發布日期：2026-05-29
現在你點擊產出 HTML 分析報告時，速度會快上很多。以前系統很笨，每次產報告都會把你的所有歷史紀錄檔從頭到尾重新讀一遍；現在它學會了「快取（cache，把讀過的資料暫存在記憶體裡）」，只要檔案沒被改過，它就會直接拿上次記住的結果來用。同時，我們對系統底層進行了大規模的「型別檢查（type-checking，確保每個變數都裝對東西的健康檢查）」，原本有 35% 的程式碼是檢查死角，現在全面覆蓋後，順手抓出了好幾個隱藏的變數錯用和重複設定問題。我們還加上了多國語系的自動防呆測試，以後只要有一國語言漏翻譯，系統在出廠前就會自動擋下，不會再發生切成日文卻默默跑出英文的怪事。就像圖書館員以前每次有人問庫存，都要把整座圖書館的書重新數一遍，現在他終於學會看庫存登記簿了；而且圖書館還趁機做了一次全棟安檢，把以前偷懶沒巡到的黑暗死角全掃了一遍，連標示牌翻譯漏字的瑕疵都抓出來了。
Generating your HTML analysis reports is now significantly faster. Previously, the system was stubbornly inefficient, re-reading every single one of your historical log files from scratch every time you asked for a report. It now employs caching (saving previously read data into memory): if a file hasn't changed, it instantly reuses what it already remembers. Meanwhile, we've subjected the system's foundation to a massive type-checking sweep (a health check ensuring every variable holds the right kind of data). We uncovered a 35% blind spot in the codebase, and by covering it, we swept out several latent bugs involving mixed-up variables and redundant settings. We also added an automated failsafe for translations, meaning if a single word is left untranslated in any language, the build will immediately fail before release, stopping the weird glitch where a Japanese menu would silently fall back to English. It's like a librarian who used to recount every single book in the building whenever you asked for inventory, but has finally learned to just read the ledger. Plus, the library just passed a full-building safety inspection that shined a light into all the neglected dark corners, fixing even the missing translations on the exit signs.
**File-level cache for the HTML report loaders**: `adapters/claude.py` and `adapters/codex.py` gain an `mtime`+`size`-keyed LRU cache (matching `history_loader`), so generating a report no longer re-parses every JSONL log on each run; the Codex adapter shares one cache between `load_entries` and `load_rate_limits`. Whole-file `OSError` / `PermissionError` / `sqlite3.Error` are now printed to stderr when `USAGE_DEBUG=1` (per-line `JSONDecodeError` stays silent).
**mypy `--strict` now covers the whole codebase**: removed the mypy exclude for `adapters/`, `analyzer/`, `ui/` and `usage_cli.py` (a ~35% type-checking blind spot), added the missing generics and function annotations, and switched `_group_by_agent` to a PEP 695 type parameter. `mypy --strict` now checks all 70 source files.
**Three cross-module functions in `adapters/claude.py` are now public API**: `get_claude_dirs`, `extract_project_from_dir`, `parse_jsonl` (previously underscore-private), dropping the matching `# type: ignore[attr-defined]` in `analyzer/reporter.py`.
Removing the mypy exclude surfaced and fixed a few latent issues: a redundant `parsed_entries` re-annotation left over from the cache change in `adapters/claude.py`, the `agent` loop variable reused with two different types in `analyzer/reporter.py` (inner accumulator renamed `agent_totals`), and a redundant `cast` in `menubar.py`.
Added coverage for `_apply_sort` with the `"time"` sort key (which maps to `None` and is handled per-command).
Added an i18n key-parity test asserting all five `i18n.json` language sections share the same key set, so a forgotten translation fails CI instead of silently falling back to English.