v0.9.3

🚀 Coding Usage v0.9.3 更新说明 / Release Notes

🇨🇳 中文

• DSH（DeepSeek Harness）用量改为主进程自行读取，不再交给 tokscale：后者的 DSH 解析器遇到一个解析不了的事件会把整份会话丢弃且不报错（本机实测：一个 21 MB、3,020 次调用、约 12.6 亿 token 的会话因末尾一个事件被整份跳过）。新读取器逐帧解压追加式 zstd、逐行容错，并同时读 DSH 应用自己的用量插件存储（~/.dsh/storages/dsh-app-plugin-usage/usage.jsonl）——转录只能覆盖仍在磁盘的会话，被应用清理掉的会话只剩那份记录；两个来源逐事件合并：转录能记账的每个 sessionId#seq 先占位，账本只补转录真没覆盖到的事件，既不会重复也不会因为一段坏尾巴 / 截断丢掉整会话。实测本机 DSH 由 tokscale 的 23.3 亿回到 112.6 亿 token / 36,151 次调用，与应用内用量插件（去重后 112.5 亿）一致
• DSH 读取器的计数修正：真实 v4 的 fork 边界是 session/end-seed 的 data.inherited（不是 header.seedLength，忽略它会把父会话前缀重算一遍）；compaction/summary 的 provider/model 在 data 上而非 data.message.source，否则这些调用会落到 dsh/unknown 而无法定价；同一 message id 的递增快照取最终值；两个来源改为逐事件合并——转录知道的每个 sessionId#seq（已计、继承、被取代，以及“早期零计数快照但同代已有最终快照”的那些）都记为已占用，账本只补转录真没覆盖到的事件，既不会重复也不会因为一段坏尾巴 / 截断丢掉整会话。已解析会话按 inode/ctime/mtime/size 缓存，重复扫从 ~5.3 秒降到 ~0.15 秒
• 随之移除 electron/dsh-aliases.ts：不再需要把版本化转录硬链接成 tokscale 认的名字

🇬🇧 English

• DSH (DeepSeek Harness) usage is now read by the main process instead of tokscale: tokscale's DSH reader discards an entire transcript when it meets one event it cannot parse, silently (measured here: a 21 MB session with 3,020 calls and ~1.26 B tokens was skipped wholesale because of one event near its end). The new reader decodes the append-only zstd frames one by one, tolerates unparsable lines, and also reads the DSH app's own usage-plugin store (~/.dsh/storages/dsh-app-plugin-usage/usage.jsonl) — transcripts only cover sessions still on disk, while sessions the app has deleted survive only in that store. The two sources are merged event by event: every sessionId#seq a transcript can account for (counted, inherited, superseded by a later snapshot) is claimed, and the store only contributes events no transcript covers — so nothing is counted twice and neither a damaged tail nor a truncated transcript can discard a session. Measured here, DSH goes from tokscale's 2.33 B to 11.26 B tokens / 36,151 calls
• Three counting fixes in that reader: the real v4 fork boundary is session/end-seed with data.inherited (not header.seedLength, which double-billed a child session's inherited prefix); a compaction/summary names its channel on data, not on data.message.source, so those calls no longer land in dsh/unknown where they could not be priced; and a repeated message id keeps its final snapshot. The two sources are now merged event by event: every sessionId#seq a transcript can account for is claimed — counted, inherited, superseded by a later snapshot, or an early counter-less snapshot of a generation that later reports — and the store only contributes events no transcript covers — so nothing is counted twice and neither a damaged tail nor a truncated transcript can discard a session. Parsed sessions are cached by inode/ctime/mtime/size, taking a repeat scan from ~5.3 s to ~0.15 s
• electron/dsh-aliases.ts is gone with it: versioned transcripts no longer need hard-linked canonical names for tokscale

---

安装包见下方 Release 附件：Windows（NSIS 安装版 / 便携版）、macOS（dmg / zip，未签名，首次打开请右键 App →「打开」）、Linux（AppImage，chmod +x 后运行）。
Installers are attached below: Windows (NSIS installer / portable), macOS (dmg / zip, unsigned — right-click the App → Open on first launch) and Linux (AppImage, chmod +x then run).