache stats (hits / misses / estimated savings), export/import, clear cache for one game, optional “prefer cache” toggle.
+- **Invalidation:** optional metadata (model id, prompt version) in the key or entry if we need to avoid reusing stale translations after prompt or model changes.