# Changelog

All notable changes to `dsh-chat-import` are documented here, newest first.

## [0.26.4] - 2026-11-02

[中文](#cn-0.26.4) | [English](#en-0.26.4)

<h3 id="cn-0.26.4">新增功能</h3>

- **新增 CodeBuddy Code 导入源**：`import_chat({ format: "codebuddy", path })` 与面板支持腾讯云代码助手 CodeBuddy Code CLI 的会话目录 `~/.codebuddy/projects/<projectKey>/`（`<sessionId>.jsonl` 逐行事件 + 同名 `.meta.json` 侧车，可传 projects 根、projectKey 目录、单个会话文件；子代理会话 `<parentSessionId>/subagents/agent-*.jsonl` 不并入主会话）。事件词汇是 OpenAI Agents SDK 条目加扩展字段：标题与 cwd 取 `custom-title` / `ai-title` 与记录 `cwd`，推理走 `reasoning` 条目，工具调用与结果按 `callId` 配对回声明它的步（未配对的孤儿结果计数 `droppedOrphanResults`），`providerData.usage` / `rawUsage` 的 input / output / cacheRead / reasoning 四桶映射为逐步 usage（camelCase 与 snake_case 两代键名都认），逐条 `timestamp` 透传成事件时间。框架注入的 user 条目（`isMeta` / `skipRun` / `startsNewUserRequest: false` / compact agent / 附带数据）不开轮、不作标题回退，只按结构化标记判定、不看内容正则。运行期条目（`turn-metrics` / `context-usage` / `credit-usage` / `file-history-snapshot` 等）不进正文，计数上报 `droppedMetaItems`。`local-jsonl` 自动识别新增该格式（`.codebuddy/projects/` 路径特征）。与 WorkBuddy 同源的解析逻辑收在共用引擎里，两个来源各留一份差异描述。
- **新增 CodeArts Desktop 导入源**：`import_chat({ format: "codearts", path })` 与面板支持华为云码道 CodeArts 桌面客户端内置 agent 的会话目录 `~/.codearts/kernel/sessions/<启动桶>/<session-uuid>/`（`history.jsonl` + `meta.json`，可传 sessions 根、启动桶、单个会话目录或直传 `history.jsonl`；根被迁移过时用 `CODEARTS_HOME` 指向）。消息正文是 Rust serde 变体（`SinglePart` / `MultiPart` 下的 `Text` / `ToolCall` / `ToolResult`），转换器摊平成标准内容块：标题与 cwd 取 `meta.json`，推理走 `reasoning_content`，工具结果按 `tool_call_id` 跨轮回看配对（未配对的孤儿结果计数 `droppedToolResults`），`token_usage` 的 prompt / completion / reasoning / cached 四桶映射为逐步 usage，逐消息 RFC3339 时间戳透传成事件时间。宿主注入的上下文消息（`synthetic: true`，如 todowrite 快照）仍进日志保序，但来源标为插件注入——UI 折叠、不进消息计数、不作标题回退，计数上报 `injectedPrompts`。`local-jsonl` 自动识别新增该格式（`.codearts/kernel/sessions/` 路径特征或 `history.jsonl` 文件名）。本机实测 3 个真实会话 342 条消息：199 次工具调用与 199 条结果全部配对，零丢弃。

<h3 id="cn-0.26.4">体验优化</h3>

- **按会话时间命名的 cwd 不再产生日期名工作区**：cwd 末段是纯时间戳时（CodeArts Desktop 的 `codearts/2026-10-30-00-40-33`、WorkBuddy Desktop 的 `WorkBuddy AI/2026-10-27-13-46-52` 这类每会话新建的工作目录）按「不可用 cwd」处理，与主目录同级——`auto` 落专用导入工作区、`per-project` 进未分组，`ungroupedReason` 报 `cwd-ephemeral`。此前每个会话都会就地建出一个日期名工作区。识别只认末段且要求整段是时间戳形态，带后缀的项目目录（`proj/2026-10-30-release`）不受影响。

<h3 id="cn-0.26.4">问题修复</h3>

- **WorkBuddy Desktop 5.9.x 改数据根后会话扫不到（0 条）**：`~/.workbuddy` 只剩 `logs/`，会话移到 `~/.workbuddy-ai/projects/`；旧实现把根写死成 `~/.workbuddy/projects`，且扫描器按 `.workbuddy/projects` 路径签名自拒，因此即使显式传 `path` 也返回 0。现在 `roots` 同时返回两代数据根（老布局继续可用），路径签名、单文件判格式与 `layoutProject` 都认 `.workbuddy` 与 `.workbuddy-ai` 两种。本机实测同一批 11 个会话从 0 条恢复到全部可见。
- **Cursor 旧对话在项目移动后仍可恢复工作区**：发现、旧扫描书签与导入优先按 composerId 查询本机 Cursor 工作区登记册，只采用现存目录；导入始终重新读取，发现缓存随 DB/WAL 指纹变化失效，失败结果不缓存。畸形登记 JSON 与文件系统读取失败计数告警，缺表或锁库仍静默降级。缺登记、旧版库或库不可读时保留 slug 解码兜底，避免仅因旧目录名失效而落入「无工作区」。
- **显式预算不再能把导入会话写死：参数 / 环境变量给出的预算超过模型窗口时按窗口钳制，并用 `budgetClampedFrom` 回报原值**：`resolveImportBudget` 对参数与 `DSH_IMPORT_CONTEXT_BUDGET` 直接原值返回，只有动态分支才看模型窗口，于是调用方传一个比窗口大的数就会把所有保护同时关掉——整段历史全量落盘，恢复对话时每轮都 `CONTEXT_WINDOW_EXCEEDED`，而自动压缩自己要发一次超窗请求也发不出去，那条会话除了重导没有别的出路。现在两条分支都按动态分支同一条上限（窗口 − 输出上限 − max(25% 窗口, 40k)）钳制并回报原值，导入结果里明说「预算被钳制」、registry 落钳后值；宿主报不出窗口时按原值使用——钳制只在已知上限上生效，不用来拦导入。同一份源实测：显式 750000 落 10,040 事件、表面 7.2 万 token；显式 500000000 落 105,883 事件、表面 849 万 token，每轮 400。

<h3 id="en-0.26.4">New Features</h3>

- **New import source: CodeBuddy Code**: `import_chat({ format: "codebuddy", path })` and the panel accept the session directories of Tencent Cloud's CodeBuddy Code CLI `~/.codebuddy/projects/<projectKey>/` (`<sessionId>.jsonl` line-per-event plus a same-named `.meta.json` sidecar) — pass the projects root, a projectKey directory, or a single session file; sub-agent sessions (`<parentSessionId>/subagents/agent-*.jsonl`) are not folded into the parent session. The event vocabulary is OpenAI Agents SDK items plus extended fields: titles and cwd come from `custom-title` / `ai-title` and the record `cwd`, reasoning from `reasoning` items, tool calls and results pair back to the step that announced the call by `callId` (unpaired orphans count as `droppedOrphanResults`), the input / output / cacheRead / reasoning buckets of `providerData.usage` / `rawUsage` map to per-step usage (both the camelCase and snake_case key generations are accepted), and per-item `timestamp` values become event times. Framework-injected user items (`isMeta` / `skipRun` / `startsNewUserRequest: false` / compact agent / additional data) open no turn and never become the title fallback — decided purely by structured markers, never by content regex. Runtime items (`turn-metrics` / `context-usage` / `credit-usage` / `file-history-snapshot` …) contribute no body text and are reported as `droppedMetaItems`. `local-jsonl` auto-detection now also recognizes this format (the `.codebuddy/projects/` path hint). The parsing logic shared with WorkBuddy lives in one engine; each source keeps only its own difference profile.
- **New import source: CodeArts Desktop**: `import_chat({ format: "codearts", path })` and the panel accept the built-in agent session directories of Huawei Cloud's CodeArts desktop client (码道 CodeArts) `~/.codearts/kernel/sessions/<bucket>/<session-uuid>/` (`history.jsonl` + `meta.json`) — pass the sessions root, a bucket, a session directory or a single `history.jsonl`; set `CODEARTS_HOME` when the `.codearts` root was moved. Message bodies are Rust serde variants (`Text` / `ToolCall` / `ToolResult` under `SinglePart` / `MultiPart`); the converter flattens them into standard content blocks: titles and cwd come from `meta.json`, reasoning from `reasoning_content`, tool results pair back to their call across turns by `tool_call_id` (unpaired orphans count as `droppedToolResults`), the prompt / completion / reasoning / cached buckets of `token_usage` map to per-step usage, and per-message RFC3339 timestamps become event times. Host-injected context messages (`synthetic: true`, e.g. todowrite snapshots) keep their place in the log but are attributed to the plugin — folded in the UI, excluded from message counts and title fallback, reported as `injectedPrompts`. `local-jsonl` auto-detection now also recognizes this format (the `.codearts/kernel/sessions/` path hint or a `history.jsonl` file name). Verified against three real local sessions (342 messages): all 199 tool calls paired with their 199 results, zero dropped.

<h3 id="en-0.26.4">Improvements</h3>

- **Session-timestamp-named cwds no longer produce date-named workspaces**: when a cwd's last segment is a bare timestamp (the per-session working directory CodeArts Desktop (`codearts/2026-10-30-00-40-33`) and WorkBuddy Desktop (`WorkBuddy AI/2026-10-27-13-46-52`) create), it counts as an unusable cwd, on par with the home directory — `auto` falls back to the dedicated import workspace, `per-project` leaves the session ungrouped and reports `ungroupedReason: "cwd-ephemeral"`. Previously every session created its own date-named workspace. Detection looks only at the last segment and requires it to be timestamp-shaped, so project directories carrying a date suffix (`proj/2026-10-30-release`) are unaffected.

<h3 id="en-0.26.4">Bug Fixes</h3>

- **WorkBuddy Desktop 5.9.x sessions invisible after the data-root rename (0 found)**: `~/.workbuddy` now holds only `logs/`; sessions moved to `~/.workbuddy-ai/projects/`. The old code hard-coded `~/.workbuddy/projects` and the scanner rejected any path not matching the `.workbuddy/projects` signature, so even an explicit `path` returned 0. `roots` now returns both data roots (the old layout keeps working), and the path signature, single-file format detection and `layoutProject` accept both `.workbuddy` and `.workbuddy-ai`. Verified locally: the same 11 sessions went from 0 to fully listed.
- **Cursor conversations can recover their workspace after a project moves**: discovery, existing scan bookmarks and imports prefer the local Cursor workspace registry by composerId, accepting only existing directories. Imports always read fresh; discovery invalidates cached results on DB/WAL fingerprint changes and does not cache failures. Malformed registry JSON and filesystem failures produce counted diagnostic warnings, while missing tables and locked databases still degrade silently. Missing entries, older databases and unreadable registries retain slug decoding as a fallback, so a stale transcript directory name alone no longer leaves a conversation without a workspace.
- **An explicit budget can no longer write a session dead: parameter / env values above the model window are clamped to it and reported as `budgetClampedFrom`**: `resolveImportBudget` returned a parameter or `DSH_IMPORT_CONTEXT_BUDGET` value untouched and only ever consulted the window for the dynamic branch, so a caller passing a number larger than the window silently disabled every layer of protection at once — the whole history was written to disk, each recovered turn exceeded the model limit with `CONTEXT_WINDOW_EXCEEDED`, and auto-compaction could not run either because it must send its own over-window request, leaving the session unrecoverable except by re-importing. Both branches now clamp to the same ceiling the dynamic branch uses (window − output cap − max(25% of window, 40k)) and echo the original value back as `budgetClampedFrom`; the import note says the budget was clamped, and the registry records the clamped value. When the host cannot report the window the value is used as-is — clamping is a guard on a known limit, never a reason to block an import. Measured on one source: an explicit 750000 budget produced 10,040 events and a 72k-token surface; an explicit 500000000 produced 105,883 events and an 8.5M-token surface that answered every turn with 400.

**Full Changelog**: [v0.26.3...v0.26.4](https://github.com/Nwflower/dsh-chat-import/compare/v0.26.3...v0.26.4)

## [0.26.3] - 2026-10-27

[中文](#cn-0.26.3) | [English](#en-0.26.3)

<h3 id="cn-0.26.3">新增功能</h3>

- **新增 Copilot VS Code 导入源**：`import_chat({ format: "copilot-vscode", path })` 和面板支持 version 3 `chatSessions/*.json`，可传 `workspaceStorage` 根、会话目录或单个文件。保留标题、源时间、推理、已保存的工具输入与结果；相邻 `workspace.json` 提供单文件夹工作区。工具参数及结果缺失数量通过 `missingToolInputs`、`missingToolOutputs` 上报。
- **新增 Copilot CLI 导入源**：`import_chat({ format: "copilot-cli", path })` 与面板支持 Copilot CLI 的逐会话事件流 `~/.copilot/session-state/<id>/events.jsonl`，可传 `session-state` 根、单个会话目录或单个文件；`.copilot` 根被迁移过时用 `COPILOT_HOME` 指向。事件流按顺序还原提问、回复、推理与按 `toolCallId` 配对的工具调用；`session.compaction_complete` 的摘要写为 DSH 原生压缩检查点（边界之前的轮标 log-only，正文保留在日志里），`fullHistory: true` 保留全量并进参数指纹。标题按首条人类提问兜底；事件格式上游无稳定承诺，未知类型、不可读行、未落地附件、未声明补出的调用与孤儿结果都计数进公开结果（`unknownEvents`、`skippedLines`、`secrets`、`droppedAttachments`、`unannouncedToolCalls`、`orphanResults`），不会让整份会话失败。
- **面板会话行新增四个行内工具：预览、正文搜索、隐藏、删除源文件，并支持拖到侧栏工作区导入**：双击会话行打开预览卡，按顺序显示提问与回复，插件注入的消息不出现，连续工具调用折成一行（如 `Bash ×2, Read`），预览只跑导入的只读 dry-run；「搜内容」把面板搜索框从标题切到正文（只覆盖提问与回复文本，工具入参与结果不计，一库多会话或整份 JSON 的来源跳过并计数），同一条消息须包含全部关键词，命中数与高亮随列表轮询更新，换词或换来源会取消上一次检索；「隐藏」在忽略表里登记为可撤销的隐藏项，重扫不再出现，撤销即恢复，且不会覆盖已有的归档 / 撤回记录；「删除文件…」在确认后把源文件移进回收站，一库多会话的来源与按自定义路径扫描出来的行一律拒绝，DSH 侧已导入的副本保留；把行拖到左侧栏的工作区上即导入到该工作区，副本时间轴平移到「现在」，排在该工作区最上面。
- **新增豆包桌面版对话导入源**：`import_chat({ format: "doubao", path })` 与面板「导入会话」都能导入豆包的会话（默认根 Windows `%LOCALAPPDATA%\Doubao\User Data`，也可直接给某个库目录）。库是 Chromium IndexedDB 的 LevelDB 目录而不是单个文件，读出正文要过三层编码：LevelDB 的 SSTable 内部键带 8 字节尾标（`WriteBatch` 的键没有）、值是 `FF 11 02`（snappy）或 `FF 11 01`（外置 Blob）包装、剥完才是 Blink 的 `SerializedScriptValue` 头加 V8 结构化序列化流。正文在一条固定主键里，而豆包把**流式生成状态**整条写在该键上（每次增量都是一次重写，生成结束再写一条清空快照），所以必须扫该键的全部历史版本、按 taskId 取最长快照才能还原完整回答——只读当前值会得到空会话。**有意的范围限制**：豆包的会话列表在服务端，本地只为最近打开过的对话保留正文；因此只列出并导入**本地确实有正文**的会话，缺正文的不列出、不导入，也不拿标题凑空壳；一个都读不到时报错并点名库路径与解不开的记录数。

<h3 id="cn-0.26.3">体验优化</h3>

- **Kimi Code 导入保持源消息顺序，减少缓存前缀失效**：同一步的并行工具结果按源转录里的完成顺序发射，不再重排成工具声明顺序；连续的同文注入也不再被去重吞掉。用一条 45 轮、327 条模型可见消息的真实 Kimi Code 会话逐条对照：修复后导入序列与源完全一致，除 DSH 自身的系统头与环境变更声明外，转换不再造成前缀偏移。
- **Qoder 导入同时覆盖 Qoder CN 的数据根**：Qoder CN 与 intl 版共用逐字相同的 transcript 格式，`qoder` 格式与面板现在同时扫描 `~/.qoder/projects` 与 `~/.qoder-cn/projects`，单文件路径也按 `~/.qoder-cn` 识别；format 枚举不变，中英导入指南的 path 形状同步。

<h3 id="cn-0.26.3">问题修复</h3>

- **Kimi Code 的宿主注入不再显示为用户消息**：`context.append_message` 的 `message.origin` 区分真人输入（`user`）与宿主注入（`injection` / `task` / `system_trigger`：日期、权限、AGENTS.md、todo 提醒，后台任务完成通知，子代理事件）。此前只按 `role:"user"` 建轮，大量工具通知和系统提醒会以用户气泡出现；现在按 `source.kind:"plugin"` 的上下文注入导入，UI 折叠显示，正文与顺序不变，不计入消息数，也不会被当作首问标题。`turn.prompt` 与紧随其后的同文 `append_message` 只算一条，注入不会重复。
- **Trae Work / Trae CN 导入读取 icube-ai-chat-storage / icube-ai-ng-chat-storage 键家族的全部装机键**：两个聊天存储键的数字后缀随安装变化（社区导出脚本按 `icube-ai-ng-chat-storage` 前缀查询 ItemTable），此前固定后缀只匹配观察到的一台安装，其它安装的会话被漏读、面板列表为空。现在按前缀吸收两个家族的全部键（带或不带 `memento/` 前缀都认），同一会话在多键下命中仍按会话 id 去重；其余 icube-ai-* 条目（遥测、UI 状态）不属于会话存储，不读取、也不因此触发「有条目却认不出」的失败。
- **Qoder CLI 导入还原源侧的上下文压缩，导入会话不再堆到模型窗口的数倍**：Qoder CLI 压缩时落两条记录（`type:"system"` + `subtype:"compact_boundary"` 的边界标记，`isCompactSummary:true` 的 user 记录承载摘要），此前两条都不认——摘要被当成普通提问（每次压缩多一个 16~20 KB 的假提问），边界之前的全部历史留在模型投影上，源侧压掉的内容全回来，每轮 `CONTEXT_WINDOW_EXCEEDED`，换模型、点继续都无效。现在摘要写进 DSH 原生压缩检查点、边界之前的轮标 log-only（正文仍在日志里），模型视角是最后一次摘要加压缩点之后的内容。`fullHistory: true` 保留旧行为（摘要按普通 user 记录导入、正文不丢），并随 qoder 一起登记进参数指纹：换值会照常报 `argsChanged`，不再被「源未变」短路径静默吃掉。同一份源实测：12 次与 47 次源侧压缩此前各生效 0 次、导入会话分别堆到 381 万与 850 万 token，修后为 57 万与 29 万。
- **与 `@deepseek-ai/dsh-tools` 0.2.2 兼容，插件在新版宿主上不再无法激活**：0.2.2 移除了 `TOOL_RUNTIME_SCHEDULER` 导出（`ToolRuntime` 成为 `ToolsService` 的别名），旧的哨兵探测在 0.2.2 宿主上必然抛错，整个插件入口无法加载。现在 `ToolsService` 或 `TOOL_RUNTIME_SCHEDULER` 任一存在即通过，两者皆无（过旧副本）仍然大声报错。
- **插件卸载 / 热重载后不再残留面板路由**：`webServer.register` 只返回 disposer，原来 11 条路由注册后全部丢弃，插件因服务重新提供或 HMR 卸载重跑时重复 `(kind, path)` 抛错、面板由已 dispose 的上下文接管（每个请求都报 `cannot get required service "fs" in inactive context`）。现在路由收集到 `ctx.effect` 里，卸载时逐条撤销；单条路径被别的插件占用时只警告，不拖垮整个面板。
- **Qoder 长会话的面板标题不再为空**：标题载体 `last-prompt` 是每轮追加的，长会话里唯一一份落在 256 KB 头窗之外；发现层改为读头 + 尾两段拼成文件序后取 `ai-title` / `last-prompt`，与导入侧同口径，并递增扫描缓存版本让旧书签自愈。
- **Hermes 导入补齐 OpenAI 形状的工具调用与独立结果行**：`tool_calls` 的嵌套 `function.name` / `function.arguments`（JSON 编码字符串逐字保留）与独立 `role:"tool"` 行都认，结果按 `tool_call_id` 挂回调用、孤儿结果计入 `droppedToolResults`；纯工具调用行正文为 `NULL` 时不再整行丢弃。同一份真实会话（155 条消息）此前只剩 58 assistant + 11 user、0 个工具事件，现在得到 79 对 tool/call 与 tool/result，6 条孤儿结果计数上报，参数逐字保真。
- **`restamp:'end'` 的时间锚点与投放参数指纹修正**：锚点改为事件时间的最大值（`restamp:true` 对称地取最小值），事件乱序时不会把「最后一个事件=现在」推到未来，`meta.createdAt` 也夹到不晚于当前；`workspaceMode` / `workspaceDir` / `restamp` 并入参数指纹，同一源换工作区再次拖放不再被「源未变」短路径当成同一份导入，未显式给出的参数不会产生伪差异。

<h3 id="en-0.26.3">New Features</h3>

- **New import source: Copilot VS Code**: `import_chat({ format: "copilot-vscode", path })` and the panel accept version 3 `chatSessions/*.json` from a `workspaceStorage` root, a session directory or a single file. Imports preserve titles, source times, reasoning and stored tool input/results; adjacent `workspace.json` supplies the workspace folder. Missing tool parameters and results are reported as `missingToolInputs` and `missingToolOutputs`.
- **New import source: Copilot CLI**: `import_chat({ format: "copilot-cli", path })` and the panel accept Copilot CLI's per-session event stream `~/.copilot/session-state/<id>/events.jsonl` — pass the `session-state` root, a session directory or a single file; set `COPILOT_HOME` when the `.copilot` root was moved. The stream replays prompts, replies, reasoning and tool calls paired by `toolCallId` in order; a `session.compaction_complete` summary becomes a native DSH compaction checkpoint (earlier turns are log-only, their text stays in the log) and `fullHistory: true` keeps the full surface, joining the args fingerprint. Titles fall back to the first user prompt. The event format has no upstream stability contract, so unknown types, unreadable lines, dropped attachments, calls inferred without a declaration and orphan results are all counted in the public result (`unknownEvents`, `skippedLines`, `secrets`, `droppedAttachments`, `unannouncedToolCalls`, `orphanResults`) instead of failing the session.
- **Four per-row tools in the import panel — preview, body search, hide and delete source — plus drag onto a workspace**: double-clicking a session row opens a preview card with the conversation's user and assistant turns in order, plugin-injected messages skipped and consecutive tool calls folded into one row (`Bash ×2, Read`); the preview runs the import's read-only dry-run. The "In content" toggle points the search box at message bodies instead of titles: user and assistant text only (tool inputs and results are excluded), sources whose transcript is a database or a one-file-many-sessions JSON are skipped and counted, every term must appear in the same message, and hit counts plus highlighting update by polling the same list; a new query or source cancels the previous run. "Hide" records a reversible hidden entry in the ignore registry, so rescans keep the row hidden and Undo restores it, without ever overwriting an existing archive or retract record. "Delete file…" moves the source to the Recycle Bin after confirmation, refusing one-file-many-sessions sources and rows that came from a custom-path scan, and keeps the imported DSH copy. Dragging a row onto a sidebar workspace imports it there with timestamps shifted so the copy's last event is now and it sorts to the top of that workspace.
- **New import source: Doubao Desktop**: `import_chat({ format: "doubao", path })` and the Import Sessions panel both import Doubao Desktop conversations (default root on Windows is `%LOCALAPPDATA%\Doubao\User Data`; a specific database directory also works). The database is a Chromium IndexedDB LevelDB **directory**, not a single file, and reading a message body crosses three encodings: LevelDB SSTable keys carry an 8-byte internal trailer (`WriteBatch` keys do not), values are wrapped as `FF 11 02` (snappy) or `FF 11 01` (payload moved to an external Blob), and only then does the Blink `SerializedScriptValue` header plus the V8 structured-serialization stream begin. Bodies live under one fixed primary key, and Doubao writes its **streaming generation state** into that key as a whole (every delta is a full rewrite, and finishing writes a snapshot with an emptied task map), so every historical version of that key must be scanned and the longest snapshot taken per task id — reading only the current value yields an empty session. **Deliberate scope limit**: Doubao keeps its conversation list server-side and caches bodies locally only for recently opened chats, so only conversations with a **locally cached body** are listed and imported; ones without a body are neither listed nor imported, and no shell session is fabricated from a title. When nothing is readable the import fails with the database path and the count of undecodable records.

<h3 id="en-0.26.3">Improvements</h3>

- **Kimi Code imports keep the source message order, so the cache prefix survives**: parallel tool results in one step are emitted in the source transcript's completion order instead of being reordered to the tool declaration order, and consecutive identical injections are no longer deduplicated away. Checked against a real 45-turn Kimi Code session with 327 model-visible messages: after the fix the imported sequence matches the source one-for-one, and apart from DSH's own system head and environment-change notice the conversion introduces no prefix shift.
- **Qoder imports now cover the Qoder CN data root**: Qoder CN shares the byte-identical transcript format with the international build, so the `qoder` format and the panel scan both `~/.qoder/projects` and `~/.qoder-cn/projects`, and a single-file path under `~/.qoder-cn` is recognised too; the format enum is unchanged and both import guides list the new path shape.

<h3 id="en-0.26.3">Bug Fixes</h3>

- **Kimi Code host injections no longer show up as user messages**: `context.append_message`'s `message.origin` separates real input (`user`) from host injections (`injection` / `task` / `system_trigger`: date, permission, AGENTS.md and todo reminders, background-task completion notices, subagent events). The converter previously opened a user turn for every `role:"user"` record, so many tool notices and system reminders appeared as user bubbles. They are now imported as `source.kind:"plugin"` context injections: collapsed by the UI, with text and order unchanged, excluded from the message count, and never used as the first-question title. A `turn.prompt` and its immediately following identical `append_message` count once, so an injection is not duplicated.
- **Trae Work / Trae CN imports read every install-specific key of the icube-ai-chat-storage / icube-ai-ng-chat-storage families**: the numeric suffix of both chat-storage keys varies per install (community export scripts query the ItemTable by the `icube-ai-ng-chat-storage` prefix), so the previously fixed suffix matched only the one installation it was observed on and other installs' sessions were missed, leaving an empty panel list. The reader now absorbs every key of both families by prefix (with or without the `memento/` prefix), a session found under several keys still deduplicates by session id, and remaining icube-ai-* entries (telemetry, UI state) are not session storage — they are not read and do not trigger the "entries exist but none could be recognised" failure.
- **Qoder CLI imports restore the source's context compactions, so an imported session no longer piles up to several times the model window**: a Qoder CLI compaction writes two records — a boundary marker (`type:"system"`, `subtype:"compact_boundary"`) and a user record carrying the summary (`isCompactSummary:true`) — and the converter recognised neither: the summary became an ordinary prompt (one bogus 16–20 KB question per compaction) while every turn before the boundary stayed on the model surface, bringing back exactly what Qoder had compacted away, so every round failed with `CONTEXT_WINDOW_EXCEEDED` regardless of the model or "continue". Summaries now become native DSH compaction checkpoints and pre-boundary turns are log-only (their full text stays in the log), so the model sees the last summary plus what follows the compaction point. `fullHistory: true` keeps the old behaviour (summary imported as an ordinary user record, nothing lost) and qoder now joins the args fingerprint, so flipping the flag reports `argsChanged` instead of being silently swallowed by the "source unchanged" fast path. On the same sources: 12 and 47 source compactions previously took effect 0 times, leaving imported sessions at 3.81M and 8.5M tokens; after the fix they are 0.57M and 0.29M.
- **Compatible with `@deepseek-ai/dsh-tools` 0.2.2, so the plugin activates on the new host again**: 0.2.2 removed the `TOOL_RUNTIME_SCHEDULER` export (`ToolRuntime` is now an alias of `ToolsService`), so the old sentinel check always threw and the whole plugin entry could not load. Either `ToolsService` or `TOOL_RUNTIME_SCHEDULER` now passes, while a copy with neither (too old) still fails loudly.
- **Panel routes no longer linger after an unload or hot reload**: `webServer.register` only returns a disposer and all 11 routes dropped it, so when the plugin was unloaded and re-applied (services re-provided, or HMR saving a file) duplicate `(kind, path)` registration threw and the panel was left on a disposed context (`cannot get required service "fs" in inactive context` on every request). Routes are now collected through `ctx.effect` and disposed one by one on unload; a path already taken by another plugin only warns instead of taking the whole panel down.
- **Long Qoder sessions no longer show an empty panel title**: the `last-prompt` carrier is appended every turn, and in a long session its only copy sits beyond the 256 KB head window. Discovery now reads the head and tail, merges them in file order and takes `ai-title` / `last-prompt` on the same terms as import; the scan cache version is bumped so old bookmarks heal.
- **Hermes imports now cover OpenAI-shaped tool calls and standalone result rows**: nested `function.name` / `function.arguments` in `tool_calls` (JSON-encoded strings kept verbatim) and standalone `role:"tool"` rows are recognised, results are attached to their call by `tool_call_id`, orphan results count as `droppedToolResults`, and a pure tool-call row with a `NULL` body is no longer dropped whole. On one real 155-message session that previously left 58 assistant + 11 user messages and zero tool events, the import now produces 79 tool/call + tool/result pairs, reports 6 orphan results, and preserves arguments byte-for-byte.
- **`restamp:'end'` anchors on the latest time and destination options join the fingerprint**: the anchor is now the maximum event time (`restamp:true` symmetrically uses the minimum), so out-of-order events cannot push "last event = now" into the future, and `meta.createdAt` is clamped to `now`; `workspaceMode` / `workspaceDir` / `restamp` are part of the args fingerprint, so dropping the same source into another workspace is no longer swallowed by the "source unchanged" fast path, while omitted options keep the old fingerprint and introduce no false differences.

**Full Changelog**: [v0.26.2...v0.26.3](https://github.com/Nwflower/dsh-chat-import/compare/v0.26.2...v0.26.3)

## [0.26.2] - 2026-10-19

[中文](#cn-0.26.2) | [English](#en-0.26.2)

<h3 id="cn-0.26.2">新增功能</h3>

- **新增 oh-my-pi（omp）对话导入源**：`import_chat({ format: "omp", path })` 与面板「导入会话」都能导入 omp 的会话（默认根 `~/.omp/agent/sessions`，一条会话一个 JSONL）。按 omp 的存储契约读：首行**定宽 256 字节的标题槽**作标题（旧文件没有槽时回退 `header.title`、再回退 `title_change`）、会话头取 id/时间/cwd、`id`/`parentId` 树只重建活动分支；`message` 的 camelCase 角色（`user` / `assistant` / `toolResult` / `bashExecution` / `pythonExecution` / `fileMention`）、`toolCall` 块（`arguments` 对象）、压缩条目（`firstKeptEntryId` 定保留窗口）按源自身 `convertToLlm` 的口径转换，压缩摘要默认导入为原生压缩检查点。`/clear` 的 `reset_boundary` 是纯标记、没有摘要载荷：与 Kimi「只有边界没有摘要」同口径，最后一次清空点之前的内容不进日志，结果里显式上报 `resetBoundaries: <N>`（不静默），`fullHistory: true` 保留全量——该开关进参数指纹。转换实现与 pi 共用同一引擎（`lib/convert/pi-family.mjs`），两个来源只保留各自的差异描述。
- **工具结果新增两个降级计数**：`resetBoundaries`（omp 的 `/clear` 次数，见上）与 `droppedInjected`（无轮可挂的注入消息数：bash / python 执行活动）。后者此前在 pi 上叫 `droppedBash` 且**没有在输出 schema 里声明**，被宿主的输出校验丢掉、用户看不到——现已声明并随结果透出。

<h3 id="cn-0.26.2">体验优化</h3>

- **`import_chat` 的常驻 schema 瘦身 59%，长文改由运行时 skill 按需加载**：schema 只留一次调用需要的东西——format 枚举、必填 `path`，以及「默认值会改变结果」的那几个参数各一行说明；30 个来源的 path 形状表、完整参数详解与重导规则下沉到 `docs/IMPORT-GUIDE.md`，并注册为宿主运行时 skill `dsh-chat-import-usage`（`import_chat` 的 description 点名它，模型在真要导入前加载一次）。口径是宿主真正发给模型的部分（name + description + parameters；`output.schema` 不上线、不计费），本机实测：`import_chat` 常驻 1633 → 663 tokens，全量 12 个工具 3720 → 2762 tokens；技能目录多出的一行约 46 tokens，故精简档净省约 0.9k/轮。设置页「精简档省多少」原先写的约 4.5k 有误，按同口径修正为约 2.1k；「关闭注入」的约 5k 修正为约 2.8k。**22 个参数一个没少**（全部保留声明，只是说明搬到了按需付费的地方）；`docs/USAGE*.md` 与中英两份 README 的文档表同步。
- **常驻成本进护栏，防止说明重新长回 schema**：`test/tools-injection.test.mjs` 的全量档 payload 上限从 15.3k 收紧到 12k，并单独钉住 `import_chat` ≤ 2.8k 字符；新增 `test/import-guide.test.mjs` 守住三条——指南的来源表与代码里的来源表逐字一致（中英两份）、每个声明的参数在指南里都有条目、schema 不再携带来源 path 形状表。

<h3 id="cn-0.26.2">其他变更</h3>

- **发布流程移入 CI**：发布不再在本地执行 `npm publish` / `gh release`。推送 `v*` tag 后，`.github/workflows/release.yml` 校验 tag 与 `package.json` 版本一致、tag 提交在 `main` 上、CHANGELOG 版本节版式正确（`scripts/release-notes.mjs`），跑完全部门禁后以 trusted publishing + provenance 发布到 npm，并用同一版本节建 / 更新 GitHub Release；重跑会跳过 npm 上已有的版本。手动 `workflow_dispatch` 只校验、不发布。

<h3 id="en-0.26.2">New Features</h3>

- **New import source: oh-my-pi (omp)**: `import_chat({ format: "omp", path })` and the Import Sessions panel both import omp sessions (default root `~/.omp/agent/sessions`, one JSONL per session). It follows omp's storage contract: the first line is a **fixed-width 256-byte title slot** used as the title (legacy files without it fall back to `header.title`, then `title_change`), the session header supplies id/time/cwd, and the `id`/`parentId` tree is rebuilt for the active branch only. The camelCase `message` roles (`user` / `assistant` / `toolResult` / `bashExecution` / `pythonExecution` / `fileMention`), `toolCall` blocks (`arguments` objects) and compaction entries (`firstKeptEntryId` marks the kept window) convert on the source's own `convertToLlm` terms, and compaction summaries become native DSH compaction checkpoints by default. `/clear`'s `reset_boundary` is a pure marker with no summary payload: by the same rule as Kimi's summary-less wire, content before the last clear point does not enter the log and the result reports `resetBoundaries: <N>` explicitly (never silently), while `fullHistory: true` keeps everything — that flag is part of the args fingerprint. Conversion shares one engine with pi (`lib/convert/pi-family.mjs`); each source keeps only its own differences.
- **Two conversion counters added to tool results**: `resetBoundaries` (omp `/clear` count, above) and `droppedInjected` (injected messages with no step to attach to: bash / python execution activity). The latter was previously `droppedBash` on pi and **not declared in the output schema**, so the host's output validation dropped it and users never saw it — it is now declared and reported.

<h3 id="en-0.26.2">Improvements</h3>

- **`import_chat`'s resident schema is 59% smaller; the long form now loads on demand as a runtime skill**: the schema keeps only what a call needs — the format enum, the required `path`, and a one-line note for the parameters whose default would otherwise change the result. The 30-source path table, the full parameter reference and the re-import rules moved into `docs/IMPORT-GUIDE.md`, registered as the host runtime skill `dsh-chat-import-usage` (named by `import_chat`'s description, loaded once before an import actually happens). Measured on what the host really sends to the model (name + description + parameters; `output.schema` never reaches the model and is not billed): `import_chat` 1633 → 663 tokens resident, all 12 tools 3720 → 2762 tokens; the extra skill-catalog line costs ~46 tokens, so Minimal nets ~0.9k saved per request. The settings page's "Minimal saves ~4.5k" was wrong and is corrected to ~2.1k on the same basis; "Off saves ~5k" is corrected to ~2.8k. **All 22 parameters are still declared** — only their explanations moved to where they are paid for on demand; `docs/USAGE*.md` and both README doc tables follow.
- **Resident cost is now guarded so explanations cannot creep back into the schema**: the full-surface payload ceiling in `test/tools-injection.test.mjs` tightened from 15.3k to 12k, plus a dedicated `import_chat` ≤ 2.8k ceiling; the new `test/import-guide.test.mjs` holds three lines — the guide's source table matches the code's table verbatim (both languages), every declared parameter has a guide entry, and the schema no longer carries the per-format path table.

<h3 id="en-0.26.2">Chores</h3>

- **Releases now run in CI**: publishing no longer happens locally (`npm publish` / `gh release`). Pushing a `v*` tag makes `.github/workflows/release.yml` check that the tag names `package.json`'s version and sits on `main`, validate the CHANGELOG section (`scripts/release-notes.mjs`), run the full gate set, publish to npm with trusted publishing and provenance, and create or update the GitHub Release from the same section; reruns skip a version already on npm. A manual `workflow_dispatch` validates without publishing.

**Full Changelog**: [v0.26.1...v0.26.2](https://github.com/Nwflower/dsh-chat-import/compare/v0.26.1...v0.26.2)

## [0.26.1] - 2026-10-17

[中文](#cn-0.26.1) | [English](#en-0.26.1)

<h3 id="cn-0.26.1">问题修复</h3>

- **Claude Code 的 fork 会话此前一条都扫不到，现在被发现并导入为独立会话**：新版 Claude Code（桌面端会话菜单与 `--fork-session`）的 fork 转录把父会话的记录**逐字拷贝**在文件开头，自己的记录从 fork 点起追加——旧的「文件名 stem == 首条记录的 sessionId」口径把整份 fork 判成子代理辅助 transcript，发现层与导入层都看不到它（本机实测 `~/.claude/projects` 下 35 份转录：4 份 fork 全部漏掉，只认出 31 份）。主转录判定改为看文件两端（文件名 stem 出现在头部**或**尾部即算该文件自己的会话），会话身份取转录里**最后**出现的 `sessionId`：fork 因此以自己的 id 落成 `import-<fork id>`，不再与父会话的 `import-<父 id>` 撞车、也不再被「文件名与 sessionId 不一致」的辅助转录判定跳过。拷贝段（父会话历史）照常进日志——fork 的上下文里本就有它，父会话与 fork 会话在 DSH 里因此重叠一份历史。
- **Codex 会话标题优先取保存的标题索引**：面板此前只按 rollout 里的首问起标题，Codex 里改过的名字显示不出来；现在优先读 `session_index.jsonl` 的 `thread_name`（按行流式读取，同一 thread 取最后一条有效记录），没有索引标题才回退首问。首问兜底同时与导入共用同一提取口径：注入的页面上下文信封被剥掉——信封独占一条消息、或与真实提问同在一段文本里都只保留提问，上下文不再变成标题。
- **Codex 会话的活动时间改为取日志里的时间戳**：面板排序此前用 rollout 文件的修改时间，复制、恢复或同步过的会话时间会整体失真；现在从每条 rollout 的头尾有界窗口取最新有效时间戳（滚过分页合并），文件修改时间只在日志里完全没有时间戳时兜底。被窗口截断的记录中的时间戳可能遗漏。
- **Windows 下 Hermes 的默认扫描根补全**：此前只扫 `~/.hermes`，而 Windows 上的 Hermes 把数据放在 `%LOCALAPPDATA%\hermes`，面板一条会话都列不出来（手动填绝对路径可以导入）。现在 Windows 同时扫 `%LOCALAPPDATA%\hermes` 与 `~/.hermes`，`LOCALAPPDATA` 未设置或为空时回退 `~/AppData/Local/hermes`；macOS 与 Linux 仍为 `~/.hermes`。每个根优先读 `state.db`，数据库不可用时回退 JSONL。

<h3 id="cn-0.26.1">其他变更</h3>

- **CI 覆盖率护栏恢复生效**：`check-coverage` 解析器此前只认 `ℹ ` 前缀，而 CI 的 Node 22 输出 `# `，从 2026-10-13 起每个 PR 的 test job 都红（一行报告都解析不到，误报「目录下限有记账但目录下没有文件」）。现在两种前缀都收，`coverage` 脚本再固定 `--test-reporter=spec`，Node 22 / 24 口径一致；同时补上解析器与 `clearSessionArtifactsForReplace` 的用例，Node 22 下的逐文件下限继续有效。

<h3 id="en-0.26.1">Bug Fixes</h3>

- **Claude Code fork sessions were invisible; they are now discovered and imported as their own sessions**: a fork transcript written by newer Claude Code (the desktop session menu and `--fork-session`) opens with a **verbatim copy** of the parent's records and appends its own from the fork point, so the old test — "file stem == the first record's sessionId" — classified the whole file as a subagent auxiliary transcript, invisible to both discovery and import (measured here: all 4 forks among the 35 transcripts under `~/.claude/projects` were missed, only 31 recognised). The main-transcript test now looks at both ends of the file (the stem at either end means the file owns that session) and session identity is the **last** `sessionId` in the transcript: a fork lands as `import-<fork id>`, no longer colliding with the parent's `import-<parent id>` and no longer skipped by the "file name does not match sessionId" auxiliary-transcript rule. The copied prefix (the parent's history) still goes into the log — a fork's context does contain it, so parent and fork overlap by one stretch of history in DSH.
- **Codex discovery prefers the saved title index**: the panel used to title sessions from the rollout's first prompt only, so a rename made in Codex was invisible. Discovery now reads `thread_name` from `session_index.jsonl` first (streamed line by line; the last valid record wins) and falls back to the first prompt only when no indexed title exists. That fallback now shares the import extraction: injected page-context envelopes are stripped — whether the envelope is its own message or shares a block with the real prompt, only the prompt is kept, so context text no longer becomes the title.
- **Codex activity time now comes from transcript timestamps**: the panel ordered sessions by rollout file mtime, so copied, restored or synced trees showed skewed times. The latest valid timestamp is now read from bounded head and tail windows of each rollout page and combined across the thread; file mtime is only a fallback when no timestamp exists at all. A timestamp inside a truncated record can be missed.
- **Hermes default discovery now covers the Windows layout**: only `~/.hermes` was scanned, while Windows Hermes stores its data in `%LOCALAPPDATA%\hermes`, so the panel listed nothing (importing by absolute path worked). Windows now scans both `%LOCALAPPDATA%\hermes` and `~/.hermes`, falling back to `~/AppData/Local/hermes` when `LOCALAPPDATA` is unset or empty; macOS and Linux still use `~/.hermes`. Each root reads `state.db` first and falls back to JSONL when the database is unavailable.

<h3 id="en-0.26.1">Chores</h3>

- **The CI coverage gate is effective again**: the `check-coverage` parser only accepted the `ℹ ` prefix while CI's Node 22 prints `# `, so every PR's test job had been red since 2026-10-13 (no report row was parsed, misreported as "a directory floor is booked but the directory has no files"). Both prefixes are now accepted and the `coverage` script pins `--test-reporter=spec` so Node 22 and 24 agree; tests now pin the parser for both prefixes and cover `clearSessionArtifactsForReplace`, keeping the per-file floors effective on Node 22.

**Full Changelog**: [v0.26.0...v0.26.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.26.0...v0.26.1)

## [0.26.0] - 2026-10-13

[中文](#cn-0.26.0) | [English](#en-0.26.0)

<h3 id="cn-0.26.0">新增功能</h3>

- **新增 `chatImport` Host Service，供第三方插件复用整条导入管线**：插件入口注册可选服务 `chatImport`（经宿主的 `ctx.provide`），暴露四个方法——`scanDiscover()`（只读发现，与 `scan_discover` 同一批扫描器与缓存）、`importPath()`（单路径或受支持的批量路径，结果形状与 `import_chat` 一致：单会话带 `sessionId`、批量带 `results`）、`importAll()`（导入发现出的全部待导入会话，返回计数 + 逐条结果 + 扫描 `warnings`）、`resumeSummary()`（Claude Code / Codex 交接摘要，只生成摘要注入当前会话，不创建也不打开新会话）。服务缺席（宿主不提供 `ctx.provide`）时不注册、不抛错；方法接受可选的 `AbortSignal`。类型声明见 `lib/index.d.ts` 的 `ChatImportService`。
- **`/import-all` 支持 `--force`**：已导入 / 归档的会话也另铸副本重导；必须带 `path` 限定范围，不限定范围的全量重导会被拒绝并说明用法。

<h3 id="cn-0.26.0">体验优化</h3>

- **`/import-all` 不再吞掉扫描失败**：某个来源 / 数据根扫描抛错时，命令结果逐来源点名（格式 + 目标 + 错误）；此前 `scan_discover` 与面板都会报告这类失败，只有命令面把它丢掉，读起来像「扫描 0 个会话」。
- **`/import-all` 的逐条失败明细带省略行**：失败超过 5 条时列出前 5 条并注明「另有 N 条未列出」，此前多余条目静默消失；一条也没扫到时空结果返回错误状态，不再以成功文案收场。

<h3 id="cn-0.26.0">其他变更</h3>

- **批量导入编排只留一份**：`/import-all` 的循环（扫描 → 逐会话导入 → registry 批提交 → 聚合计数）此前在命令面与服务面各有一份，已收进 `lib/chat-import-service.mjs` 的 `runChatImportAll`，命令只解析参数并渲染文案（`formatAllImport`）。`--force`、逐条 `results`、扫描 `warnings` 因此在两条入口上同口径。

<h3 id="en-0.26.0">New Features</h3>

- **New `chatImport` host service for third-party plugins**: the plugin entry registers an optional `chatImport` service (through the host's `ctx.provide`) exposing four methods — `scanDiscover()` (read-only discovery on the same scanners and cache as `scan_discover`), `importPath()` (a single path or a supported batch path, with the same result shape as `import_chat`: `sessionId` for a single session, `results` for a batch), `importAll()` (import every pending discovered session, returning counts, per-session results and scan `warnings`), and `resumeSummary()` (Claude Code / Codex handoff summary; it only produces the summary for the current session and neither creates nor opens a new one). The service is skipped without error when the host has no `ctx.provide`, and every method takes an optional `AbortSignal`. Types: `ChatImportService` in `lib/index.d.ts`.
- **`/import-all` gained `--force`**: already-imported / archived sessions are re-imported as fresh copies; it requires a `path` to bound the scope, and an unbounded invocation is rejected with the usage line.

<h3 id="en-0.26.0">Improvements</h3>

- **`/import-all` no longer swallows scan failures**: when a source or data root fails to scan, the command names it (format + target + error). `scan_discover` and the panel already reported these failures; only the command dropped them, so the outcome read like "scanned 0 sessions".
- **`/import-all` per-session failures now carry an elision line**: beyond five failures the first five are listed plus "N more not listed" instead of silently disappearing, and an empty scan reports an error state rather than a success line.

<h3 id="en-0.26.0">Chores</h3>

- **One implementation of the batch orchestration**: the `/import-all` loop (scan → per-session import → registry batch commit → counting) existed once in the command layer and once in the service; it now lives in `runChatImportAll` (`lib/chat-import-service.mjs`) and the command only parses arguments and renders text (`formatAllImport`). `--force`, per-session `results` and scan `warnings` therefore behave identically on both entry points.

**Full Changelog**: [v0.25.2...v0.26.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.25.2...v0.26.0)

## [0.25.2] - 2026-10-06

[中文](#cn-0.25.2) | [English](#en-0.25.2)

<h3 id="cn-0.25.2">体验优化</h3>

- **面板不再把大 DSH 会话列成「(无标题)」**：DSH 来源里压缩后超过 256KB 的会话日志原本整份不解压、按布局目录名兜底，标题恒空且工作区落到 `--D-Build-…--` 这类布局名——同一工作区在面板里被劈成两组，按标题搜索也搜不到。现在这类条目的标题、工作区与创建时间由宿主的会话 header 与持久投影缓存补齐（与宿主侧栏显示的是同一个标题）：本机实测 `dsh4` 来源无标题由 214 条降到 87 条（剩下的只有从未发过消息的空会话与子代理会话），`dsh-claude-style` 由 63 + 43 两组并回 106 一组。来源日志自己读到的标题（含改名）永远优先；宿主未挂投影缓存或该会话没有缓存行时退回原行为，只记一行警告。

<h3 id="cn-0.25.2">问题修复</h3>

- **grokbuild 的 `fullHistory` 开关不再静默失效**：该参数决定是否写原生压缩检查点，但编排把参数指纹硬编码成空表——换值重导时「源未变」短路径照常跳过，开关永远不生效且不报任何变化。现在接通参数指纹：已导入过 `fullHistory: false` 的会话改用 `true` 重导会报 `argsChanged`（需 `force` 才真重导）；从不传该参数的调用方指纹不变，零差异。
- **面板不再认领或连带删除别人的样式表**：宿主客户端模块系统按 `data-plugin` 记账——前端模块物化时把文档里所有无标记的 `<style>` 认领到正在物化的包名下，并在该包换 revision 时整批删除。插件设置导航的遮罩样式表此前没有标记，两个方向都会出事：自己那张被下一个物化的包收走，别人的无标记样式表在插件重载时被连带删掉（实测删掉过 dsh-meme 的两张，表情包界面整份失去样式）。现在该样式表带 `data-plugin` / `data-plugin-css` 标记并按自己的键认回既有元素，无标记的外来样式表被改写到不匹配任何包名的值上（本插件之后出现的那些由常驻观察者照看）。

<h3 id="cn-0.25.2">其他变更</h3>

- 体量治理按 D6 继续：`lib/imports.mjs` 的宿主 `sessionPersistence` ABI 移进 `lib/host-session.mjs`（该文件的既定职责就是「按宿主契约怎么写」），`imports.mjs` 只留 registry 与决策；会话骨架（轮队列 + 待落压缩检查点）从 13 个转换器收进 `lib/convert/ir.mjs` 的 `createTurns()`，落盘序列收成 `writeOneSession()` / `appendTail()`。`lib/` 下最大的手维护文件由 932 行降到 803 行。
- 来源清单三张表（发现层描述符 / `import_chat` 的 format 表 / 导入 spec）互相对账成测试门禁；`fullHistory` 另有「spec 声明 ↔ 实现 ↔ 参数指纹」三面对账，已知缺口（vibe / trae）显式记录、不得被悄悄实现。
- 新增两个机械护栏：`check:tests` 保证每个来源至少一个测试锚点且单测文件不超停止线，`check:coverage` 逐文件与逐目录设下限（全局数字会盖住单文件空洞）；`test/*.test.mjs` 按主题拆分，最大的由 5284 行降到 819 行。
- 测试共用件继续收口（夹具载入口径 / 建库原语 / 会话不变量断言 / 临时目录骨架），删掉零消费者的死代码与已失效的 eslint 全局声明；`docs/USAGE` 双语修正 17 处指向已不存在工具名的调用示例，来源数 27 → 29。

<h3 id="en-0.25.2">Improvements</h3>

- **The panel no longer lists large DSH sessions as "(untitled)"**: a DSH session log above 256KB compressed used to be left undecoded and fall back to its layout directory name — no title, and the workspace name became a layout key like `--D-Build-…--`, splitting one workspace into two panel groups and hiding those sessions from title search. Such entries now take their title, workspace and creation time from the host's session header plus its persisted projection cache (the same title the host sidebar shows): measured on this machine, untitled `dsh4` entries dropped from 214 to 87 (what remains is only never-used empty sessions and subagent sessions), and `dsh-claude-style` merged from 63 + 43 back into a single group of 106. A title the source log itself yields (including a rename) always wins; when the host does not mount the projection cache, or the session has no cache row, the previous behaviour returns with a single warning line.

<h3 id="en-0.25.2">Bug Fixes</h3>

- **grokbuild's `fullHistory` switch no longer fails silently**: the parameter decides whether native compaction checkpoints are written, but the orchestration hard-coded an empty args fingerprint — re-importing with a changed value took the "source unchanged" short path, so the switch never took effect and nothing was reported. The fingerprint is now wired up: re-importing a session previously imported with `fullHistory: false` using `true` reports `argsChanged` (`force` is needed to actually re-import); callers that never pass the parameter have an unchanged fingerprint and see zero difference.
- **The panel no longer claims or deletes other packages' stylesheets**: the host's client module system keeps its books by `data-plugin` — when a front-end module materializes it claims every unmarked `<style>` in the document under the materializing package's name and deletes them in bulk when that package swaps revisions. The plugin's settings-nav scrim stylesheet carried no marker, which broke in both directions: its own sheet was taken over by the next materializing package, and other packages' unmarked sheets were deleted when the plugin reloaded (measured: two of dsh-meme's were removed and its whole UI lost its styling). The sheet now carries `data-plugin` / `data-plugin-css` and reclaims existing elements by its own key instead of adding a second one, while unmarked foreign sheets are rewritten to a value matching no package name (a resident observer handles those appearing after the plugin loads).

<h3 id="en-0.25.2">Chores</h3>

- Size governance continues per D6: the host `sessionPersistence` ABI moved out of `lib/imports.mjs` into `lib/host-session.mjs` (whose stated job is "how to write against the host contract"), leaving `imports.mjs` with the registry and the decisions; the session skeleton (turn queue plus pending compaction checkpoints) moved out of 13 converters into `createTurns()` in `lib/convert/ir.mjs`, and the write sequence collapsed into `writeOneSession()` / `appendTail()`. The largest hand-maintained file under `lib/` drops from 932 to 803 lines.
- The three source tables (discovery descriptors / `import_chat`'s format table / import specs) are now cross-checked by a test gate; `fullHistory` additionally has a three-way "spec declaration ↔ implementation ↔ args fingerprint" check, with known gaps (vibe / trae) recorded explicitly so they cannot be implemented by accident.
- Two new mechanical guards: `check:tests` keeps at least one test anchor per source and holds unit files under the size stop line, and `check:coverage` sets per-file and per-directory floors (a global number hides a single-file hole); `test/*.test.mjs` was split by topic, the largest dropping from 5284 to 819 lines.
- Test helpers keep consolidating (fixture loading, database primitives, session-invariant assertions, temp-directory scaffolding); zero-consumer dead code and stale eslint globals were removed; both `docs/USAGE` languages had 17 call examples pointing at tool names that no longer exist corrected, and the source count went 27 → 29.

**Full Changelog**: [v0.25.1...v0.25.2](https://github.com/Nwflower/dsh-chat-import/compare/v0.25.1...v0.25.2)

## [0.25.1] - 2026-10-05

[中文](#cn-0.25.1) | [English](#en-0.25.1)

<h3 id="cn-0.25.1">问题修复</h3>

- **导入被忽略墓碑拦下时不再只剩一个「跳过」**：服务端把墓碑原因按结构化字段透出（`reason`：`archived` / `retracted` / `workspace-deleted`），批量计数把 `ignored` 从 `skipped` 里拆出来；面板结果行单列「忽略 N（原因）」，工具与命令面文案改为「未导入：该源有忽略记录（原因）」。此前这类导入返回 200、`ok: true`，条目状态 `ignored` 被并进「跳过 N」——一条会话都没建，读起来却像已处理。面板另给一个「仍然导入」的动作出口（一次性 `force` 越权，不解除墓碑；永久解除仍是 `/unignore`），结果行与 Toast 都点名原因。
- **`skipped` 计数口径收窄**：批量结果的 `skipped` 只统计「无可导入内容」，被忽略墓碑挡下的条数计入新的 `ignored`；面板路由摘要另带 `ignoredReasons`（去重后的原因码），面板据此点名原因。

<h3 id="en-0.25.1">Bug Fixes</h3>

- **An import blocked by an ignore tombstone no longer reads as a bare "skipped"**: the server now exposes the tombstone reason as a structured field (`reason`: `archived` / `retracted` / `workspace-deleted`), and batch counting splits `ignored` out of `skipped`; the panel's result line reports "ignored N (reason)" and the tool / command wording became "not imported: this source has an ignore record (reason)". Previously such an import answered 200 with `ok: true` while its `ignored` status was folded into "skipped N" — no session had been created, yet it read as handled. The panel gained an **Import anyway** action (a one-shot `force` that does not clear the tombstone; `/unignore` is still the permanent way out), and both the result line and the toast name the reason.
- **Narrower `skipped`**: in batch results `skipped` now counts only "nothing importable", while tombstone-blocked entries count toward the new `ignored`; the panel route's summary also carries `ignoredReasons` (de-duplicated reason codes) so the panel can name them.

**Full Changelog**: [v0.25.0...v0.25.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.25.0...v0.25.1)

## [0.25.0] - 2026-10-05

[中文](#cn-0.25.0) | [English](#en-0.25.0)

<h3 id="cn-0.25.0">体验优化</h3>

- **扫描失败不再静默**：某个来源 / 数据根扫描抛错时，`scan_discover` 结果新增 `warnings: [{ format, target, error }]`，面板在列表上方点名扫描失败的来源；失败结果不进 30 秒缓存，下次扫描会重试。此前任何异常（含程序错误）都被当成「没有会话」。
- **发现层少做重复 IO**：单次扫描内目录列举与 stat 记忆化，不给格式的目录探测、dsh / dsh4 共用的会话目录不再被各扫描器重复遍历（测试宿主上 Claude 数据根的列举次数由 15 次降为 1 次）；Claude 转录的「只读尾部」改为按偏移只读末尾窗口（此前流过整份文件）；DSH 明文日志只读头尾两段，`.zstd` 经宿主有界读取。
- **书签真正生效**：pi、Cline 旧版任务与 Codex 分页链的扫描书签生效，未变文件不再每次整读（Codex 分页的页级与链级书签此前同键互相覆盖、永不命中）。书签版本升至 4，升级后首次扫描整体重建。
- `node:sqlite` 改为首次真正打开数据库时才加载：插件激活与只用 JSONL 来源的会话不再在宿主日志里打印 SQLite ExperimentalWarning；模块缺席（不满足 engines 的 Node 构建）时大声报错并在扫描 warnings 里点名，不再静默显示为「没有会话」。
- `restore_bundle`：未变的备份包先走 registry 短路径，不再整读、解析并校验指纹。
- 启动时 `ignores.json` 只在内容变化时写盘，不再按每个已归档会话重写并 fsync 一次。
- 面板多选导入时，kilocode / goose / zed / crush / trae 只导所选会话（此前整库导入）；`import_chat` 的 `sessionIds` 说明与 `index.d.ts` 同步列出这些格式。
- `/import` 的结果文案与 `import_chat` 工具同源，补回「仅追加了跳过的行」「回填」两类跳过原因，批量计数带单位（「N 个会话」）。
- 独立 CLI 复用插件的 registry 路径与离线体检，`export-md` 接受 `.zstd` 与 `session.vN` 代次日志。
- 英文界面下，面板请求失败与偏好读取失败的提示不再显示中文。

<h3 id="cn-0.25.0">问题修复</h3>

- **interchange 文档里的工具结果落点**：写在 step `content` 里的 `tool-result` 块现在按 `toolCallId` 派生进 `step.toolResults`（显式列表优先，与 `tool-call` 块的派生对称），正文不再残留宿主 V4 已退休的结果包装——按转换指南写出的这类文档此前被宿主 v4 codec 整份拒载（`... must not contain a released tool-result wrapper`），而预览与全部降级计数都读作 0、不给任何提示；写在 `promptBlocks` 或某个结果内层的结果块无处安放，整块丢弃并计 `skippedBlocks`。存量日志里的残留包装由 `verify_session` 的 `retired-tool-result-wrapper` 点名到事件 seq 与槽位，并给出 `force` 重导的修复提示。
- **图片降级计数**：`imagesDegraded` 不再把转换层计数重复相加、同时丢掉宿主层（附件落地失败 / `storeImages:false`）的计数，两者现在同口径相加。
- **opencode fork 目录导入**：mimocode / teleagent 按目录导入时不再退回 opencode 转换器与「OpenCode · …」标题；mimocode / kilocode / teleagent 的 registry 记录不再把来源格式记成 `opencode`。
- **`/import` 来源名**：别名表由 `import_chat` 的格式表派生，`/import kilocode`、`antigravity`、`dsh4` 不再报「未知来源」。
- **DSH V4 会话**：默认扫描（不给路径）不再漏掉 V4 会话日志；dsh4 条目与 dsh 一样按工作区键回退项目名（此前恒为空）。
- **`storeImages:false` / `DSH_IMPORT_STORE_IMAGES=0` 对所有来源生效**：此前 chatgpt、grokbuild、hermes、kimi、备份包、Codex 分页链、vibe 与全部 SQLite 来源都忽略它；这些来源的 restamp、cwd 补全与预算比对也与标准来源同口径。
- **SQLite 来源重导**：crush / goose / zed / hermes 只有 WAL 边车变化时也能识别为「源已变」；对未变的库传 `replace: true` 会真正覆盖（此前被短路径跳过）。
- **批量计数**：`replace` 覆盖重导计入 imported（grokbuild 与 `restore_bundle` 批量此前记为 skipped）；vibe 批量条目带 `path`。
- **输出 schema**：`import_chat` 声明 `replaced` 状态、落盘结果的 `cwdRemap`、多会话条目的 `staleRegistry`，以及多会话来源摊平到批量顶层的 `validation` 结构校验报告；目录批量的逐文件条目透出 `validation` / `staleRegistry`（此前被条目白名单丢掉）；`restore_bundle` 补齐它实际会返回的字段（`argsChanged` / `budgetChanged` / `images` / 续写与压缩计数等）。宿主按 schema 校验返回值，此前这些情形会让整次调用被判失败。
- **并发与落盘**：purge 修剪 registry 改走串行读-改-写，不再与并发导入互相覆盖；vibe 批量导入等待 registry 提交完成，写盘失败不再成为未处理的 rejection；Codex 分页导入在读页之前先比对文件状态。
- **POSIX 路径**：Claude 的 `~/.claude.json` cwd 映射与 Reasonix 桌面版 `.titles.json` 不再在 Linux / macOS 上拼出反斜杠路径而静默失效。
- **上传文件名 sanitize 跨平台一致**：Windows 风格的路径串（`..\..\evil.jsonl`）在 Linux / macOS 宿主上同样只取最后一段（此前 basename 按运行平台认分隔符，整串残留为 `....evil.jsonl`）。
- **解析错误不再回显内容**：畸形行明细与 generic / 探测层 / 导出校验 / 备份包还原的报错不再带出行内容（含行中段出错的形态）。pi 与 DSH 来源改走共享逐行解析器，畸形行明细与疑似 secret 位置照常上报。
- **图片识别**：OpenAI / Kimi 的 `image_url: { url }` 对象形态被识别（data URL 图片不再降级为占位）；Kimi 旧格式 wire 中拿不到字节的图片计入 `imagesDegraded`。
- generic 文档发射原生压缩检查点时报告 `compacted`；单文件 `db.sqlite` 能识别为 zcode；WorkBuddy 不带 `<user_query>` 的提问去标签后不再把相邻词粘连；Cline / Crush / Goose 带小数毫秒的时间戳取整，带小数的秒级时间戳保留毫秒精度（此前截断到整秒）。
- `export_chat` 的说明不再宣称 claude 导出会写 imports registry；`lib/index.d.ts` 与运行时对齐（格式枚举、导入参数与报告字段、状态值、同步读库函数），并有一致性测试防止再漂移。
- kilocode 的来源标签为「Kilo Code」（导入标题不再以短名打头）。
- 面板：清除导入会话前先清空旧错误提示；「清理上传暂存」确认框执行后关闭；鼠标离开会话行后分组头高亮复位。
- 迁移提示：短会话（如 headless 单轮）在发现跑完前结束时，不再打印「迁移提示注入失败：cannot get required service "systemPrompt" in inactive context」，也不记提示记忆（下一个会话照常提示）。
- 重导 DSH 会话时，无标题事件的会话从首条用户消息派生的标题与其他来源同口径（超 80 字符截断并以「…」收尾，此前是 80 字符硬截）。

<h3 id="cn-0.25.0">其他变更</h3>

- 体量热点按 D6 拆完：`lib/discovery.mjs`（2629 行）拆为 `lib/discovery/` 按来源族的模块并由来源描述符表派生全部清单，`lib/tools.mjs`（1663 行）按工具分组拆为 `lib/tools/`，`lib/import-core.mjs` 的共享状态机拆出 `import-state` / `import-batch` / `host-session`；`lib/` 下最大的手维护文件回到 1000 行以内。
- 测试假宿主的真实磁盘模式不再把测试未显式给定的来源根读成执行机器的真实目录（`import_agents` 用例显式传 `claudeRoot`）——此前装了 Claude Code 技能的机器上 `npm test` 会多出候选而失败。
- 同口径逻辑收口为共享件（D21）：标题 / 时间 / 正文抽取、调用与结果整理、SQLite 只读打开、源未变短路径、批量计数、原子写、测试假宿主各只留一份；面板 React 调用统一为 `h()`，请求统一经 `postJson`。
- 发布面：`files` 补上 `lib/discovery`、`lib/tools`；`npm run build` 从发布入口沿模块图检查可达模块全部在发布集合内。
- CI 只跑一遍测试（覆盖率步骤兼跑）并单独校验 client bundle 新鲜度；`check-leaks` 按 `git ls-files` 扫描受版本管理的文件；`check-linux-compat` 按查找点就近判定分隔符归一；eslint 按生成的跨片全局名单逐片段检查面板源码。

<h3 id="en-0.25.0">Improvements</h3>

- **Scan failures are no longer silent**: when a source or data root throws during a scan, the `scan_discover` result carries `warnings: [{ format, target, error }]` and the panel names the failed sources above the list; failed results are not cached for the 30-second TTL, so the next scan retries. Previously any exception (programming errors included) read as "no sessions".
- **Less repeated IO in discovery**: directory listings and stats are memoized within one scan, so format-less directory probes and the session directory shared by dsh / dsh4 are no longer walked once per scanner (on the test host the Claude data root is listed once instead of 15 times); the "tail only" read for Claude transcripts now reads just the trailing window at an offset (it used to stream the whole file); plain DSH logs read only head and tail, and `.zstd` logs go through bounded host reads.
- **Bookmarks actually hit**: scan bookmarks now apply to pi, legacy Cline tasks and Codex paged chains, so unchanged files are no longer re-read on every scan (Codex page and chain bookmarks used to share a key, overwrite each other and never hit). The bookmark version is bumped to 4; the first scan after upgrading rebuilds them.
- `node:sqlite` is loaded only when a database is actually opened, so plugin activation and sessions that only use JSONL sources no longer print the SQLite ExperimentalWarning in the host log; a Node build without the module (below engines) now fails loudly and names itself in the scan `warnings` instead of silently reading as "no sessions".
- `restore_bundle`: an unchanged bundle takes the registry short path first instead of reading, parsing and fingerprint-checking the whole file.
- At startup `ignores.json` is written only when its content changes, instead of being rewritten and fsynced once per archived session.
- Multi-select imports in the panel now import only the selected sessions for kilocode / goose / zed / crush / trae (they used to import the whole database); the `import_chat` `sessionIds` description and `index.d.ts` list these formats.
- `/import` uses the same result wording as the `import_chat` tool, restoring the "only skipped lines appended" and "backfilled" skip reasons, and batch counts name their unit ("N sessions").
- The standalone CLI reuses the plugin's registry path and offline doctor, and `export-md` accepts `.zstd` and `session.vN` logs.
- In the English UI, the panel's request-failure and preferences-failure messages are no longer shown in Chinese.

<h3 id="en-0.25.0">Bug Fixes</h3>

- **Where tool results belong in interchange documents**: a `tool-result` block written inside a step's `content` is now derived into `step.toolResults` by `toolCallId` (the explicit list wins, symmetric to the `tool-call` derivation), so message content no longer keeps the result wrapper the host V4 codec retired — documents written from the conversion guide used to be rejected wholesale (`... must not contain a released tool-result wrapper`) while the preview and every degradation counter read zero. A `tool-result` block anywhere else (`promptBlocks`, or nested inside another result's `content`) has nowhere to go and is dropped, counted as `skippedBlocks`. Stored logs with a leftover wrapper are named by `verify_session`'s `retired-tool-result-wrapper` check with the event seq and slot, plus a `force` re-import repair hint.
- **Image degradation count**: `imagesDegraded` no longer double-counts the converter's count while dropping the host-side count (attachment save failures / `storeImages:false`); both are now added together.
- **opencode-fork directory imports**: importing a mimocode or teleagent directory no longer falls back to the opencode converter and "OpenCode · …" titles, and mimocode / kilocode / teleagent registry records no longer store `opencode` as the source format.
- **`/import` source names**: the alias table is derived from the `import_chat` format table, so `/import kilocode`, `antigravity` and `dsh4` no longer report "unknown source".
- **DSH V4 sessions**: a default scan (no path) no longer misses V4 session logs, and dsh4 entries fall back to the workspace key for the project name like dsh entries (it was always empty).
- **`storeImages:false` / `DSH_IMPORT_STORE_IMAGES=0` now applies to every source**: chatgpt, grokbuild, hermes, kimi, bundles, Codex paged chains, vibe and all SQLite sources used to ignore it; these sources now also restamp, fill in cwd and compare budgets the same way as the standard sources.
- **SQLite re-imports**: crush / goose / zed / hermes detect a WAL-only change as "source changed", and `replace: true` on an unchanged database really replaces (it used to be skipped by the short path).
- **Batch counts**: a `replace` re-import counts as imported (grokbuild and `restore_bundle` batches used to report it as skipped); vibe batch entries carry `path`.
- **Output schemas**: `import_chat` declares the `replaced` status, `cwdRemap` on saved results, `staleRegistry` on multi-session entries, and the `validation` structural-check report that multi-session sources flatten onto the batch top level; per-file entries of directory batches now pass `validation` / `staleRegistry` through (the entry whitelist used to drop them); `restore_bundle` declares the fields it actually returns (`argsChanged` / `budgetChanged` / `images` / append and compaction counts, …). The host validates results against the schema, so these cases used to fail the whole call.
- **Concurrency and persistence**: purge prunes the registry through the serialized read-modify-write chain instead of racing concurrent imports; vibe batch imports wait for the registry commit, so a failed write is no longer an unhandled rejection; Codex paged imports compare file stats before reading pages.
- **POSIX paths**: Claude's `~/.claude.json` cwd mapping and Reasonix desktop `.titles.json` no longer build backslash paths on Linux / macOS and silently fail.
- **Upload filename sanitization is platform-consistent**: a Windows-style path string (`..\..\evil.jsonl`) reduces to its last segment on Linux / macOS hosts too (`basename` used to honour only the running platform's separators, leaving the whole string as `....evil.jsonl`).
- **Parse errors no longer echo content**: malformed-line details and generic / detection / export-verification / bundle-restore errors no longer carry line content (including mid-line error forms). pi and DSH sources go through the shared line parser, so malformed lines and suspected-secret locations are reported.
- **Image recognition**: OpenAI / Kimi `image_url: { url }` object forms are recognised (data-URL images are no longer degraded to placeholders); images without bytes in legacy Kimi wire tool results count toward `imagesDegraded`.
- generic documents report `compacted` when they emit native compaction checkpoints; a single `db.sqlite` file is recognised as zcode; WorkBuddy prompts without `<user_query>` no longer glue adjacent words when tags are stripped; Cline / Crush / Goose timestamps with fractional milliseconds are rounded, and fractional *seconds* keep their millisecond precision (previously truncated to whole seconds).
- The `export_chat` description no longer claims claude exports write the imports registry; `lib/index.d.ts` matches the runtime (format unions, import parameters and report fields, status values, synchronous database readers), with a consistency test against drift.
- kilocode's source label is "Kilo Code" (imported titles no longer start with the short name).
- Panel: purging imported sessions clears the previous error first; the "clean upload staging" confirmation closes after it runs; the group header highlight resets when the pointer leaves a session row.
- Migration hint: when a short session (e.g. a single headless turn) ends before discovery finishes, the plugin no longer logs "migration hint injection failed: cannot get required service "systemPrompt" in inactive context", and it does not record the hint as shown, so the next session still gets it.
- Re-importing a DSH session derives the title from the first user message (when no title event exists) with the same policy as every other source (truncated past 80 characters with a trailing "…", previously a hard 80-character cut).

<h3 id="en-0.25.0">Chores</h3>

- The size hot spots are split per D6: `lib/discovery.mjs` (2629 lines) into `lib/discovery/` modules per source family with every list derived from one source-descriptor table, `lib/tools.mjs` (1663 lines) into `lib/tools/` per tool group, and the shared state machine of `lib/import-core.mjs` into `import-state` / `import-batch` / `host-session`; the largest hand-maintained file under `lib/` is back under 1000 lines.
- The fake host's real-disk mode no longer lets a test read source roots it did not pass explicitly (the `import_agents` cases now pass `claudeRoot` explicitly) — on a machine with Claude Code skills installed, `npm test` used to find extra candidates and fail.
- Same-semantics logic lives in one shared place (D21): titles / time / text extraction, call-and-result tidying, read-only SQLite access, the unchanged-source short path, batch counting, atomic writes and the test fake host each exist once; panel React calls go through `h()` and requests through `postJson`.
- Publishing: `files` now includes `lib/discovery` and `lib/tools`; `npm run build` walks the module graph from the package entry points and fails if a reachable module would not be published.
- CI runs the test suite once (the coverage step runs it) and checks client-bundle freshness as its own step; `check-leaks` scans tracked files via `git ls-files`; `check-linux-compat` checks separator normalization per lookup site; eslint checks panel fragments with a generated cross-fragment globals list.

**Full Changelog**: [v0.24.1...v0.25.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.24.1...v0.25.0)

## [0.24.1] - 2026-10-04

[中文](#cn-0.24.1) | [English](#en-0.24.1)

<h3 id="cn-0.24.1">新增功能</h3>

- **从文件导入**：面板导入页新增「从文件导入」区，任意本地文件或目录都能进导入管线——**填本机路径回车**（桌面端 host 直读，无大小限制；路径是目录时用**宿主内置预设样式的 Modal** 问「是否搜索子文件夹」，选「仅当前目录」先扫一层、选「包含子文件夹」再递归），或点「**选择…**」用系统文件框挑文件（浏览器只有 `File` 对象时分片上传到暂存）。**不注册拖放**：把会话文件拖到 DSH 窗口会被宿主当作「给当前对话加附件」，面板不与宿主抢这个手势。阶段固定为**识别 → 只读预览 → 导入**：预览卡片给出识别格式与判据、标题、轮 / 消息 / 工具调用计数、cwd、时间与降级计数；目录展开为可勾选条目，「导入所选」逐条导入。命令面同步为 `/import auto <路径>`（`local-jsonl` 同义），工具面 `import_chat({ format: "local-jsonl" })` 的 dry-run 结果现在带 `detectedFormat` / `detectedBy` / `failures`。
- **三级探测与失败出路**：显式格式 > 内容标记（interchange 文档 / `.dshbundle` 便携包，只扫文件头）> 路径特征排序后逐格式试跑。识别失败不再是一句「未识别」——卡片给一句结论「所有对话格式解析器全部解析失败（N 个）」，**失败清单默认收起**（展开后每行「解析器：原因」，等宽两列、可滚动；或一次复制整段摘要），未识别条目不再显示恒为 0 的轮/消息/工具调用计数、也不重复「未能识别」；结论下面一行是出路：把 interchange v1 文档规范当作 **Skill 指令**交给自己的 Agent 转换一次再导入（文案：「请使用 … 作为 Skill 指令，让你的 Agent 将此格式转成受支持的对话文件格式。」）。面板只做自动识别与 DSH 落点（不摆格式 / 目标下拉；覆盖与转投属低频精确控制，留在工具与命令面）。识别不出或 0 轮的文件禁用导入，绝不产出侧栏看不见的空壳会话。
- **interchange v1 文档成为一等导入格式**（`parseFormat: "generic"`，契约见 docs/INTERCHANGE.md §5）：带 `"interchange": "dsh-chat-import"` 与 `version: 1` 的 turns 文档可直接导入，是长尾工具与脚本产出的落点。版本不符整体拒绝；未知内容块、拿不到字节的图片、畸形轮步、孤儿工具结果、非法用量全部计数上报。`.dshbundle.json` 在文件导入里被识别为备份包并转 `restore_bundle`（双层指纹校验），不会被当成普通转录。
- **上传通道**：`POST /api-import/upload/init|chunk|complete` 三步，按 (sha256, size) 幂等——刷新或断线后从已收字节续传，同一文件重复上传零重传；整文件指纹校验通过才产出可导入路径。配额单文件 256MiB、暂存 2GiB、未完成 24 小时回收；文件名 sanitize 且落点固定在 `$DSH_HOME/dsh-chat-import/uploads/`。导入后的暂存件保留（重导语义以它为源键），维护路由可清理未被 registry 引用的件。
- **Agent 转换指南与失败摘要任务模板**：新增双语文档 `docs/INTERCHANGE-GUIDE.md` / `.zh-CN.md`——给 Agent 的 Skill 指令（任务说明、最小合法示例、字段必填表、自检清单、常见错误后果表），`docs/INTERCHANGE.md` 保持维护者视角的协议规范并在头部互链。识别失败的「复制失败摘要」从「路径 + 原因清单」升级为可直接发给 Agent 的任务模板（任务说明 + 指南链接 + 源路径 + 六条产出要求 + 逐解析器失败原因），面板出路链接按界面语言分链到对应语言的转换指南。
- **转换指南注册为宿主 skill**：插件把 `docs/INTERCHANGE-GUIDE.md` 注册为运行时 skill `dsh-chat-import-convert`（`skills` 是可选宿主服务，旧宿主 / headless 缺席时记警告跳过，不影响导入），DSH 内的 Agent 可直接调用它拿到转换指令；识别失败的提示小字改为「请复制失败摘要给 Agent」，任务模板引导优先调用该 skill，GitHub 链接作兜底。
- **新增 Mistral Vibe 会话导入**：新来源 `vibe`（`import_chat({ format: 'vibe' })`、`/import vibe <路径>`、面板来源列表与 `scan_discover`）。读取 Vibe CLI 的 `~/.vibe/logs/session/<session_dir>/`（`messages.jsonl` + `meta.json`，`VIBE_HOME` 可覆盖数据根），会话目录本身即一个会话；用户提问开轮，助手消息成步（`reasoning_content` 进推理块，`tool_calls` 与 `role: 'tool'` 结果按 `tool_call_id` 配对，未配对的结果计数丢弃），内联 base64 图片落为图片块、拿不到字节的降级为占位文本并计数。`context_boundary = "compaction"` 导入为 DSH 原生压缩检查点。两个文件的联合 stat 指纹驱动未变跳过 / 增量续写，`meta.json` 缺失或头部有解析不了的行时标题退回首条用户消息。
- **逐记录时间戳与 provider 用量透传**：源转录里真实存在的逐条时间戳成为会话事件时间（宿主统计据此折出真实的逐步模型耗时与工具耗时；源没有逐记录时间的会话这些耗时为 0），provider 回报的 token 用量（Claude Code 的 `message.usage`、opencode / kilocode / mimocode 的 `message.data.tokens`、Reasonix v2 的行级 `usage` 等）写入 `assistant/message.data.usage` 供宿主 token 统计折叠；事件时间只前进不倒退，`input` / `output` 不是非负整数时整份丢弃，不污染统计。**首 token 延迟与输出速度任何源都导不了**——外部转录只记录完成态消息、没有流块时间戳，不伪造。导出 opencode 方向对称：事件带 usage 时如实回填 `tokens`，没有才写 0 并计入 `usage-unknown`。

<h3 id="cn-0.24.1">问题修复</h3>

- **Qoder 导入不再产出侧栏看不见的空会话**：新版 Qoder 把直连提问写成块数组 content（`[{ type: "text", … }]`），此前转换器只认字符串提问，数组提问被当成工具结果静默丢弃，整段对话 0 轮导入、会话被宿主判为 blank。现在不含 `tool_result` 的块数组按 text 块拼接为提问并开新轮，含 `tool_result` 的数组照旧只做调用配对；旧版字符串提问不受影响。

<h3 id="en-0.24.1">New Features</h3>

- **Import from file**: the import tab gains an "Import from file" area — any local file or directory can enter the import pipeline by **typing a local path and pressing Enter** (the desktop host reads it directly, no size limit; for a directory the host's built-in preset **Modal** asks whether to search subfolders — "this folder only" scans one level, "include subfolders" rescans recursively) or by pressing **Choose…** to pick files with the OS dialog (chunked upload into staging when the browser only holds a `File` object). The panel registers **no drop target**: dropping a session file onto the DSH window is taken by the host as "attach to the current conversation", so it does not fight the host for that gesture. The flow is fixed: **detect → read-only preview → import**. The preview card shows the detected format and criterion, title, turn / message / tool-call counts, cwd, timestamps and degradation counts; a directory expands into checkable entries and "Import selected" imports them one by one. The command surface mirrors it as `/import auto <path>` (`local-jsonl` is the same), and the tool surface's dry-run now returns `detectedFormat` / `detectedBy` / `failures` for `import_chat({ format: "local-jsonl" })`.
- **Three-level detection with a way out**: explicit format > content marker (interchange document / `.dshbundle` backup, sniffed from the file head only) > path hints ordering the parsers that are then tried one by one. A failed detection is no longer a single "unrecognized" line — the card leads with "Every conversation-format parser failed (N)", the **failure list starts collapsed** (expand for a scrollable monospace "parser: reason" table, or copy the whole summary in one go), an unrecognised entry no longer shows all-zero turn/message/tool-call counts nor repeats "unrecognized", and the line below the verdict is the way out: hand the interchange v1 specification to your own agent as a **skill instruction** and let it convert the file once (the panel wording: "Use … as a skill instruction and have your agent convert this format into a supported conversation file format."). The panel itself does auto-detection and DSH landing only (no format / target dropdowns; overriding and transferring are low-frequency precise controls and stay on the tool and command surfaces). Files with no detection or 0 turns disable the import button; a blank session that never shows up in the sidebar is never produced.
- **The interchange v1 document is now a first-class import format** (`parseFormat: "generic"`, contract in docs/INTERCHANGE.md §5): a turns document carrying `"interchange": "dsh-chat-import"` and `version: 1` imports directly, which is the landing spot for long-tail tools and scripted conversions. A mismatched version is rejected wholesale; unknown content blocks, images without usable bytes, malformed turns or steps, orphan tool results and invalid usage are all counted and reported. A `.dshbundle.json` is recognised as a backup archive during file import and routed to `restore_bundle` (two-layer fingerprint check) instead of being treated as a plain transcript.
- **Upload channel**: `POST /api-import/upload/init|chunk|complete` — idempotent per (sha256, size), so a refresh or a dropped connection resumes from the received byte count and re-uploading the same file transfers nothing; a path only becomes importable after the whole-file fingerprint checks out. Limits: 256 MiB per file, 2 GiB of staging, incomplete uploads reclaimed after 24 hours; names are sanitized and staging stays under `$DSH_HOME/dsh-chat-import/uploads/`. Imported staging files are kept (the re-import semantics key on them) and a maintenance route removes the ones no registry entry references.
- **Agent conversion guide and failure-summary task brief**: new bilingual docs `docs/INTERCHANGE-GUIDE.md` / `.zh-CN.md` — skill instructions for agents (task statement, minimal valid document, required-field tables, a pre-delivery checklist, and a mistake→consequence table), while `docs/INTERCHANGE.md` stays the maintainer-facing protocol specification and cross-links to them. On a failed detection, "Copy failure summary" now copies a ready-to-send agent task brief (task statement + guide link + source path + six output requirements + per-parser failure reasons) instead of a bare path-and-reasons list, and the panel's way-out link points at the conversion guide in the UI's language.
- **Conversion guide registered as a host skill**: the plugin registers `docs/INTERCHANGE-GUIDE.md` as the runtime skill `dsh-chat-import-convert` (`skills` is an optional host service — older or headless hosts skip with a logged warning, imports unaffected), so an agent inside DSH can invoke the guide directly; the failed-detection hint now reads "Copy the failure summary to your agent", and the copied task brief points at the skill first with the GitHub link as fallback.
- **Mistral Vibe session import**: new source `vibe` (`import_chat({ format: 'vibe' })`, `/import vibe <path>`, the panel source list and `scan_discover`). It reads the Vibe CLI's `~/.vibe/logs/session/<session_dir>/` (`messages.jsonl` plus `meta.json`; `VIBE_HOME` overrides the data root), where each session directory is one session. User prompts open turns, assistant messages become steps (`reasoning_content` lands as reasoning blocks, `tool_calls` pair with `role: 'tool'` results by `tool_call_id`, unpaired results are dropped and counted), inline base64 images become image blocks while images without usable bytes degrade to a placeholder and are counted. `context_boundary = "compaction"` imports as a native DSH compaction checkpoint. A composite stat fingerprint over both files drives the unchanged-skip / incremental-append paths, and a missing `meta.json` or unparseable head lines fall back to the first user message for the title.
- **Per-record timestamps and provider usage pass through**: timestamps that genuinely exist in the source become session event times (the host folds them into real per-step model and tool durations; sessions whose source has no per-record times report 0 there rather than an invented value), and provider-reported token usage (Claude Code's `message.usage`, opencode / kilocode / mimocode's `message.data.tokens`, Reasonix v2's per-line `usage`, and more) is written to `assistant/message.data.usage` for the host's token statistics; event time only moves forward, and a non-negative-integer check drops the whole usage object rather than polluting the stats. **Time to first token and output speed are unavailable from every source** — external transcripts only record completed messages and carry no stream-chunk timestamps, so nothing is fabricated. The opencode export direction is symmetric: usage is written back as `tokens` when present, otherwise 0 plus a `usage-unknown` degradation.

<h3 id="en-0.24.1">Bug Fixes</h3>

- **Qoder imports no longer produce blank sessions hidden from the sidebar**: recent Qoder versions write direct user prompts as a block-array content (`[{ type: "text", … }]`). The converter only recognised string prompts, so array prompts were silently discarded as tool results, the whole conversation imported with 0 turns and the host treated the session as blank. A block array without `tool_result` now becomes the prompt (text blocks joined) and opens a new turn, an array carrying `tool_result` is still only paired with its calls, and older string prompts are unaffected.

**Full Changelog**: [v0.24.0...v0.24.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.24.0...v0.24.1)

## [0.24.0] - 2026-10-03

[中文](#cn-0.24.0) | [English](#en-0.24.0)

<h3 id="cn-0.24.0">新增功能</h3>

- **新增 Trae Work 会话导入**：新来源 `trae`（`import_trae` / `import_chat({ format: 'trae' })`、面板来源列表、`scan_discover`）。只读打开 Trae 的 VS Code 风格 `state.vscdb`，从 `ItemTable` 的已知会话键（主键 `memento/icube-ai-agent-storage`，另有旧版/变体回退键）提取会话并按会话 id 去重；用户与助手消息成为可继续的 DSH 会话，智能体的计划步骤（思考 / 工具 / 参数 / 结果）落成可读文本，只有工具行的会话不产生空回合。`path` 接受 Trae 的 `User` 根、`workspaceStorage`、`globalStorage` 或单个 `state.vscdb`；默认发现根覆盖 Trae / Trae CN / TRAE SOLO / TRAE SOLO CN 在 Windows、macOS、Linux 上的 `User` 目录。目录模式只展开这两种已知布局、不递归扫描用户目录：从没开过 Trae 对话的工作区库静默跳过，整个目录都没有会话时明确报错，有会话条目却一条都认不出的库计为失败（疑似存储格式变化）。重导沿用一库多会话的增量语义。

<h3 id="cn-0.24.0">其他变更</h3>

- SQLite 库类来源（opencode / mimocode / teleagent / kilocode / zcode / goose / zed / crush / trae）的 dry-run 预览与各自的导入编排同住 `lib/sources/<src>.mjs`，行为不变。
- `test/panel-toast.test.mjs` 读入面板 bundle 后先归一 CRLF：Windows 上 `core.autocrlf=true` 检出时该用例此前恒失败，连带 pre-push 钩子拦下推送。

<h3 id="en-0.24.0">New Features</h3>

- **Trae Work session import**: new source `trae` (`import_trae` / `import_chat({ format: 'trae' })`, the panel source list, and `scan_discover`). Trae's VS Code-style `state.vscdb` is opened read-only; sessions are taken from known `ItemTable` keys (primary `memento/icube-ai-agent-storage`, plus fallback keys for older and variant builds) and de-duplicated by session id. User and assistant messages become a resumable DSH session, the agent's plan steps (thought / tool / arguments / result) are rendered as readable text, and tool-only sessions produce no empty turns. `path` accepts a Trae `User` root, `workspaceStorage`, `globalStorage`, or a single `state.vscdb`; default discovery covers the `User` directories of Trae / Trae CN / TRAE SOLO / TRAE SOLO CN on Windows, macOS and Linux. Directory mode expands only those two known layouts and never walks the user's files: workspace databases that never held a Trae chat are skipped silently, a directory with no sessions at all fails loudly, and a database whose session entries cannot be recognised counts as a failure (a likely storage-format change). Re-imports follow the usual per-database incremental semantics.

<h3 id="en-0.24.0">Chores</h3>

- Dry-run previews for the SQLite-backed sources (opencode / mimocode / teleagent / kilocode / zcode / goose / zed / crush / trae) now live next to their import orchestration in `lib/sources/<src>.mjs`; behavior is unchanged.
- `test/panel-toast.test.mjs` normalizes CRLF after reading the panel bundle: on Windows with `core.autocrlf=true` the case always failed and blocked the pre-push hook.

**Full Changelog**: [v0.23.2...v0.24.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.23.2...v0.24.0)

## [0.23.2] - 2026-10-01

[中文](#cn-0.23.2) | [English](#en-0.23.2)

<h3 id="cn-0.23.2">新增功能</h3>

- **新增 opencode 2.x 的会话导入**：`import_opencode` / `import_chat({ format: 'opencode' })` 现在同时覆盖 opencode 的两个存储世代——V1（`session`/`message`/`part`）与 V2（opencode 2.x 的 `session_v2`/`session_message`，与 V1 共用同一个 `opencode.db`）。读取层按库里的表自动分派：两代产出同一种中间结构，导入、预览与面板发现全部复用；同一会话不会被导入两次，两代都没有的库明确报错，而不是显示成「没有会话」。V2 的压缩边界取最近一条 `completed` 的 `compaction` 行（与源工具自己仍让模型看见的内容起点一致），检查点正文为 `summary` + `recent`；未完成的压缩正文按普通内容保留，不丢。实测（本机 opencode 2.0.21 迁移后的库）：4 个会话在 V1/V2 两侧读出的会话数、消息数与工具调用数一致，压缩会话正确产出 1 个原生压缩检查点。

<h3 id="en-0.23.2">New Features</h3>

- **opencode 2.x session import**: `import_opencode` / `import_chat({ format: 'opencode' })` now cover both of opencode's storage generations — V1 (`session`/`message`/`part`) and V2 (opencode 2.x's `session_v2`/`session_message`, which shares the same `opencode.db`). The reader dispatches on the tables present; both generations produce the same intermediate shape, so import, preview and panel discovery are shared. A session is never imported twice, and a database with neither generation fails loudly instead of looking like "no sessions". V2's compaction boundary is the latest `completed` `compaction` row — the same point the source tool itself keeps visible to the model — with the checkpoint body taken from `summary` + `recent`; unfinished compactions keep their text as ordinary content. Measured on a migrated opencode 2.0.21 database: the same four sessions read from both generations agree on session, message and tool-call counts, and the compacted session produces exactly one native compaction checkpoint.

**Full Changelog**: [v0.23.1...v0.23.2](https://github.com/Nwflower/dsh-chat-import/compare/v0.23.1...v0.23.2)

## [0.23.1] - 2026-10-01

[中文](#cn-0.23.1) | [English](#en-0.23.1)

<h3 id="cn-0.23.1">问题修复</h3>

- **Claude Code 自动压缩后的续跑内容不再丢失**：源工具压缩后会直接接着做原任务、之后可能再没有人类提问，而转换器在压缩摘要处把「当前轮」置空，紧跟其后的 assistant 记录因无轮可挂被整段丢弃——实测一份 18499 行、压缩 6 次的转录丢掉 **2632 / 5962 条** assistant 记录（最后一次压缩后约 2.5 小时的工作全部消失）。现在这些记录归入一个空 prompt 的压缩边界轮，检查点就是该轮的 user 侧消息，与会话正好停在压缩点时的既有处理一致。
- **会话事件时间改为源记录时间戳**：事件 `time` 取最近一个已知的源时间且不倒退（来源没有时间戳时保持 `meta.createdAt` 不变），轮、步、工具结果与压缩检查点的时间都经 IR 传递。此前全部事件都等于 `meta.createdAt`，会话列表的「最后活动时间」因此一直停在创建时间。
- **带上下文压缩的会话能在 0.2.x 宿主上打开**：替换标记的拼写属于**宿主 runtime 世代**、与目标 header 代次无关——`dsh-session` ≥ 0.2.0 的 runtime 与 V3/V4 两代 released codec 都只认 `{ op:'replace', startSeq, endSeq }`，≤ 0.1.x 只认 `{ start, end }`。此前一律写旧拼写，V4 宿主以 `replacement start must be a non-negative safe integer` 拒载整份日志；现在按宿主原生代次双向归一（V4 源日志回灌旧宿主会改回旧拼写），面板显式选 V3 落点同样跟随宿主代次。
- **V4 压缩检查点改用宿主的生产者 kind `compact-checkpoint`**：`shapeMessageSources` 逐条对齐宿主 `producerKind()`（改名表 + 同名生产者表 + system 角色特例）。此前写成 `plugin:compact`，宿主虽能载入，但 `isCompactCheckpointSource` 认不出这是压缩检查点——会话引用投影会直接丢掉摘要，trajectory 视图也不把它折叠成压缩节点。
- **`verify_session` 不再对自己刚导入的 V4 压缩会话误报**：`compaction-checkpoint-source` / `compaction-checkpoint-op` 两项此前只认 V3 形状（`{kind:'plugin',plugin:'compact'}` + `{op,start,end}`），读回 V4 日志（生产者 kind + `startSeq/endSeq`）必然报错；现在 source 与替换端点都按 V3/V4 双形状校验，口径与 `toolResultOf` 一致。
- **DSH → DSH 重导不再把压缩检查点退化成追加**：`convertDshJsonl` 的替换范围重映射只读 `surfaceOp.start/end`，V4 源日志的 `startSeq/endSeq` 找不到端点 → 退化成 `append`（摘要变成普通 user 消息、模型看到全量历史）。现在两种拼写都识别，重排时保持源拼写。

<h3 id="cn-0.23.1">其他变更</h3>

- 本版是 0.22.2 之后第一个发布到 npm 的版本：0.23.0 只在仓库里归版、没有单独发布，它的改动随本版一起到达，内容见下方 0.23.0 节。
- CI：headless 冒烟的 mock LLM 补上 `dsh-llm-deepseek` ≥ 0.1.7 的 Messages wire（`POST {base}/v1/messages` 的 SSE 事件流，事件形状照抄宿主自己的 `tests/mock-server`）——此前只有 OpenAI 兼容的 `/chat/completions`，冒烟自适配器换代起恒红（`DeepSeek Messages request failed (404)`）；`package-lock.json` 里 `@deepseek-ai/dsh-tools` 的 peer 范围与 `package.json` 同步，锁文件漂移护栏恢复绿。

<h3 id="en-0.23.1">Bug Fixes</h3>

- **Claude Code auto-compaction no longer drops the work that continues after it**: after compacting, the source tool keeps working on the same task and may never ask another question, while the converter cleared the "current turn" at the summary record — every assistant record that followed had no turn to attach to and was dropped wholesale. Measured on one 18,499-line transcript with 6 compactions: **2,632 of 5,962** assistant records lost, including roughly 2.5 hours of work after the last compaction. They now form a zero-prompt compaction-boundary turn whose checkpoint is that turn's user-side message, matching the existing handling when a session stops exactly at the boundary.
- **Event times now come from the source records**: an event's `time` follows the latest known source timestamp and never goes backwards (a source without timestamps keeps `meta.createdAt`). Turn, step, tool-result and compaction times are carried through the IR. Every event used to carry `meta.createdAt`, so a session list's "last activity" was stuck at its creation time.
- **Sessions with context compaction open on 0.2.x hosts**: the replacement marker's spelling belongs to the **host runtime generation**, not the target header generation — the `dsh-session` ≥ 0.2.0 runtime and both released V3/V4 codecs accept only `{ op:'replace', startSeq, endSeq }`, while ≤ 0.1.x accepts only `{ start, end }`. Writing the old spelling made a V4 host reject the whole log with `replacement start must be a non-negative safe integer`; the plugin now normalizes in both directions from the host's native generation (a V4 source log written back to an old host gets the old spelling), and an explicit V3 landing follows the host generation too.
- **V4 compaction checkpoints use the host's producer kind `compact-checkpoint`**: `shapeMessageSources` now mirrors the host's `producerKind()` (rename table, same-name table, and the role-sensitive system-prompt case). The old `plugin:compact` was loadable but `isCompactCheckpointSource` did not recognise it, so session-reference projections dropped the summary and the trajectory view did not fold the checkpoint.
- **`verify_session` no longer false-alarms on the plugin's own freshly imported V4 sessions**: `compaction-checkpoint-source` / `compaction-checkpoint-op` only accepted the V3 shapes (`{kind:'plugin',plugin:'compact'}` plus `{op,start,end}`), so a V4 log (producer kind plus `startSeq/endSeq`) always reported a problem; source and replacement endpoints are now read in either shape, the same policy as `toolResultOf`.
- **DSH → DSH re-imports no longer degrade a compaction checkpoint to an append**: `convertDshJsonl`'s replacement-range remap read only `surfaceOp.start/end`, so a V4 source log's `startSeq/endSeq` had no endpoints to map and fell back to `append` (the summary became a plain user message and the model saw the full history). Both spellings are recognised now, and the source spelling is preserved through the remap.

<h3 id="en-0.23.1">Chores</h3>

- This is the first release on npm since 0.22.2: 0.23.0 was versioned in the repository but never published on its own, so its changes arrive with this release — see the 0.23.0 section below.
- CI: the headless smoke's mock LLM now serves the Messages wire of `dsh-llm-deepseek` ≥ 0.1.7 (`POST {base}/v1/messages` SSE, event shapes copied from the host's own `tests/mock-server`); it previously had only the OpenAI-compatible `/chat/completions`, so the smoke had been red since the adapter moved (`DeepSeek Messages request failed (404)`). `package-lock.json` now matches `package.json` for the `@deepseek-ai/dsh-tools` peer range, restoring the lockfile-drift guard.

**Full Changelog**: [v0.22.2...v0.23.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.22.2...v0.23.1)

## [0.23.0] - 2026-09-28

[中文](#cn-0.23.0) | [English](#en-0.23.0)

<h3 id="cn-0.23.0">新增功能</h3>

- 面板「导入历史」新增「**清理空工作区**」按钮（`POST /api-import/workspaces/cleanup`）：移除本插件建过、已无成员的导入工作区登记，幂等；对应用户在侧栏里点不动的空分组。
- **导入完成弹落点提示**：面板发起的导入成功后，用官方 Toast（`shell.overlay`，顶部居中，约 6 秒）报出「导入完成 → <工作区路径>（本次新建该工作区）；另有 N 个未归组」；转投弹写出文件路径。切走 tab / 收起右栏也不会漏掉落点。**面板内原先那条底部黄条（跳过提示 + 「忽略警告」）也并入同一条 Toast**：官方 `Toast` 支持 `actions`，于是变成「已跳过 N 个…」+「重新导入为新会话」按钮（= force 另铸新会话，带动作的横幅停留 15 秒）。宿主没有官方 `Toast` 组件的旧版本上自动退化为同位置、同样带动作按钮的自绘横幅（`require` 包在 `try/catch` 里，见 docs/architecture.md D17）。
- `/attach-workspaces` 改为按**会话自己的 cwd** 重新规划归组，不再回退源目录；cwd 不可用的会话如实报告未归组（header 是 append-only、事后改不了 cwd，要换落点用 `force: true` 重导）。
- **Codex Desktop「导入外部 agent 会话」的 rollout：展平的工具调用还原为真正的工具消息**。Codex Desktop 把外部 agent（实测 Claude Code）的 transcript 写成 rollout 时，工具调用**没有** `function_call` 记录，而是展平成 assistant 正文里的文本信封（`[external_agent_tool_call: Bash]…[/external_agent_tool_call]`、`[external_agent_tool_result[: error]]…`，信封可与正文混排在同一个文本块里）。这类 rollout 此前导入后工具调用全是散文——python 命令里的 `# 注释` 行被 markdown 渲染成巨型标题、Edit 与普通发言无法区分；现在按行扫描切段并还原为 `tool/call` + `tool/result`（结果按出现顺序 FIFO 配对、`: error` 还原为 `isError`）。**还原不虚构**：载荷键名与内容照抄信封——Codex 自己把 `file_path` 写成 `file`、并丢掉 Edit/Write 的正文，本插件不补；未闭合信封 / 认不出的载荷留在正文并计入 `malformed`，找不到调用的结果保留原文并计入 `orphanResults`。导入结果新增 `externalAgent: { calls, results, orphanResults, malformed }`（工具渲染同步）；存量旧形状由 `verify_session` 新增的 `flattened-tool-envelope` 点名（源未变时普通重导会被幂等闸跳过，需 `force: true`）。

<h3 id="cn-0.23.0">问题修复</h3>

- **修「导入成功但对话列表里找不到」**：宿主只接受「会话 cwd 与工作区路径相等」的挂接（`@deepseek-ai/dsh-workspace` 的 `attachSession` 读 header 校验，不匹配即抛错），而插件原先在创建会话之后才归组、失败时回退到**源 transcript 目录**——那条路在宿主上必然被拒，于是每次都在侧栏留下一个**空工作区**、会话仍停在最底部的「未分组」，异常还只写进 `console.error`（本机 `$DSH_HOME/logs` 为空，用户看不到）。现在归组**前置到创建之前**：目标工作区路径先写进 header cwd，创建后再挂接。本机实测同一批三条导入中，两条 cwd = 主目录的会话各留一个空工作区（创建时间与导入时间差 16–22ms），第三条 cwd 命中已有工作区的照常归组——现已一致。
- **落点不再按 transcript 的一面之词造目录**：cwd 是否是可用目录交给宿主 `create` 判定（realpath + isDirectory），插件只为自己拥有的专用导入工作区 `mkdir -p`。cwd 不在本机（跨机器导入）/ 拿不到（含主目录，沙箱 ACL 不允许）时落到专用导入工作区 `$DSH_HOME/dsh-chat-import-workspace` 并改写 cwd，会话因此一定能被看到；源 transcript 目录不再被建成工作区。
- **`workspaceMode` 三态真正生效**：`auto`（默认，按上述规则）/ `per-project`（不改写 cwd，宁可「未分组」也不伪造 cwd）/ `dedicated`（一律落专用工作区，`workspaceDir` 可覆盖目录）。此前 `dedicated` 与源目录回退在宿主上都是失效路径。
- **归组结果进公开结果**：新增 `workspace` / `workspaceMode` / `workspaceCreated`（侧栏多出的分组）与 `ungrouped` + `ungroupedReason`（会话已导入但停在「未分组」）；面板摘要、批量汇总、工具渲染与结果 schema 同步。
- **`restore_bundle` 不再谎报落点**：原先自己算出 `groupedTo` 并声称「回退归组到 bundle 目录」，现改为照抄真实落点；`restoreNote` 给出原 cwd、落点或未归组原因。
- **空工作区清理真正生效**：宿主现只有 `workspaceRegistry.delete(id)`（实体无 `remove`），旧写法恒不命中；改为 `delete(id)` 并覆盖旧实现为源 transcript 目录误建的空工作区，且只删**成员为 0** 的工作区登记（目录与会话日志保留）。

<h3 id="en-0.23.0">New Features</h3>

- The panel's History tab gains a **Clean empty workspaces** button (`POST /api-import/workspaces/cleanup`): it removes import-workspace registrations this plugin created that have no members, and is idempotent — the empty groups in the sidebar that cannot be clicked away.
- **A landing toast now appears when an import finishes**: after a panel-triggered import the official Toast (`shell.overlay`, top-center, ~6 s) reports "Import done → &lt;workspace path&gt; (workspace created); N left ungrouped"; transfers toast the written file path. Switching tabs or collapsing the right sidebar no longer hides where it landed. **The panel's old bottom banner (skip notice + "ignore warning") moved into the same toast**: the official `Toast` accepts `actions`, so it is now "Skipped N …" plus a **Re-import as a new session** button (`force`, minting a new session) on a 15-second hold. On host versions without the official `Toast` component the plugin degrades to a self-drawn banner in the same place, action button included (the `require` is wrapped in `try/catch`; see docs/architecture.md D17).
- `/attach-workspaces` now re-plans grouping from **each session's own cwd** instead of falling back to the source directory; sessions whose cwd is unusable are honestly reported as ungrouped (the header is append-only, so its cwd cannot be changed after the fact — re-import with `force: true` to get a copy under a new landing point).
- **Codex Desktop "import an external agent session" rollouts: flattened tool calls are restored as real tool messages.** When Codex Desktop writes a foreign agent transcript (in practice Claude Code) into a rollout, the tool calls carry **no** `function_call` records — they are flattened into text envelopes inside assistant body text (`[external_agent_tool_call: Bash]…[/external_agent_tool_call]`, `[external_agent_tool_result[: error]]…`, sometimes mixed with prose in the same text block). Such rollouts used to import as pure prose: a `#` comment line inside a python command rendered as a giant markdown heading, and an Edit was indistinguishable from chatter. The converter now scans line by line, splits by segment and restores `tool/call` + `tool/result` (results pair FIFO by order of appearance; `: error` becomes `isError`). **Restoration invents nothing**: payload keys and content are copied from the envelope — Codex itself renamed `file_path` to `file` and dropped the Edit/Write bodies, and the plugin does not fabricate them; unclosed envelopes and unrecognised payloads stay in the body and count as `malformed`, results with no call are kept verbatim and counted as `orphanResults`. The import result gains `externalAgent: { calls, results, orphanResults, malformed }` (wired into tool rendering), and `verify_session` names the stale shape via its new `flattened-tool-envelope` check (an unchanged source is skipped by the idempotence gate, so re-import with `force: true`).

<h3 id="en-0.23.0">Bug Fixes</h3>

- **Fixed "the import succeeded but the conversation is not in the list"**: the host only accepts an attach whose session cwd equals the workspace path (`attachSession` in `@deepseek-ai/dsh-workspace` validates the stored header and throws on a mismatch), while the plugin used to group *after* creating the session and fell back to the **source transcript directory** on failure — a path the host always rejects. The net effect was a fresh **empty workspace** per import plus a session stuck at the very bottom under "Ungrouped", with the exception only reaching `console.error` (this machine's `$DSH_HOME/logs` is empty, so nothing was visible). Grouping now happens **before creation**: the target workspace path is written into the header cwd first, and the session is attached after it exists. Measured on this machine: of three imports in one batch, the two whose cwd was the home directory each left an empty workspace behind (created 16–22 ms after the import), while the third, whose cwd matched an existing workspace, grouped correctly — all three now behave the same.
- **No more directories invented from a transcript's word**: whether a cwd is a usable directory is decided by the host's own `create` (realpath + isDirectory); the plugin only `mkdir -p`s the dedicated import workspace it owns. A cwd that is absent locally (cross-machine import) or unavailable (including the home directory, which the sandbox ACL refuses) lands in the dedicated import workspace `$DSH_HOME/dsh-chat-import-workspace` with the cwd rewritten, so the session is always findable; source transcript directories are no longer turned into workspaces.
- **`workspaceMode` now really has three behaviours**: `auto` (default, rules above) / `per-project` (never rewrite cwd — "Ungrouped" is preferred over a fabricated cwd) / `dedicated` (always the dedicated workspace, `workspaceDir` overrides it). Previously both `dedicated` and the source-directory fallback were dead paths on the host.
- **The grouping outcome is part of the public result**: `workspace` / `workspaceMode` / `workspaceCreated` (the group that appeared in the sidebar) and `ungrouped` + `ungroupedReason` (imported but sitting under "Ungrouped"), wired through the panel summary, batch totals, tool rendering and output schemas.
- **`restore_bundle` no longer reports a landing point that never happened**: it used to derive `groupedTo` itself and claim a fallback "grouped to the bundle directory"; it now reports the real landing point, with `restoreNote` naming the original cwd plus the landing point or the reason it stayed ungrouped.
- **Empty-workspace cleanup actually works**: the host now only exposes `workspaceRegistry.delete(id)` (entities have no `remove`), so the old call never matched; it now deletes by id and also covers the empty workspaces older versions created for source transcript directories — and only workspaces with **zero members** (directories and session logs are kept).

**Full Changelog**: [v0.22.2...v0.23.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.22.2...v0.23.0)

## [0.22.2] - 2026-09-28

[中文](#cn-0.22.2) | [English](#en-0.22.2)

<h3 id="cn-0.22.2">其他变更</h3>

- 本版是 0.19.7 之后第一个发布到 npm 的版本：0.20.0、0.21.0、0.22.0、0.22.1 只在仓库里归版、没有单独发布，它们的改动都随本版一起到达，内容见下方各节。
- README 的导入面板截图换成 DSH Claude Style 的 DeepSeek 配色，工作区与会话均为示例数据；「使用」一节注明用 DSH Claude Style 主题时，「导入会话」按钮收在左下角的账号菜单里。

<h3 id="en-0.22.2">Chores</h3>

- This is the first release on npm since 0.19.7: 0.20.0, 0.21.0, 0.22.0 and 0.22.1 were versioned in the repository but never published on their own, so all of their changes arrive with this release — see the sections below.
- The README's import panel screenshots now use the DeepSeek palette of DSH Claude Style, with sample workspaces and sessions; the Usage section notes that under the DSH Claude Style theme the "Import sessions" button sits in the account menu at the bottom left.

**Full Changelog**: [v0.22.1...v0.22.2](https://github.com/Nwflower/dsh-chat-import/compare/v0.22.1...v0.22.2)

## [0.22.1] - 2026-09-28

[中文](#cn-0.22.1) | [English](#en-0.22.1)

<h3 id="cn-0.22.1">问题修复</h3>

- **与 DSH 0.2.0-rc.1 兼容**：`peerDependencies` 里 `@deepseek-ai/dsh-tools` 原先写死 `^0.1.0-rc.6`（等价于 `<0.2.0`），而宿主按「每个 `@deepseek-ai/dsh*` peer 范围都要匹配运行时版本」的规则检查（见 dsh-app-boot 文档），于是插件在 DSH 0.2.0-rc.1 上被标为不兼容、拒绝导入。范围放宽为 `>=0.1.0-rc.6 <0.3.0`：0.1.x 与 0.2.x 两条 rc 线都覆盖，0.3 起仍会被拦下（未验证的 API 代次不静默放行）。
- 已验证新版宿主 API 面：从宿主 `app.asar` 读取 `@deepseek-ai/dsh-tools@0.2.0-rc.1`，我们仅有的两处依赖（`defineTool` 与 `TOOL_RUNTIME_SCHEDULER`）都在——`TOOL_RUNTIME_SCHEDULER` 仍是 `Symbol('@deepseek-ai/dsh-tools.scheduler')`，`defineTool` 与 0.1.0-rc.8 逐行同构（参数、选项名、校验一致）。加载时的 ABI 守卫报错文案同步改为新范围。

<h3 id="en-0.22.1">Bug Fixes</h3>

- **Compatible with DSH 0.2.0-rc.1**: the `peerDependencies` range for `@deepseek-ai/dsh-tools` was pinned to `^0.1.0-rc.6` (i.e. `<0.2.0`), while the host enforces that every `@deepseek-ai/dsh*` peer range matches its runtime version (see the dsh-app-boot docs) — so the plugin was flagged incompatible on DSH 0.2.0-rc.1 and refused import. The range is now `>=0.1.0-rc.6 <0.3.0`, covering both the 0.1.x and 0.2.x rc lines while still rejecting 0.3+ (unverified API generations are not silently allowed through).
- Verified the new host API surface by reading `@deepseek-ai/dsh-tools@0.2.0-rc.1` out of the host's `app.asar`: both of our only two imports are present — `TOOL_RUNTIME_SCHEDULER` is still `Symbol('@deepseek-ai/dsh-tools.scheduler')`, and `defineTool` is line-for-line identical to 0.1.0-rc.8 (same parameters, option names and validation). The load-time ABI guard message now names the new range.

**Full Changelog**: [v0.22.0...v0.22.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.22.0...v0.22.1)

## [0.22.0] - 2026-09-28

[中文](#cn-0.22.0) | [English](#en-0.22.0)

<h3 id="cn-0.22.0">新增功能</h3>

- **Claude 富结果 sidecar 可选并入**：`import_claude({ includeToolUseResult: true })` 把模型看不到的 `toolUseResult` 渲染成文本追加在工具结果之后——编辑补丁（`structuredPatch` / `bashEditDiff`）→ diff 文本，交互问答（`questions` / `answers`）→ 问答对，其余标量（退出码、耗时、持久化输出路径等）→ 一行紧凑 JSON。默认关（本机 4001 条样本里 1260 条带原文补丁、1115 条带 old/new 字符串，全量并入会明显撑大日志）；体积键（`originalFile` / `content` / `stdout` / `stderr`）与可见正文里已有的内容不重复搬，单值超 2000 字符跳过。开关进参数指纹（换值须 `force` 重导），结果里 `toolUseResultsMerged` 上报实际并入条数。

<h3 id="cn-0.22.0">问题修复</h3>

- **V3 落点不再写出旧宿主读不出的图片块**：附件引用是当前世代（V4）宿主的概念，此前只要宿主提供附件服务就一律落地，「导入到 → DSH（V3）」或续写一条 V3 会话时会产出 V3 宿主打不开的图片块。现在目标代次低于 4 时图片（含已是引用的块）一律降级为 `[image]` 占位并计入 `imagesDegraded`；目标代次只取权威来源（显式覆盖的代次 / 目标会话 header 自己的代次），不用「推断出来的宿主代次」——宿主没有版本信号时推断值会退化成转换层默认的 3。
- **转投（导入到 Claude Code / Codex / Kimi / opencode）的图片口径**：目标能承载图片（Claude / Codex）时按常规路径落附件、导出时读回字节写进目标格式，并在结果里用 `attachmentsOrphaned` 点名「中间会话已撤回、这些字节留在附件存储里无法回收」的张数；目标导不出图片（Kimi 的 wire 只认自有 blob、opencode 的 file part 另需外部文件）时不再落字节，免得白白留下无法回收的数据。
- **导出解引用的能力判定修正**：原本用 `attachmentService()`（要求 `saveImage`）判断能否读附件，而读面只需要 `readImage`——「只能读」的宿主面会被误判成读不了，把有字节的图片降级为占位。改为直接调 `readAttachmentBytes`。
- **面板补上图片计数显示**：`lib/panel.mjs` 已把 `images` / `imagesDegraded` 传回客户端，但客户端结果行不读它们（空载荷）；现补两档文案（中英）。
- **校验器新增 `inline-image-data` 检查**：`validateSessionEvents` 此前不看内容块，「图片块带内联 data（落地步骤被绕过）」这类回归没有任何守卫能发现；现在递归进 tool-result 内层 content（V3 wrapper / V4 一级形状都覆盖）点名。
- **dsh 源同一 turn 多条带图 user/message 时 `promptBlocks` 被覆盖**：改为累积（文本仍取第一条，IR 语义不变）。dsh 源的事件是原样透传，落盘产物不受影响，这是 IR 层（预算估算 / 再合成）的正确性修复。

<h3 id="cn-0.22.0">其他变更</h3>

- 清理 grokbuild 早返回分支里语义已废的 `images: 0` 死字段；`docs/architecture.md` 新增 D15（富结果 sidecar）并补齐 D14 的「V3 目标」「附件不可回收」两条代价；USAGE 双语补「转投与图片」与 `includeToolUseResult` 说明。

<h3 id="en-0.22.0">New Features</h3>

- **Optional Claude tool-result sidecar import**: `import_claude({ includeToolUseResult: true })` renders the model-invisible `toolUseResult` into text appended after the tool result — edit patches (`structuredPatch` / `bashEditDiff`) become diff text, interactive questions/answers become Q&A pairs, and the remaining scalars (exit codes, durations, persisted output paths, …) become one compact JSON line. Off by default (among 4001 sampled records, 1260 carry a source patch and 1115 carry old/new strings; importing all of it grows the log noticeably); bulk keys (`originalFile` / `content` / `stdout` / `stderr`) and anything already present in the visible body are never duplicated, and single values longer than 2000 characters are skipped. The flag is part of the argument fingerprint (changing it requires `force`), and the result reports `toolUseResultsMerged`.

<h3 id="en-0.22.0">Bug Fixes</h3>

- **V3 targets no longer receive image blocks the old host cannot read**: attachment refs are a current-generation (V4) host concept, so previously any host offering the attachment service got them stored — "import to DSH (V3)" or appending to a V3 session produced image blocks a V3 host cannot read. Below generation 4, images (including already-referenced blocks) now degrade to `[image]` placeholders counted in `imagesDegraded`; the target generation is taken only from authoritative sources (an explicit override or the target session's own header) and never from an inferred host version — without a version signal that inference collapses to the converter default of 3.
- **Transfer images (import to Claude Code / Codex / Kimi / opencode)**: when the target can carry images (Claude / Codex) they are stored through the normal path and read back into the target format, with `attachmentsOrphaned` reporting how many bytes the retracted intermediate session leaves behind in the attachment store; when the target cannot express images at all (Kimi's wire only accepts its own blob refs, opencode's file parts need external files) nothing is stored, avoiding data that can never be reclaimed.
- **Export dereference capability check fixed**: it used `attachmentService()` (which requires `saveImage`) to decide whether it could read attachments, while the read path only needs `readImage` — a read-only host surface was misjudged as unable to read, degrading images that had bytes to placeholders. It now calls `readAttachmentBytes` directly.
- **The panel shows the image counters**: `lib/panel.mjs` already returned `images` / `imagesDegraded`, but the client result line ignored them (dead payload); two new labels (zh + en) render them now.
- **New `inline-image-data` validator check**: `validateSessionEvents` never inspected content blocks, so a regression where image blocks keep their inline `data` (the landing step bypassed) had no guard; it now walks tool-result inner content (both the V3 wrapper and the V4 flat shape) and names it.
- **dsh source: `promptBlocks` was overwritten when one turn carried several image-bearing `user/message` records**: they now accumulate (the text still comes from the first, unchanged IR semantics). Since dsh events pass through verbatim, the written product is unaffected — this fixes IR-level correctness (budget estimation and re-synthesis).

<h3 id="en-0.22.0">Chores</h3>

- Removed the dead `images: 0` field from grokbuild's early-return branch; `docs/architecture.md` gained D15 (tool-result sidecar) and D14 now records the V3-target and non-reclaimable-attachment costs; USAGE documents "Transfer and images" plus `includeToolUseResult` in both languages.

**Full Changelog**: [v0.21.0...v0.22.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.21.0...v0.22.0)

## [0.21.0] - 2026-09-28

[中文](#cn-0.21.0) | [English](#en-0.21.0)

<h3 id="cn-0.21.0">新增功能</h3>

- **图片导入为宿主附件，不再一律降级成 `[image]` 文本**：中间结构新增 `image` 块（与宿主 `ContentBlockMap` 的 `text/reasoning/image/tool-call/tool-result` 对齐），落盘前经宿主附件服务（`ctx.attachments`）把字节存成不可变对象，会话日志里只留 `attachmentId` 引用——**base64 永不进日志**。本机真实数据实测：抽样 40 个 Claude 会话落成 297 张（43MB），Codex 25 张，grokbuild 33 张。已是附件引用的会话（`dsh`/`dsh4` 源、导出再导入）原样带过，不重复存。结果新增 `images`（落成附件张数）与 `imagesDegraded`（仍以占位导入的张数）；`storeImages: false` 或环境变量 `DSH_IMPORT_STORE_IMAGES=0` 可只留占位、不写附件存储（单会话上限 500 张，超出部分降级并计数）。
- **反向导出把图片写回目标格式**：`export_claude` 产出 Claude 的 `{type:'image', source:{type:'base64',…}}`，`export_codex` 产出 `{type:'input_image', image_url:'data:…'}`；读不回字节的块以 `[image]` 占位导出并计入 `degradations` 的 `attachment-skipped`。`export-md` 在 Markdown 里标注附件引用与展示名。
- 转换层新增 `lib/convert/image.mjs`（图片来源 → IR 图片块：媒体类型归一、data URL 解析、魔数嗅探、上限保护）与宿主层 `lib/attachments.mjs`（落地 / 解引用 / 读回），`docs/INTERCHANGE.md` 的中间结构一节补齐 `image` / `promptBlocks` / `model` / `aborted` / `compaction` / `shadowed` 字段与两种图片块状态。

<h3 id="cn-0.21.0">问题修复</h3>

- **用户提问里的图片不再静默丢弃**：Claude / Codex / Pi 的提问消息此前只取 `text` 块拼 prompt，同一条消息里的图片连计数都没有（本机 Claude 样本里就有「两张截图 + 一句话」的提问）——现在图片进 `turns[i].promptBlocks`，合成的 `user/message` 同时带文本与图片块。
- **助手消息里的图片与 Kimi 工具结果里的媒体不再静默丢弃**：Claude 助手消息的 `image` 块、Codex 助手消息的图片、Kimi `ToolResult.return_value.output` 里的 `image_url`（`blobref:` 引用，插件取不到字节）此前直接跳过，现在有字节的落成附件、取不到字节的以占位导入并计入 `imagesDegraded`。
- 图片字节不再被复制进日志文本：源转录正文里本来含 base64 文本（如工具结果的 2KB 预览）时，那是源内容；导入产出的图片块只走附件引用，落盘前统一替换并校验。

<h3 id="cn-0.21.0">其他变更</h3>

- `eslint.config.mjs` 声明 `Buffer` 为只读全局（Node 全局此前未登记，新增的 base64 编解码代码因此报 `no-undef`）。
- 导出降级的 `attachment-skipped` 计数口径扩展为「未知块类型 + 读不回字节的图片」（`lib/convert/interchange.mjs` 新增 `unavailableImages` 汇入口）。

<h3 id="en-0.21.0">New Features</h3>

- **Images are imported as host attachments instead of collapsing to `[image]` text**: the intermediate structure gained an `image` block (matching the host's `ContentBlockMap`: `text/reasoning/image/tool-call/tool-result`), and before writing, the host attachment service (`ctx.attachments`) stores the bytes as immutable objects while the session log keeps only an `attachmentId` reference — **base64 never reaches the log**. Measured on this machine's real data: 40 sampled Claude sessions yield 297 stored images (43MB), Codex 25, grokbuild 33. Sessions that already carry attachment refs (`dsh`/`dsh4` sources, re-imported exports) pass through unchanged and are never stored twice. Results gained `images` (blocks stored as attachments) and `imagesDegraded` (blocks still imported as placeholders); `storeImages: false` or `DSH_IMPORT_STORE_IMAGES=0` keeps placeholders only and writes nothing to the attachment store (500 images per session, the excess degrades and is counted).
- **Reverse export writes images back into the target format**: `export_claude` emits Claude's `{type:'image', source:{type:'base64',…}}`, `export_codex` emits `{type:'input_image', image_url:'data:…'}`; blocks whose bytes cannot be read are exported as `[image]` placeholders and counted under `attachment-skipped` in `degradations`. `export-md` marks the attachment reference and display name in Markdown.
- The converter layer gained `lib/convert/image.mjs` (image source → IR image block: media-type normalization, data-URL parsing, magic-byte sniffing, size guard) and the host layer `lib/attachments.mjs` (materialize / dereference / read back); the exchange-format docs now spell out `image` / `promptBlocks` / `model` / `aborted` / `compaction` / `shadowed` and both image-block states.

<h3 id="en-0.21.0">Bug Fixes</h3>

- **Images inside user prompts are no longer dropped silently**: Claude / Codex / Pi prompt messages used to keep only `text` blocks when building the prompt, without even counting the images in the same message (one sampled Claude session opens with two screenshots plus one sentence) — images now go into `turns[i].promptBlocks` and the synthesized `user/message` carries both text and image blocks.
- **Images in assistant messages and media in Kimi tool results are no longer dropped silently**: Claude assistant `image` blocks, Codex assistant images and Kimi's `image_url` entries inside `ToolResult.return_value.output` (a `blobref:` reference the plugin cannot resolve) used to be skipped outright; blocks with bytes now become attachments, and blocks without bytes are imported as placeholders counted in `imagesDegraded`.
- Image bytes are no longer copied into log text: when a source transcript itself contains base64 text (such as a 2KB tool-output preview) that stays source content, while image blocks produced by the import only ever reach the log as attachment references, substituted and verified before writing.

<h3 id="en-0.21.0">Chores</h3>

- `eslint.config.mjs` declares `Buffer` as a read-only global (the Node global was not registered, so the new base64 code flagged `no-undef`).
- The export degradation counter `attachment-skipped` now covers "unknown block types plus images whose bytes cannot be read" (new `unavailableImages` input in `lib/convert/interchange.mjs`).

**Full Changelog**: [v0.20.0...v0.21.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.20.0...v0.21.0)

## [0.20.0] - 2026-09-27

[中文](#cn-0.20.0) | [English](#en-0.20.0)

<h3 id="cn-0.20.0">体验优化</h3>

- **移除双向增量同步，只保留导入与导出**：删掉 `sync_to_claude` 工具、面板「同步」页与 `/api-import/sync` 路由、设置页的同步分区、`sync.json` / `outbound.json` 配置与定时巡检，工具面 13 → 12。反向导出不受影响（`export_chat` 只写新文件，从不改写源转录）。`$DSH_HOME/dsh-chat-import/` 下历史的 `sync.json` / `outbound.json` 以及 imports registry 里的 `writeback` / `exports` 字段成为惰性残留，不再被读取，也不静默删除用户数据。
- **重导语义重做**：不再「记录存在即跳过、源增长一律续写」，改为按「这条 DSH 会话还是不是我上次写完的样子」分流——源未变 → 跳过（不重读）；源增长且未在 DSH 续聊 → 只把新增轮次追加进同一会话（`appended`）；源增长但已在 DSH 续聊 → **另铸新副本**（`reimported.reason: "continued-in-dsh"`），旧会话一字不改；DSH 侧日志比基线短（被外部截短）→ 不写、跳过并报 `storedShrunk`；0.20.0 之前的记录没有基线 → 不可判定，保守另建一次副本并在落盘后回填基线。读不到 DSH 日志长度时既不追加也不复制（跳过并报 `appendedSkipped`）。`force: true` 与显式 `sessionId` 变更仍恒另铸副本。
- **重导产生的副本不再丢账**：另铸副本时旧会话收进 registry 的 `record.copies`，`list_imported_sessions`、`retract_import`、`/attach-workspaces`、清理与 `doctor` 都能枚举到它；删除单条副本只摘掉该条，删除主记录时把最新副本提升为主记录（会话还在，账也还在）。
- 面板与命令：已导入行的按钮与 tooltip 改为重导语义（未导入则新建会话、已续聊则另建副本），跳过 Toast 的动作从「忽略警告」改为「重新导入为新会话」，批量结果摘要单列「重导为新会话 N 个」；工具栏移除「仅选未导入 / 仅选已导入」两个跨页勾选按钮——重导语义下前者的作用与「全选 + 逐行导入」重叠，后者批量重导只会整批空转（源未变）或产出副本（已续聊），按钮名也不描述实际动作；批量拉取新增轮次仍可用 `/import-all`。

<h3 id="cn-0.20.0">问题修复</h3>

- 修复 **重导时把导入轮次追加进用户已续聊的会话**：源文件增长后重导会把新增轮次 append 进那条会话，位置落在用户自己的提问之前，事后无法拆开；而「我就想再导一份」的用户在未变文件上又只有静默跳过一条路。现在已续聊即另铸副本，用户那条会话保持原样；想显式要新副本时随时可用 `force: true`。

<h3 id="cn-0.20.0">其他变更</h3>

- 删除随写回失效的导出层死代码：`tailClaudeEvents` / `serializeClaudeJsonlTail` / `verifyClaudeJsonl` / `serializeCodexJsonlTail` 与 `lib/export/grokbuild.mjs` 整个文件；`./export.mjs` 子路径不再导出这些名字（破坏性契约变化）。
- registry 的展开口径收口到 `lib/imports.mjs` 的 `registryEntries` / `recordEntries`：撤回 / 清理 / 体检 / 工作区挂载 / 面板历史原先各自展开一份（同一逻辑 8 处副本），加副本记账必漏改。
- 输出字段 `forceImported` 更名为 `reimported`（带 `reason`：`continued-in-dsh` / `baseline-missing` / `forced` / `session-id-changed`），新增 `storedShrunk` 与批量 `reimported` 计数；`import_chat` 与 `restore_bundle` 的输出 schema、`lib/index.d.ts` 类型面同步。

<h3 id="en-0.20.0">Improvements</h3>

- **Removed two-way incremental sync; import and export only**: the `sync_to_claude` tool, the panel's Sync page and the `/api-import/sync` route, the sync section in Settings, the `sync.json` / `outbound.json` config and the interval watcher are gone (tool surface 13 → 12). Reverse export is unaffected (`export_chat` only ever writes new files and never rewrites source transcripts). Existing `sync.json` / `outbound.json` files under `$DSH_HOME/dsh-chat-import/` and the registry's `writeback` / `exports` fields become inert leftovers — nothing reads them, and user data is never deleted silently.
- **Re-import policy reworked**: no longer "skip when a record exists, always append when the source grew", but "is this DSH session still exactly what the import wrote?" — unchanged source → skipped (no re-read); source grew and the DSH session was not touched → only the new turns are appended (`appended`); source grew but you already chatted in that DSH session → a **new copy** is created (`reimported.reason: "continued-in-dsh"`) and the old session is left byte-identical; the DSH log is shorter than the baseline (externally truncated) → nothing is written, skipped and reported as `storedShrunk`; records written before 0.20.0 have no baseline → undecidable, so one conservative copy is made and the baseline is backfilled afterwards. When the DSH log length cannot be read at all, neither append nor copy happens (skipped as `appendedSkipped`). `force: true` and an explicit `sessionId` change still always mint a copy.
- **Copies produced by re-imports are no longer lost from the books**: the superseded session lands in the registry's `record.copies`, so `list_imported_sessions`, `retract_import`, `/attach-workspaces`, purge and `doctor` all still enumerate it; deleting a single copy only drops that entry, and deleting the main record promotes the newest copy to main (the session stays, and so does the bookkeeping).
- Panel and commands: the imported row's button and tooltip now describe the re-import policy (new session when unimported, new copy when you already chatted), the skip toast's action is "Re-import as a new session" instead of "Ignore warning", and the batch summary reports "N re-imported as new" separately; the toolbar drops the two cross-page "Select unimported" / "Select imported" buttons — under the re-import policy the former overlaps "select all plus the unimported rows", while the latter only no-ops in bulk (nothing changed) or mints copies for sessions you already chatted in, and its label never described the action; pulling new turns in bulk is still available through `/import-all`.

<h3 id="en-0.20.0">Bug Fixes</h3>

- Fix **re-importing appending imported turns into a session you had already chatted in**: when the source file grew, the new turns were appended to that session, landing *before* your own question and impossible to untangle afterwards; meanwhile a user who simply wanted a second copy only had a silent skip on an unchanged file. A continued session now gets a fresh copy while the session you used stays untouched, and `force: true` remains the explicit way to ask for a copy.

<h3 id="en-0.20.0">Chores</h3>

- Removed the export-layer dead code left behind by the write-back: `tailClaudeEvents` / `serializeClaudeJsonlTail` / `verifyClaudeJsonl` / `serializeCodexJsonlTail` and the whole `lib/export/grokbuild.mjs`; the `./export.mjs` subpath no longer exports those names (breaking contract change).
- Registry flattening is now centralized in `lib/imports.mjs` (`registryEntries` / `recordEntries`): retract / purge / doctor / workspace attach / panel history each used to expand the registry themselves (8 copies of the same logic), which copy tracking would inevitably have missed.
- The `forceImported` result field is renamed to `reimported` (with `reason`: `continued-in-dsh` / `baseline-missing` / `forced` / `session-id-changed`), and `storedShrunk` plus a batch `reimported` counter are added; the `import_chat` and `restore_bundle` output schemas and the `lib/index.d.ts` type surface follow.

**Full Changelog**: [v0.19.7...v0.20.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.7...v0.20.0)

## [0.19.7] - 2026-09-27

[中文](#cn-0.19.7) | [English](#en-0.19.7)

<h3 id="cn-0.19.7">新增功能</h3>

- 新增 **导入结果透出转换层保真 / 降级计数**：`import_chat` 的返回值与渲染正文现在带上 `metaMessages`（isMeta 记录数）、`images`（以 `[image]` 占位导入的图片数）、`backendToolCalls`（Grok Build 后端工具调用数）、`droppedToolResultBlocks`（无法映射的结果块数）、`droppedMalformedOutputs`（Codex 工具输出里的未知块类型数）、`droppedMalformedArgs`（Codex 未能转成标准 JSON 的 `custom_tool_call` 参数数）——六个字段均为可选、>0 才占键，输出 schema（单文件与批量条目）、`SingleImportResult` / `BatchItemResult` 类型同步登记。渲染正文按「失败要大声」逐条列出（如「图片占位 294 张」「无法映射的结果块 1 个」），此前这些计数只停在转换器返回值、工具结果里看不到。

<h3 id="cn-0.19.7">问题修复</h3>

- 修复 **`import_claude` 把绝大多数工具结果导成空**：`tool_result.content` 实测 29494/33254 = **88.7% 是纯字符串**（shell 输出等），转换器此前只映射块数组，字符串内容全部落成空结果（本机 180 份转录实测 **24449 条空结果**）。现在三形态都落地：字符串 → 单文本块；数组逐块映射，其中 `image` 块以 `[image]` 占位并计入新字段 `images`（实测 294 张，base64 永不进日志）；缺失 → 空数组交由合成层兜底。未知块类型计数进新字段 `droppedToolResultBlocks`（不静默）。旧新对比：空结果 24449 → 1（仅剩的 1 条是 `tool_reference` 未知块，已计数上报）。
- 修复 **`import_claude` 把一次响应拆成多个步骤**：Claude Code 按流式增量把一次 API 响应写成多条 assistant 记录（同一 `message.id`，thinking / text / tool_use 各占一行），转换器此前每行各成一步，正文与工具调用被切碎。现在同 id 且中间只隔元数据记录（`mode` / `last-prompt` / `ai-title` 等运行期旁路）的行并回同一步，content 块按到达顺序落在同一步内；间隔里出现 user / assistant 会话记录立即断开——同 id 跨会话记录重复出现实测 1868 对，属另一次真实消息，粘连会跨轮错并。旧新对比：本机 180 份转录步骤数 **56814 → 26858**（合并掉 29956 个碎片步），一次响应只投影一条 `assistant/message`。
- 修复 **`import_codex` 把工具输出导成 JSON 转义串**：`function_call_output.output` 实测 968/1045 = **92.6% 是块数组**（`[{type:'input_text',text…}]`，偶含 `input_image`），转换器此前整体 `JSON.stringify`，导入后是一串转义 JSON、可读性全无。现在块数组逐块映射：文本块拼接为可读文本、`input_image` 以 `[image]` 占位（实测 25 张，data URL / base64 永不进日志）、未知块类型计数进新字段 `droppedMalformedOutputs`；`{"output":[…]}` 信封（字符串或对象）同样处理，纯字符串形态行为不变。旧新对比：转义串结果 **876 → 0**。

<h3 id="en-0.19.7">New Features</h3>

- Add **fidelity / degradation counters to import results**: `import_chat` results and rendered text now carry `metaMessages` (isMeta records), `images` (images imported as `[image]` placeholders), `backendToolCalls` (Grok Build backend tool calls), `droppedToolResultBlocks` (tool-result blocks that could not be mapped), `droppedMalformedOutputs` (unknown Codex tool-output block types) and `droppedMalformedArgs` (Codex `custom_tool_call` arguments that could not be converted to standard JSON). All six are optional and only present when > 0, and are registered in the output schema (single result and batch item) plus the `SingleImportResult` / `BatchItemResult` types. The rendered text lists them ("图片占位 294 张", "无法映射的结果块 1 个", …) so they are loud in the tool result instead of stopping at the converter's return value.

<h3 id="en-0.19.7">Bug Fixes</h3>

- Fix **`import_claude` importing the vast majority of tool results as empty**: 29494 of 33254 (88.7%) `tool_result.content` values in real local transcripts are plain strings (shell output and friends), while the converter only mapped block arrays — every string result landed as an empty result (24449 empty results measured across 180 local transcripts). All three shapes now land: a string becomes a single text block; an array maps block by block, with `image` blocks becoming `[image]` placeholders counted in the new `images` field (294 measured; base64 never enters the log); a missing value stays an empty array for the synthesizer to pad. Unknown block types are counted in the new `droppedToolResultBlocks` field instead of being swallowed. Old vs new on identical data: empty results 24449 → 1 (the remaining one is a `tool_reference` unknown block, counted loudly).
- Fix **`import_claude` splitting one response into multiple steps**: Claude Code writes a single API response as several `assistant` records during streaming (one line each for thinking / text / tool_use, all sharing one `message.id`), and the converter used to turn every line into its own step, shredding the response text and its tool calls. Records with the same id separated only by metadata records (`mode` / `last-prompt` / `ai-title` and other runtime side channels) now merge back into one step with content blocks in arrival order; a `user` or `assistant` conversation record between them still breaks the group — the same id repeating across conversation records (1868 pairs measured) is a genuinely different message, and fusing those would merge separate turns. Old vs new on identical data: 56814 steps → **26858** across 180 local transcripts (29956 fragment steps merged away), one `assistant/message` per response.
- Fix **`import_codex` importing tool output as escaped JSON**: 968 of 1045 (92.6%) `function_call_output.output` values are block arrays (`[{type:'input_text',text…}]`, occasionally with `input_image`), which the converter used to `JSON.stringify` wholesale, so the imported result was unreadable escape soup. Block arrays are now mapped block by block: text blocks join into readable text, `input_image` becomes an `[image]` placeholder (25 measured; data URLs / base64 never enter the log), and unknown block types are counted in the new `droppedMalformedOutputs` field; `{"output":[…]}` envelopes (string or object) get the same treatment, and plain-string outputs behave exactly as before. Old vs new on identical data: escaped-JSON results 876 → 0.

**Full Changelog**: [v0.19.6...v0.19.7](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.6...v0.19.7)

## [0.19.6] - 2026-09-26

[中文](#cn-0.19.6) | [English](#en-0.19.6)

<h3 id="cn-0.19.6">问题修复</h3>

- 修复 **`import_grokbuild` 在真实 `chat_format_version: 1` 转录上丢失几乎全部工具活动**：旧实现把 `tool_result` / `reasoning` / `backend_tool_call` 整类丢弃、assistant 行只读 content 块不读顶层 `tool_calls`——本机 6 个真实会话实测 **473 次工具调用无一进日志**。重写后按真实格式逐类落地：顶层 `tool_calls` 逐项进步骤（实测 473/473 调用与结果全配对，跨 step 晚到的结果按 `tool_call_id` 归位）；`reasoning` 行的明文摘要前置到下一 assistant 步骤（`encrypted_content` 密文永不读取）；`tool_result` 图片以 `[image]` 占位并新增计数字段 `images`（实测 33 张，base64 不进日志）；`backend_tool_call` 只计数（新字段 `backendToolCalls`，实测 1 次）不映射——其结果不在转录里，映射会破坏配对不变量；`synthetic_reason` 注入行（system_reminder / `<user_info>` 环境块）不再被当提问开轮（只含注入的会话正确导入为 0 轮）；`compaction_meta` 交接摘要导入为原生压缩检查点（provider 标签 `grok-build`），`prior_turn_interrupt: 'mid_turn_abort'` 把上一轮标 aborted；逐行 `model_id` 落到 `step.model`，会话级取 `current_model_id` 兜底。发现层同口径修标题兜底（跳过 synthetic 行、剥 `<user_query>` 信封），并新增 `GROK_HOME` 环境变量覆盖默认 `~/.grok` 根。
- 修复 **`import_claude` 把 `isMeta` 记录当提问开轮**：Claude Code 把上下文回执、后台命令输出、图片占位、skill 正文、压缩续接提示等写成 `isMeta: true` 的 user 记录，此前一律开成新轮——本机 179 份转录实测 **198 条**，超一半落在一轮中间把对话切碎成假轮。现在 `isMeta` 记录永不开轮、不参与标题；文本按时间顺序前置到下一个 assistant 步骤 content 开头（搭在 content 块内部、不产生独立消息，不破坏 tool_calls 与 tool 消息的配对），会话末尾残余追加到最后一步，识别数计入新返回字段 `metaMessages`（实测 198/198 与独立统计一致）。
- 修复 **`import_codex` 把 AGENTS.md 注入当首问**：首条 user 消息为 `[# AGENTS.md instructions for …, <environment_context>…]` 的 rollout 此前产出 0 步假轮、标题退化为 `# AGENTS.md instructions for D:\…`（本机 11 份实测全部中招）。现在注入块走公共注入前缀表过滤（与发现层面板同一真相源），全注入消息不开轮、标题取首个真实提问；`importSystemPrompt` 开启时 `# AGENTS.md` 块收进 systemPrompt。
- 修复 **claude / codex / grokbuild 三源的步骤模型归属只记会话第一条**：会话中途换模型后所有步骤仍记成旧模型。现在每条 assistant 步骤带自己的 `step.model`（claude 取 `message.model`、codex 逐条 `turn_context` 更新、grokbuild 取 `model_id`），会话级仍取第一条兜底。

<h3 id="en-0.19.6">Bug Fixes</h3>

- Fix **`import_grokbuild` dropping nearly all tool activity on real `chat_format_version: 1` transcripts**: the old converter discarded `tool_result` / `reasoning` / `backend_tool_call` records wholesale and read only content blocks of assistant lines, never the top-level `tool_calls` — across 6 real local sessions, **all 473 tool calls were missing from the log**. The rewrite maps each record kind faithfully: top-level `tool_calls` become steps (473/473 calls paired with results, late arrivals re-attached by `tool_call_id`); plaintext `reasoning` summaries lead the next assistant step (`encrypted_content` is never read); `tool_result` images become `[image]` placeholders with a new `images` counter (33 measured locally; base64 never enters the log); `backend_tool_call` is only counted (new `backendToolCalls` field, 1 measured) — its results are not in the transcript, so mapping them would break the pairing invariant; `synthetic_reason` injection lines (system reminders / `<user_info>` environment blocks) no longer open turns (injection-only sessions correctly import as 0 turns); `compaction_meta` hand-off summaries import as native compaction checkpoints (provider tag `grok-build`); `prior_turn_interrupt: 'mid_turn_abort'` marks the previous turn aborted; per-line `model_id` lands on `step.model` with `current_model_id` as the session-level fallback. The discovery layer applies the same title fallback (skips synthetic lines, strips `<user_query>` envelopes), and a new `GROK_HOME` environment variable overrides the default `~/.grok` root.
- Fix **`import_claude` opening a turn for every `isMeta` record**: Claude Code writes context receipts, background command output, image placeholders, skill bodies and post-compaction continuation prompts as `isMeta: true` user records, which used to each open a new turn — **198 such records across 179 local transcripts**, more than half landing mid-turn and shredding the conversation into fake turns. `isMeta` records now never open a turn and never become the title; their text is prepended, in arrival order, to the next assistant step's content (inside the content blocks — no standalone message, so tool_calls/tool pairing is undisturbed), leftovers at end of session append to the last step, and the recognized count is reported in the new `metaMessages` field (198/198 matching an independent count locally).
- Fix **`import_codex` treating AGENTS.md injection as the first prompt**: rollouts whose first user message is `[# AGENTS.md instructions for …, <environment_context>…]` used to produce a 0-step fake turn and a title degenerating to `# AGENTS.md instructions for D:\…` (all 11 affected local rollouts). Injection blocks are now filtered through the shared injection prefix table (same source of truth as the discovery panel), an all-injection message opens no turn and the title falls to the first real prompt; with `importSystemPrompt` on, the `# AGENTS.md` block is kept in systemPrompt.
- Fix **claude / codex / grokbuild attributing every step to the session's first model**: after a mid-session model switch all steps still carried the old model. Each assistant step now carries its own `step.model` (claude reads `message.model`, codex tracks every `turn_context`, grokbuild reads `model_id`); the session-level model remains the first record as fallback.

**Full Changelog**: [v0.19.5...v0.19.6](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.5...v0.19.6)

## [0.19.5] - 2026-09-26

[中文](#cn-0.19.5) | [English](#en-0.19.5)

<h3 id="cn-0.19.5">问题修复</h3>

- 修复 **`import_zed` / `import_crush` / `import_continue` / `import_zcode` 的压缩摘要只作 `reasoning` 块、模型上下文与源不一致**：这四源此前把压缩摘要前置成 reasoning 块，却仍把压缩前的全量历史留在模型视角里——源侧（Zed「用摘要替换整段历史」、Crush `summary_message_id`、Continue `conversationSummary`、zcode `compactBoundary` 的保留窗口）早就把那些内容压掉了，于是导入后的会话比源胖一大截，超预算还会被预算裁剪吃掉中间（reported 的 Codex 同类问题）。现在四源默认都发 **DSH 原生压缩检查点**：日志照常保全量历史，边界处发射原生 `compaction/*` 事务（摘要进检查点、其前的轮 log-only），模型视角 = 摘要 + 压缩点之后的内容。边界口径按各源事实：Zed 是 `Compaction` 消息的位置；Crush 是 `is_summary_message` 消息；Continue 是带 `conversationSummary` 的 item 之后；zcode 是 `compactBoundary.keptMessageCount` 划出的保留窗口起点（无 `compactBoundary` 时不猜，退回摘要 reasoning 块）。压缩点之前没有可遮蔽内容时发不出检查点（宿主不变式要求 `shadowedSeqs` 非空），摘要由合成层退回 reasoning 块承载，不丢正文。`fullHistory: true` 时不发检查点、摘要回到既有 reasoning 形态，四源都已进参数指纹（换值重导走 args-changed）。

- 修复 **Cline 的压缩侧车从未被读取、DSH 会话重导丢检查点**：Cline 的 `<sessionId>.compaction.json`（SessionCompactionState）此前既不在收集范围内也不被解析，压缩后的会话导入后模型看到的是全量历史（源侧早已折叠）；读进来后按 `source_message_count` 把被折叠的 canonical 消息标 log-only、摘要进原生检查点，`fullHistory: true` 时不发检查点。同一提交修掉两处相关缺口：**重导 DSH 会话**（`format: 'dsh' | 'dsh4'`）此前 `DURABLE` 白名单不含 `compaction/*`、还会把检查点的替换写成追加，压缩过的 DSH 会话重导一次就丢检查点、摘要退化成普通 user 消息——现在三类压缩事件原样保留、替换范围与 `shadowedSeqs` 与 `sourceEventSeqs` 一起重映射（V3/V4 的 `plugin:compact` 生产者标记双向归一，源日志畸形时退回 append 并由 `verify_session` 的 compaction-* 检查点名）；以及 `import_claude({ fullHistory: true })` 会把 `isCompactSummary` 摘要记录整条丢掉（改动前它是按普通 user 记录导入的）——现在 fullHistory 下摘要记录照常可见。

<h3 id="en-0.19.5">Bug Fixes</h3>

- Fix **`import_kimi` still windowing the conversation at a compaction point**: Kimi's compaction markers (modern `context.apply_compaction`, legacy `CompactionBegin/End` pairs) used to be handled by *windowing* — the turns before the cut, plus the steps already built in the straddling turn, were deleted from the imported session, so the pre-compaction history was gone for good (neither replayable nor exportable back to the source). The modern carrier carries a summary, so it is now imported as a **native DSH compaction checkpoint**: the full history stays in the log, one native `compaction/*` transaction is emitted at the cut (the summary becomes the checkpoint), and the straddling turn is **split in two** (the earlier part is log-only, the later part becomes the boundary turn). The model sees the summary plus everything after the cut, matching the source; several compactions produce several chained checkpoints. The legacy wire carries only a marker pair with the summary written to `context.jsonl` (never into the wire) → there is no summary to restore, so that case keeps the window slice and now reports `compactionSummaryMissing: true` explicitly instead of inventing summary text. `fullHistory: true` emits no checkpoints (the model sees everything) and feeds the args fingerprint (changing it re-imports via args-changed).
- Fix **`import_zed` / `import_crush` / `import_continue` / `import_zcode` keeping the compacted-away history in the model's context** (their summary was only a leading `reasoning` block): the source tools already dropped that content (Zed "replaces the full history with the summary", Crush's `summary_message_id`, Continue's `conversationSummary`, zcode's `compactBoundary` retention window), so the imported session was much fatter than the source and could still have its middle eaten by budget trimming after import (the same defect class as the Codex report). All four now emit **native DSH compaction checkpoints** by default: the log keeps the full history, the boundary emits a native `compaction/*` transaction (summary into the checkpoint, earlier turns log-only), and the model sees the summary plus everything after the boundary. The boundary follows each source's own facts: Zed's `Compaction` message position, Crush's `is_summary_message`, Continue's item carrying `conversationSummary` (boundary after it), and zcode's retained-window start (`compactBoundary.keptMessageCount`; without a `compactBoundary` the converter does not guess and keeps the reasoning block). When nothing before the boundary can be shadowed, no checkpoint can be emitted (the host invariant requires non-empty `shadowedSeqs`), so the synthesizer falls back to carrying the summary as a reasoning block rather than losing it. `fullHistory: true` emits no checkpoints and restores the previous reasoning-block form; all four sources feed it into the args fingerprint (changing it re-imports via args-changed).

- Fix **Cline's compaction sidecar never being read, and DSH session re-imports losing their checkpoints**: `<sessionId>.compaction.json` (the SessionCompactionState) was neither collected nor parsed, so a compacted Cline session imported with the full history visible even though the source had already folded it away; the sidecar is now read and `source_message_count` marks the folded canonical messages log-only with the summary becoming a native checkpoint (`fullHistory: true` emits none). The same commit closes two related gaps: **re-importing a DSH session** (`format: 'dsh' | 'dsh4'`) used to drop every compaction event (`DURABLE` did not list `compaction/*`) and to rewrite the checkpoint's replacement as an append, so one round-trip lost the checkpoints and turned the summary into a plain user message — all three compaction events are now preserved with the replacement range, `shadowedSeqs` and `sourceEventSeqs` remapped together (the V3/V4 `plugin:compact` producer marker is normalized in both directions, a malformed source log falls back to append and gets named by `verify_session`'s compaction-* checks); and `import_claude({ fullHistory: true })` dropped the `isCompactSummary` record entirely (it used to be imported as a plain user record) — the summary record is visible again in fullHistory mode.

**Full Changelog**: [v0.19.4...v0.19.5](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.4...v0.19.5)

## [0.19.4] - 2026-09-26

[中文](#cn-0.19.4) | [English](#en-0.19.4)

<h3 id="cn-0.19.4">新增功能</h3>

- 新增 **上下文压缩导入为 DSH 原生压缩事件**：来源工具的压缩（Claude Code 的 `compact_boundary` + `isCompactSummary` user 记录与旧格式 `summary` 记录、Codex 的 `compacted` 信封、Pi 的 `compaction` 条目、opencode 的 `compaction` part + 摘要消息）此前是「切窗口」——只导压缩点之后的对话，压缩前的历史在导入时就消失了；现在改为**日志保全量历史**，并在每个压缩边界发射一次宿主原生的压缩事务（`compaction/start` → `compaction/summary` → 带 `surfaceOp:{op:'replace'}` 与 `source:{kind:'plugin',plugin:'compact'}` 的检查点 `user/message` → `compaction/end`，被遮蔽的 surface 节点全部进 `sourceEventSeqs` 溯源）。模型的投影因此是「摘要检查点 + 压缩点之后的对话」，与源工具压缩后的真实上下文一致；一次会话压缩多次就发多个链式检查点。导入结果带 `compacted: true` 与 `compactions: <N>`（检查点数），`fullHistory: true` 时不发检查点（模型看到全量历史）并已进参数指纹（换值重导走 args-changed）。摘要正文取不到时**不发检查点**（不做任何标记），绝不静默丢掉前半段。实测（本机真实转录）：Codex 一份 6579 条记录、压缩 10 次的 rollout 此前约 **8411412 tokens**、超 366000 预算被裁掉中间 5 轮（796 条消息、741 次工具调用），现在导入为 19 轮 / 867 条消息 / 794 次工具调用 / 10 个检查点、**零裁剪**；Claude 压缩过的转录（抽检 1 / 1 / 4 次压缩三份）与 Pi、opencode 的合成夹具同样零违规通过宿主 `foldSurface` 与 `dsh-compaction` 不变式。配套：受遮蔽轮**不计预算、不裁剪、不丢弃**（预算只作用于检查点之后的有效段），`verify_session` 新增压缩事务契约校验（括号配对、遮蔽范围与 `shadowedSeqs` 一致、检查点标记与括号 id 一致、溯源不缺被遮蔽节点）。
### 问题修复

- 修复 **`import_claude({ compacted: true })` 在现代转录上从不生效**：该参数只认 2.0.x 旧格式的 `type:"summary"` 记录，而现代 2.x 的压缩载体是 `system/compact_boundary` + 紧随的 `isCompactSummary` user 记录（**摘要正文就在这条 user 记录里**）——本机 105 份主转录里 17 份压缩过，实测 `summary` 记录 0 条、`isCompactSummary` 44 条，于是这个开关静默退化成全量导入。现在两代载体都认（同一会话压缩多次就发多个检查点，见上条），`compacted: true` 保留为兼容别名（默认行为即尊重压缩）。连带修掉一个会随它暴露的缺陷：标题载体（`custom-title` / `ai-title`）落在压缩边界之前时会被切片丢掉，压缩导入的标题会退化成尾部首问——实测 17 份压缩转录**全部**是这个情形，故标题载体改为在全量记录上先扫。
- 修复 **`export_chat({ format: 'claude' })` 把标题写成 `ai-title`（生成标题载体）**：Claude Code 把 `custom-title` 当用户自定义标题、把 `ai-title` 当它随对话自动生成改写的记录——导出成 `ai-title` 等于把 DSH 标题标成「生成标题」，在那边的会话里续聊后可能被它自己生成的新标题覆盖。现在写 `custom-title`（位置不变：首个 user 记录之后），与 `/rename` 的形态一致（实测本机 24 份被重命名过的真实转录正是「只有 custom-title、没有 ai-title」）。导出→导入往返标题逐字保持。
- 修复**粘贴信封 `pasted_content` 进了标题**：`<pasted_content id="…">正文</pasted_content>` 是宿主给「粘贴进来的正文」加的信封，此前整串标记被当成话题（如 `Claude · <pasted_content id="1b70"> …`）。现在标题路径（转换层首问兜底 + 发现层面板标题）剥掉信封只取正文；信封里没有正文时照旧用原文。
- 修复 **Claude 转录的标题归因错误——`/rename` 的自定义标题被丢弃、注入块被当成标题**：转换层与发现层此前只认 2.0.x 旧格式的 `summary` 记录与 `ai-title`，**不认现代 2.x 的 `custom-title` 记录**（`/rename` 写入的自定义标题，本机 2668 条），于是被重命名过的会话导入后标题退回首问原文（常常是 `<system-reminder>` / `<pasted_content>` 注入块）。现在两处都按权威度取载体：`custom-title`（后到者胜 = 最近一次重命名）> 旧格式 `summary` > `ai-title`（取首个：实测 62 份带 ai-title 的转录里 3 份尾部被 worktree 名覆盖）> 首个**非注入**提问；Claude 的发现层额外用尾部读补扫（实测 25 份 custom-title 全部落在文件最后 64KB 内、62 份 ai-title 里 9 份只能从尾部看到），`<local-command-stdout>`（斜杠命令输出回灌）并入注入前缀表。整场只有本地命令的会话落「Claude · 未命名 · 日期」而非注入块，转换层显式给出的空标题也不再被 `pinSourcedSessionTitle` 回退成原始首问。本机实测（104 份有回合的主转录）：26 份标题被纠正；24 份带自定义标题的会话修复前只有 1 份与 Claude Code 自身标题一致，修复后 24/24 一致（超 80 字符按统一规则截断）。`/resume-claude` 的交接摘要标题同源修正。
- 修复**已归档并删除的会话在导入面板显示「同步」而非「导入」**：`resolveImportStatus` 只看 imports registry 记录与归档集，不看宿主里该会话是否还在——会话被删除后 registry 记录仍在、归档集也已清掉，于是被判成「已导入」→ 面板显示「同步」，可会话没了无从同步、也无法重新导入。现在把宿主持久化会话 id 集合（`persistedIds`，发现层本就传入）一并交给状态判定：记录指向的会话**已不在宿主** → `not-imported`（显示「导入」，重导走 `decideItem` 的 `staleRegistry` 重建路径）；**已删除优先于已归档**（归档后又被删除同样是 `not-imported`）。本机实测：96 条「registry 有记录、磁盘上没有会话」的条目从 `imported` 纠正为 `not-imported`。
- 修复 **0.1.7 上设置页不可用**：0.1.7 把插件设置从 `settings.section`（设置页「每功能一页」）搬到「设置 → 插件」的插件页（`plugins.bundle.config` 槽，键 = 包名），值走客户端服务 `configForms`（命名空间 = profile 条目 id，控制器带值 + 写队列 + revision 栅栏）。插件此前只注册旧席位、只走自建 fenced 路由，于是新宿主上设置页看不到、改不动。现在按宿主世代分流（对齐 dsh-claude-style）：`configForms` 在场 → 注册 `plugins.bundle.config`（键 `dsh-chat-import`）并逐字段 `set()` 读写；缺席 → 保留 `settings.section` 整页与 `/api-import/prefs` 路由。两处席位互斥注册，避免同一设置在两处各长一份。（补充：浏览器 boot 给插件条目的是**随机 id**，所以命名空间从 configForms 已服务的名单里挑，挑不到才用 patch 声明的 `import-claude`；官方表单没 ready 或被拒时回落 fenced 路由——这正是「选项显示但存不进去」的修复。）

<h3 id="en-0.19.4">New Features</h3>

- New: **context compaction is imported as native DSH compaction events**. A source tool's compaction (Claude Code's `compact_boundary` + `isCompactSummary` user record and legacy `summary` records, Codex's `compacted` envelopes, Pi's `compaction` entries, opencode's `compaction` parts plus summary message) used to be handled by *windowing* — only the conversation after the compaction point was imported, so the pre-compaction history vanished at import time. The log now keeps the **full history** and one host-native compaction transaction is emitted per boundary (`compaction/start` → `compaction/summary` → the checkpoint `user/message` carrying `surfaceOp:{op:'replace'}` and `source:{kind:'plugin',plugin:'compact'}` → `compaction/end`, with every shadowed surface node listed in `sourceEventSeqs` for provenance). The model therefore sees "summary checkpoint + everything after the boundary", matching the source tool's real post-compaction context; several compactions produce several chained checkpoints. The result carries `compacted: true` and `compactions: <N>` (checkpoint count); `fullHistory: true` emits no checkpoints (the model sees everything) and feeds the args fingerprint (changing it re-imports via args-changed). No summary text → no checkpoint (nothing is marked, the earlier half is never dropped silently). Measured on real transcripts here: a 6579-record Codex rollout compacted 10 times used to measure roughly **8411412 tokens** and have 5 turns (796 messages, 741 tool calls) trimmed away against a 366000 budget — it now imports as 19 turns / 867 messages / 794 tool calls / 10 checkpoints with **no trimming at all**; three sampled compacted Claude transcripts (1 / 1 / 4 compactions) plus the Pi and opencode fixtures pass the host `foldSurface` and `dsh-compaction` invariants with zero violations. Supporting changes: shadowed turns are **not counted, not cropped and not dropped** by budget trimming (the budget only applies to the segment after the checkpoint), and `verify_session` gained a compaction-transaction contract check (bracket pairing, shadow range matching `shadowedSeqs`, checkpoint marker matching the bracket id, provenance covering every shadowed node).
### Bug Fixes

- Fix **`import_claude({ compacted: true })` never taking effect on modern transcripts**: the flag only knew the 2.0.x `type:"summary"` record, while the modern 2.x carrier is `system/compact_boundary` plus the `isCompactSummary` user record that follows it (**the summary text lives in that user record**) — 17 of the 105 main transcripts on this machine were compacted, with 0 `summary` records and 44 `isCompactSummary` records, so the flag silently degraded to a full import. Both carrier generations are now recognized (several compactions emit several checkpoints, see above) and `compacted: true` is kept as a compatible alias (respecting compaction is the default now). This also fixes a defect it would have exposed: title carriers (`custom-title` / `ai-title`) sitting before the compaction boundary were removed by the slice, degrading the title to a tail prompt — all 17 compacted transcripts measured are in exactly that shape, so title carriers are now scanned over the full record set.
- Fix **`export_chat({ format: 'claude' })` writing the title as `ai-title` (the generated-title carrier)**: Claude Code treats `custom-title` as the user's own title and `ai-title` as a record it generates and rewrites as the conversation evolves — exporting as `ai-title` labels the DSH title "generated" and it can be overwritten by a fresh generated title after continuing there. The export now writes `custom-title` (same position: right after the first user record), matching what `/rename` produces (all 24 renamed transcripts measured on this machine carry custom-title and no ai-title). Export → import round-trips the title verbatim.
- Fix **the pasted-content envelope leaking into titles**: `<pasted_content id="…">text</pasted_content>` is the envelope the host wraps pasted text in, and the whole markup used to become the topic (e.g. `Claude · <pasted_content id="1b70"> …`). Title paths (the converter's first-prompt fallback and the discovery layer's panel title) now strip the envelope and keep the text; when the envelope is empty the original text is kept.
- Fix **misattributed Claude transcript titles — `/rename` custom titles dropped, injection blocks used as titles**: the converter and the discovery layer only knew the 2.0.x `summary` record and `ai-title`, and ignored the modern 2.x `custom-title` record (the title written by `/rename`, 2668 of them on this machine), so renamed sessions imported with the first prompt as their title — often a `<system-reminder>` / `<pasted_content>` injection block. Both layers now pick carriers by authority: `custom-title` (last one wins = the most recent rename) > legacy `summary` > `ai-title` (first one wins: in 3 of the 62 transcripts carrying ai-title the tail had been overwritten with a worktree name) > the first **non-injected** prompt; Claude discovery additionally scans the file tail (all 25 custom-title records measured sit in the last 64 KB, and 9 of 62 ai-title records are only visible from the tail), and `<local-command-stdout>` (slash-command output fed back in) joined the injection prefix list. A session holding nothing but local commands now lands on "Claude · 未命名 · date" instead of an injection block, and an explicit empty title from the converter is no longer reverted to the raw first prompt by `pinSourcedSessionTitle`. Measured on this machine (104 main transcripts with turns): 26 titles corrected; of the 24 sessions with a custom title only 1 matched Claude Code's own title before, and 24/24 match now (longer than 80 characters is truncated by the shared rule). `/resume-claude` handoff summaries read the same carriers.
- Fix **archived-then-deleted sessions showing as "Sync" instead of "Import" in the import panel**: `resolveImportStatus` looked only at the imports registry record and the archive set, never at whether the session still exists in the host — after a delete the registry record remained while the archive set was already cleared, so the row was judged "imported" and showed **Sync**, even though a deleted session cannot be synced and could not be re-imported. The host's persisted session id set (`persistedIds`, already passed into discovery) now feeds the status: a record whose session is **no longer in the host** → `not-imported` (the panel shows **Import**, and a re-import takes `decideItem`'s `staleRegistry` rebuild path); **deleted takes precedence over archived** (archived-then-deleted is `not-imported` too). Measured on this machine: 96 entries with "a registry record but no session on disk" were corrected from `imported` to `not-imported`.
- Fix **the settings page being unusable on 0.1.7**: 0.1.7 moved plugin settings from `settings.section` (the Settings page's "one page per feature") to the plugin page under "Settings → Plugins" (the `plugins.bundle.config` slot, keyed by package name), with values served through the client `configForms` service (namespace = profile entry id; the controller carries the value, a write queue and a revision fence). The plugin only registered the old seat and only used its own fenced route, so on the new host the settings were neither visible nor writable. It now splits by host generation (matching dsh-claude-style): with `configForms` present it registers `plugins.bundle.config` (key `dsh-chat-import`) and reads/writes field-by-field through `set()`; without it the `settings.section` full page and the `/api-import/prefs` route are kept. The two seats register exclusively, so one setting never grows a copy in two places. (Note: the browser boot assigns plugin entries a generated id, so the namespace is picked from the namespaces `configForms` already serves, falling back to the patch-declared `import-claude`; when the official form is not ready or refuses, the fenced route takes over — that is the fix for "options show but cannot be saved".)

**Full Changelog**: [v0.19.3...v0.19.4](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.3...v0.19.4)

## [0.19.3] - 2026-09-22

[中文](#cn-0.19.3) | [English](#en-0.19.3)

<h3 id="cn-0.19.3">问题修复</h3>

- 修复 **0.1.7 上设置绑定被过早缓存为 `none`、面板读不到值**（0.19.2 的真机回归）：cordis 在 `apply` 期把 fiber 置为 state 1，`settings.describe()` 会跳过本条目，于是绑定解析成 `none` 并被缓存——apply 结束后即便条目已进名单，所有读取仍停在 `none`（实测面板 `available:false`，而同一响应里的 `probe.namespaceState` 已是 `configured:import-claude`）。现在 `none` 不再缓存（后续读取重新解析），且 `injectTools` 初值改从插件自己的 `config`（volatile 实时引用）读，变更通知同时订阅 `loader/volatile-update`（本 fiber 精确）与 `settings/document-updated`。

<h3 id="en-0.19.3">Bug Fixes</h3>

- Fix **the settings binding being cached as `none` too early on 0.1.7, leaving the panel unable to read values** (a real-host regression in 0.19.2): cordis puts the fiber in state 1 during `apply`, so `settings.describe()` skips the entry and the binding resolves to `none` — which was then cached, so every later read stayed `none` even after the entry appeared (measured: the panel returned `available:false` while the same response's `probe.namespaceState` was already `configured:import-claude`). `none` is no longer cached (later reads re-resolve), the initial `injectTools` value now comes from the plugin's own `config` (a live volatile ref), and change notifications subscribe to both `loader/volatile-update` (exact for this fiber) and `settings/document-updated`.

**Full Changelog**: [v0.19.2...v0.19.3](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.2...v0.19.3)

## [0.19.2] - 2026-09-22

[中文](#cn-0.19.2) | [English](#en-0.19.2)

<h3 id="cn-0.19.2">新增功能</h3>

- 导入面板的设置偏好响应（`/api-import/prefs`）新增只读 **`probe` 自检块**：宿主设置服务的关键事实（`hasSettings` / `hasDescribe` / `hasRegister`、命名空间名单、`namespaceState`）一次看全，供 0.1.5 / 0.1.7 设置模型差异排障。
- 新增 [docs/SETTINGS-MIGRATION.md](docs/SETTINGS-MIGRATION.md)（[中文](docs/SETTINGS-MIGRATION.zh-CN.md)）：DSH 0.1.5 → 0.1.7 设置页迁移说明，含实测报错原文，可直接转发给其他插件作者。

### 问题修复

- 修复 **0.1.7 宿主上导入偏好设置整体失效**：0.1.7 删除了 `settings.register()`（命名空间改为 profile 条目 id、schema 改为插件导出的 `Config`），插件此前仍走旧注册路径，注册直接抛错。现在一套代码兼容两版：宿主设置服务有 `describe()` 且名单含本插件的**裸条目 id** → 按条目 id 读写（0.1.7）；只有 `register` / `get` → 自持命名空间 `chat-import`（0.1.5）；两者皆无 → 读默认、写不持久化（`available:false`）。
- 修复**设置命名空间带种类前缀导致每次保存都 409 `settings-conflict`**：0.1.7 的 loader 把 `ctx.fiber.entry.id` 报成 `<kind>:<id>`，而设置服务按裸 id 建索引；现在剥离前缀（拿不到时回退 patch 声明的 `import-claude`）。
- 修复**插件以软链 / 本地 link 安装时 `@deepseek-ai/schemastery` 解析不到**：解析锚点改为运行中的 harness bin，再退回普通解析；并对 `.volatile()` 做能力探测——旧宿主（无 volatile）不再在模块加载期抛错报废整个插件。
- 插件入口导出 `Config`（三个字段标 `volatile`）并声明 `settings.configure({ auto: false })`（本插件自带面板，不生成宿主自动页）；`package.json` 增补 `icon`（相对路径、包内、约 1 KB、带彩度，适配宿主 `<img>` 渲染拿不到 `currentColor`）。

<h3 id="en-0.19.2">New Features</h3>

- The import panel's settings response (`/api-import/prefs`) gained a read-only **`probe` self-check block**: the host settings service's key facts (`hasSettings` / `hasDescribe` / `hasRegister`, the namespace list, `namespaceState`) in one response, for diagnosing the 0.1.5 / 0.1.7 settings-model difference.
- New [docs/SETTINGS-MIGRATION.md](docs/SETTINGS-MIGRATION.md) ([中文](docs/SETTINGS-MIGRATION.zh-CN.md)): a DSH 0.1.5 → 0.1.7 settings-page migration guide with the verbatim measured errors, ready to forward to other plugin authors.

### Bug Fixes

- Fix **import preferences being entirely broken on a 0.1.7 host**: 0.1.7 removed `settings.register()` (the namespace became the profile entry id and the schema the plugin's exported `Config`), while the plugin still took the old registration path and threw on registration. One codebase now supports both: when the host settings service has `describe()` and its list contains the plugin's **bare entry id** → read/write by entry id (0.1.7); only `register` / `get` → the plugin-owned namespace `chat-import` (0.1.5); neither → defaults are read and writes are not persisted (`available:false`).
- Fix **every save returning 409 `settings-conflict` because the namespace carried a kind prefix**: the 0.1.7 loader reports `ctx.fiber.entry.id` as `<kind>:<id>`, while the settings service indexes by the bare id; the prefix is now stripped (falling back to the patch-declared `import-claude`).
- Fix **`@deepseek-ai/schemastery` not resolving when the plugin is installed as a symlink / local link**: the resolution anchor is now the running harness bin, with ordinary resolution as a fallback; `.volatile()` is capability-probed, so an older host (no volatile) no longer throws at module load and takes the whole plugin down.
- The plugin entry now exports `Config` (all three fields marked `volatile`) and declares `settings.configure({ auto: false })` (this plugin ships its own panel, so no host auto page is generated); `package.json` gained `icon` (relative path, inside the package, ~1 KB, coloured, since the host renders it as `<img>` and never provides `currentColor`).

**Full Changelog**: [v0.19.1...v0.19.2](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.1...v0.19.2)

## [0.19.1] - 2026-09-22

[中文](#cn-0.19.1) | [English](#en-0.19.1)

<h3 id="cn-0.19.1">新增功能</h3>

- 会话读写按**宿主会话格式代次**分流：写入 V3 宿主时工具结果是 `role: 'user'` 内的 `tool-result` 包装，写入 V4 宿主时是顶层 `role: 'tool'` 消息；读取侧（反向导出、校验、Markdown 渲染）统一走同一个工具结果入口，两种形状都能读。已装 V3 宿主的行为与产物不变。
- `verify_session` 新增三类 **V4 迁移风险**检查：`unadvertised-tool-call`（调用无 assistant 内容块广告）、`cross-step-result`（结果闭合在调用所在 step 之外）、`duplicate-tool-result`（同一调用多条结果）。这三类日志在宿主升 V4 时会整份拒载，现在可在升级前用 `force: true` 重导修复。
- 新增**忽略（墓碑）表**：归档会话、撤回 / 删除导入、删除工作区都会自动把对应源登记为忽略，重扫、`/import-all` 与自动同步不再把它带回来；取消归档自动解除。新增命令 **`/ignores`**（查看）、**`/ignore <sessionId|sourcePath>`**、**`/unignore <sessionId|sourcePath|all>`**；`force: true` 仍可越权导入一次（不解除墓碑）。忽略表落盘 `$DSH_HOME/dsh-chat-import/ignores.json`，文件损坏按空表降级，不阻塞导入。
- DSH 来源与「导入到」目标各拆成 **V3 / V4 两项**：来源「DSH V3 会话格式」（日志 v0–v3）与「DSH V4 会话格式」（v4+）共用 `$DSH_HOME/sessions`、按日志文件名代次分流；目标「DSH（V3 会话格式）」「DSH（V4 会话格式）」的**默认项按探测到的宿主版本自动选中**。选定代次后 `header.version` 与事件形状一起产出，因此在 V4 宿主上能真写出一条 `session.v3.jsonl.zstd`。面板 API 的 `format` 与 `import_chat` 的 `format` 因此新增 `dsh4`。
- 来源与目标**都是 DSH 且代次不同**时，面板底部多一个反色的「**导入所选并归档旧会话 (N)**」按钮：按目标代次导入所选，导入成功后把对应的**源会话**在宿主里归档（宿主没有归档 API 时如实回报）。DSH 来源现在也列出导入产物目录（`import-<id>`），因此「把已导入的会话迁移到另一代次」这条路走得通。

### 体验优化

- 导入面板的来源下拉与会话行改用**官方品牌标**（商标归各自权利人）：18 个取自 @lobehub/icons（MIT），Reasonix / Continue / Zed 取自各自 GitHub 仓库，ChatGPT / MimoCode 用 lobehub 的 OpenAI / XiaomiMiMo 标。mark 按实测 ink 包围盒归一化后放进 16px 槽位、字标字面高统一——整列图标一样大、文字都从同一列起；手绘标只剩 3 份（WorkBuddy / TeleAgent / Crush）。
- 多选入口从「点行首来源标」改为「**点整行任意处**」：行首 22px 方图只作来源标识与选中态指示，行内导入 / 同步按钮仍只导入、不勾选。
- 来源列表把 **DSH 两项置顶**（紧随「全部来源」），排在 25 个外部来源之前。
- 幂等跳过不再静默：出现 `already-imported` 条目时面板浮出 Toast「已跳过 N 个可能已导入的对话」，点强调色动作「**忽略警告**」即以 `force: true` 重导这批（另铸新会话）。
- 导入后重扫不再闪空（旧列表留屏，新扫描首批到达时整批替换）；单条导入直接本地补状态、不触发重扫；列表按 `itemKey` 去重，同一会话不会再出现成对的两行、共享 hover。
- 「导入所选并归档旧会话」在窄面板下改用短标签（不再折行），完整标题留在悬停提示里；底部操作区恒贴底。

### 问题修复

- 修复 **0.19.0 引入的 DSH 会话正文丢失**：0.19.0 把 `.zstd` 解码改成 `node:zlib` 原生 zstd，但它**只解第一帧、其余静默丢弃**（截断帧也不报错）——宿主按「一条事件一次 flush」写日志，磁盘上的 `session.vN.jsonl.zstd` 是多帧拼接（本机实测一条 6.5MB 压缩 / 33MB 明文日志有 1962 帧），于是解出来只剩 `{"type":"session",…}` 一行 → 转换 0 轮 → 整份导入按 `skipped` 收场。现改回 `fzstd` 多帧解码（且对截断 / 畸形载荷抛错）；代价是同步解码，本机实测 129KB / 492KB 明文 30ms、6.8MB / 33MB 1.8s。
- 修复**面板 / `/import` / `/import-all` 导入 DSH 来源时「只归档旧会话、不建新会话」**：导入编排漏传转换器的读取钩子，`.zstd` 正文被当二进制读出；现在与工具层同口径透传 `readDshText`。归档也改为**只在导入成功后执行**（归档不可逆、宿主无取消归档面），未归档条数经响应 `archiveSkipped` 如实显示。
- 修复**面板导入后要刷新页面才看得到新会话**：默认目标等于宿主原生代次时，导入此前走只落盘、不进内存会话表的直写路径；现在改走宿主公开的 `agents.create`，新会话即时进列表且与正常会话一样可续聊。
- 修复 **V4 宿主读不回导入日志**：V4 退役 `source.kind = 'plugin'`，导入自产的上下文注入与 system head 写 V4 时按宿主规则改写为生产者自有 kind（`system-prompt` / `plugin:<name>`），V3 保持原样。实测通过宿主自己的 `assertReleasedV4Relationships` / `assertReleasedV4Header` / `assertV4RowAdmission`（13 行 0 失败）。
- 修复 **V3 导入的会话在 V4 宿主上打不开**（`system/message requires a protected first surface head`）：现在在首个 `step/start` 之后写一条空 `system/message` head，位置与形状对齐宿主自己的 v2→v3 迁移器。`verify_session` 新增 `system-head-missing` 点名 0.19.0 之前导入的存量会话，`force: true` 重导即可修复（head 必须是 surface 首事件，旧日志无法原地补写）。
- 导入会话的工具结果统一**闭合在调用所在 step 内**；无广告调用的**孤儿结果**与同一调用的**重复结果**丢弃并计数（`orphanToolResults` / `duplicateToolResults`，零值不占键）。此前这两类日志续聊被模型 API 拒绝，宿主升 V4 也会 fail-closed 拒载。
- 修复 **DSH 来源永远为空**：发现层此前把宿主已有会话一律隐藏，而 DSH 来源的条目本来就是宿主自己的会话日志；现对 DSH 来源不再隐藏。
- 修复 Antigravity 源的 **Windows 绝对工作目录被丢弃**（`C:\…` / `C:/…` 此前不算绝对路径），导入后不再因 header 缺 cwd 而挂载工作区失败。
- 修复**面板「DSH V4 会话格式」报「未知来源」**（面板的来源映射表漏了 `dsh4`）。
- 修复**重导时 registry 指向的会话已不存在**被静默跳过：现在点名 `staleRegistry` 并照常重建，幽灵记录不再静默。
- 修复**读 DSH 源日志时把本插件注入的环境变更声明当成提问 / 标题**（V3 与 V4 两种 source 形状都识别）。
- 修复**宿主没挂 settings 服务时插件 apply 失败**：导入偏好退回默认值，不再在 `settings.register` 上抛错。
- 修复深色主题下导入面板的**弹层透出背后列表**：下拉弹层与页码网格补上宿主菜单的背景模糊与官方投影，sticky 分组头与历史确认框改用不透明底色。

### 其他变更

- 宿主广告**未知的更高会话格式版本（V5+）**时，写盘前大声告警一次（仍按已知最高版本产出）：形状分支是「一版一支」，不假设 `>= 4` 都同形。
- **撤回 / 删除后重导不再自动发生**：`retract_import` 与清理（purge）现在写入永久墓碑，重导同一源返回 `ignored`；恢复用 `/unignore`。归档会话不再另铸后缀副本，归档即写 `archived` 墓碑（取消归档解除）。
- `lib/convert/core.mjs` 按提案拆分（958 → 186 行，回到停止线以内）：校验、形状契约、预算裁剪、事件合成分别拆到 `validate.mjs` / `shape.mjs` / `trim.mjs` / `events.mjs`，对既有 `from './core.mjs'` 引用零影响。

<h3 id="en-0.19.1">New Features</h3>

- Session reads and writes are split by the **host session format generation**: on a V3 host tool results are written as a `tool-result` wrapper inside a `role: 'user'` message, on V4 as a top-level `role: 'tool'` message; the read side (reverse export, verification, Markdown rendering) goes through one tool-result entry point and understands both shapes. An installed V3 host keeps producing exactly what it did before.
- `verify_session` gained three **V4 migration risk** checks — `unadvertised-tool-call` (a call with no advertising assistant content block), `cross-step-result` (a result closed outside its call's step) and `duplicate-tool-result` (more than one result for the same call). A V4 host refuses such logs wholesale; they can now be fixed with a `force: true` re-import before the upgrade.
- **Ignore (tombstone) table**: archiving a session, retracting/purging an import, or removing a workspace now auto-registers the affected sources as ignored, so rescans, `/import-all` and the automatic sync skip them; unarchiving clears the archive tombstone. New commands **`/ignores`**, **`/ignore <sessionId|sourcePath>`** and **`/unignore <sessionId|sourcePath|all>`**; `force: true` still imports once despite a tombstone without clearing it. The table lives at `$DSH_HOME/dsh-chat-import/ignores.json`; a damaged file degrades to an empty table without blocking imports.
- DSH is split into **V3 / V4 entries** on both the source and the "Import to" side: the sources "DSH V3 session format" (logs v0–v3) and "DSH V4 session format" (v4+) share `$DSH_HOME/sessions` and are separated by the generation in the log filename; the targets "DSH (V3 session format)" / "DSH (V4 session format)" **default to the detected host version**. The chosen generation drives both `header.version` and the event shape, so a V4 host can genuinely write a `session.v3.jsonl.zstd`. The panel API's `format` and `import_chat`'s `format` therefore gained `dsh4`.
- When **both the source and the target are DSH and their generations differ**, the bottom bar grows an inverted **"Import selected and archive old sessions (N)"** button: the selected sessions are imported in the target generation and each matching **source session** is archived in the host once its import succeeded (reported honestly when the host exposes no archive API). DSH sources now also list their import-product directories (`import-<id>`), so "migrate an already-imported session to the other generation" works.

### Improvements

- Source brand marks are now the **official** ones (trademarks belong to their owners): 18 from the @lobehub/icons static SVG package (MIT), Reasonix / Continue / Zed from their own GitHub repositories, ChatGPT / MimoCode from lobehub's OpenAI / XiaomiMiMo marks. Each mark's ink box is normalised into a 16px slot and every wordmark shares one cap height, so the column is one size and all text starts at the same x; only three hand-drawn marks remain (WorkBuddy / TeleAgent / Crush).
- Multi-select moved from the leading source mark to **the whole row**: the 22px square only names the source and shows selection state, and the per-row import / sync button still imports without toggling.
- The source list **pins the two DSH entries** right after "All sources", ahead of the 25 external sources.
- Idempotent skips are no longer silent: when `already-imported` entries appear the panel shows a toast "Skipped N possibly-already-imported conversation(s)" with an accent **"Ignore warning"** action that re-imports them with `force: true` (minting new sessions).
- Re-scanning after an import no longer flashes empty (the old list stays until the first batch of the new scan replaces it); a single import patches its row locally instead of triggering a rescan; the list de-duplicates by `itemKey`, so one session can no longer appear as a paired pair of rows sharing a hover.
- "Import selected and archive old sessions" falls back to a short label on a narrow panel (no more wrapping), with the full title in the tooltip; the bottom action bar always stays pinned to the bottom.

### Bug Fixes

- Fix the **DSH conversation body loss introduced in 0.19.0**: 0.19.0 switched `.zstd` decoding to `node:zlib`'s native zstd, which **decodes only the first frame and silently drops the rest** (not even a truncated frame errors). The host writes one frame per flushed event, so a `session.vN.jsonl.zstd` on disk is multi-frame (measured here: a 6.5 MB compressed / 33 MB plaintext log has 1962 frames) and only the `{"type":"session",…}` line survived — zero turns, so the whole import was reported `skipped`. Decoding is back on `fzstd` (multi-frame, and it throws on truncated/malformed payloads); the trade-off is synchronous decoding, measured here at 30 ms for 129 KB / 492 KB plaintext and 1.8 s for 6.8 MB / 33 MB.
- Fix **panel / `/import` / `/import-all` imports of DSH sources archiving the old sessions without creating new ones**: the import orchestration dropped the converter's read hook, so the `.zstd` bytes were read as binary text; it is now forwarded exactly as the tool layer does. Archiving also now happens **only after a successful import** (archiving is irreversible and the host has no unarchive API), and the number left unarchived is reported as `archiveSkipped`.
- Fix **imported sessions requiring a page refresh to appear**: when the default target equals the host's native generation the import used to write directly through a path that only lands bytes on disk and never enters the in-memory session table; it now goes through the host's public `agents.create`, so the new session appears immediately and stays continuable like a native one.
- Fix **V4 hosts refusing to read back imported logs**: V4 retires `source.kind = 'plugin'`, so the import's own context injection and system head are rewritten on the V4 write path to producer-owned kinds (`system-prompt` / `plugin:<name>`), while V3 is untouched. Verified against the host's own `assertReleasedV4Relationships` / `assertReleasedV4Header` / `assertV4RowAdmission` (13 rows, 0 failures).
- Fix **V3-imported sessions that a V4 host refuses to open** (`system/message requires a protected first surface head`): imports now write an empty `system/message` head right after the first `step/start`, matching the host's own v2→v3 migrator. `verify_session` gained a `system-head-missing` check that names sessions imported before 0.19.0; a `force: true` re-import fixes them (the head must be the first surface event, so an existing log cannot be repaired in place).
- Tool results are now closed **inside the step of their call**; **orphan results** (a transcript starting mid-conversation) and **duplicate results** for the same call are dropped and counted (`orphanToolResults` / `duplicateToolResults`, zero values take no key). Such logs used to be refused by the model API on continuation and fail-closed by the host's V4 migration.
- Fix **DSH sources being permanently empty**: discovery used to hide every session the host already had, but DSH sources *are* the host's own session logs; they are no longer hidden.
- Fix Antigravity's **Windows absolute working directory being dropped** (`C:\…` / `C:/…` were not treated as absolute), which made the workspace attach fail after import for lack of a header `cwd`.
- Fix the panel reporting **"unknown source" for "DSH V4 session format"** (the panel's source map was missing `dsh4`).
- Fix a re-import being silently skipped when **the registry's session no longer exists**: it is now reported as `staleRegistry` and rebuilt, so ghost records are no longer silent.
- Fix **reading a DSH source log mistaking the plugin's own environment-change note for a prompt / title** (recognised in both V3 and V4 source shapes).
- Fix the **plugin failing to apply when the host has no settings service**: import preferences fall back to their defaults instead of throwing in `settings.register`.
- Fix the import panel's dropdown **leaking the list behind it** in dark theme: the dropdown and page grid now carry the host menu's blur and official elevation, and the sticky group header / history dialog use an opaque surface.

### Chores

- When the host advertises an **unknown, newer session format version (V5+)**, the write path warns loudly once and emits the highest known shape: the shape branches are one-per-version and it does not assume everything from 4 on is the same.
- **Re-import after retract/delete no longer happens automatically**: `retract_import` and purge now write permanent tombstones and re-importing the same source reports `ignored`; use `/unignore` to lift. Archived sessions are no longer re-importable as a suffixed copy — archiving writes an `archived` tombstone (cleared by unarchiving).
- `lib/convert/core.mjs` was split as proposed (958 → 186 lines, back under the stop line): validation, shape contracts, budget trimming and event synthesis moved to `validate.mjs` / `shape.mjs` / `trim.mjs` / `events.mjs` with zero impact on existing `from './core.mjs'` imports.

**Full Changelog**: [v0.19.0...v0.19.1](https://github.com/Nwflower/dsh-chat-import/compare/v0.19.0...v0.19.1)

## [0.19.0] - 2026-09-21

[中文](#cn-0.19.0) | [English](#en-0.19.0)

<h3 id="cn-0.19.0">新增功能</h3>

- 会话发现支持「全部来源」流式加载：扫描结果按发现顺序逐条推入列表，首屏不再等全量扫描结束；扫描完成后一次性重排为最近活跃倒序。
- 导入面板新增**筛选：路径**（原工作区筛选，标签化）与**筛选：时间**（24 小时 / 7 天 / 30 天 / 不筛选，按最后活跃或创建时间过滤）。
- 分页档位改为 **500 / 2000 / 全部**（默认 500）：「全部」即不分页，10 万行实测 DOM 恒为 19 行 / 351 节点、悬停与勾选 0.6ms。
- 页码改成「第 x / y 页」控件，点开在底栏上方弹出**页码网格**（点数字直接跳页，页多时网格自滚动并停在当前页附近）。

### 体验优化

- 会话列表改为**窗口化渲染**：行高 28px + 行距 1px 与组头 34px 都是固定值，可见区间纯算术得出，只挂载可视区上下各一屏，其余用等高占位块撑住（滚动高度与 sticky 组头行为不变）。
- 会话行抽成 memo 组件、回调走 ref 转存、派生数据全部 `useMemo`：悬停或勾选一次只重建受影响的那一行；大档位下两处「每次渲染 O(n)」开销收敛，10 万行时一次悬停从 8ms 降到 0.6ms。
- 会话列表改成一行式，与皮肤的工作区列表同一套口径（行高 / 圆角 / 标题字号与配色一致）：行首来源工具标既是来源标识也是多选勾选位，右侧相对时间；上下文 / 分支 / 导入状态收进悬停提示，单条导入按钮悬停时才出现在时间位置。
- 扫描状态与分页条合并成列表下方一条：扫描中显示「已发现 N 个」，完成后显示页码与总数；总数不足一档时连「每页」选择器一起隐藏。
- 工具栏动作按钮的折叠判据改为「动作按钮组实测可用宽度」而不是面板宽度：用隐藏探针量出文字形态所需宽度再比对，文字形态不再折行（挤不下走省略号兜底）。
- 列表窗口化的上下余量从「±10 行」提高到「±一屏」，快滚不再露白。
- **发现层不再为消息条数整读**（面板已不展示该字段）：SQLite 源（opencode 系 / zcode / hermes）改走会话摘要读取器，只查 session 表加每会话一条「最近消息时间」聚合，不再逐会话读出 message/part 正文并逐 cell 解析——本机 zcode 单次同步阻塞 185ms → 1ms，正是面板卡顿的来源之一。
- DSH 会话的 `.zstd` 正文改走 **node:zlib 原生异步 zstd 解码**（libuv 线程池，Node < 22.15 自动回退 fzstd）：本机 60 个会话实测同步阻塞 3.8s → 异步 0.2s，事件循环不再被顶住。
- 发现层尾部读取（claude / kimi 的 context token）改为 chunks 数组滚动窗口：原实现每块都对整条尾串全量复制，大 transcript 的尾部读取开销主要在这块 memcpy。
- 本机实测（含 26 种来源、约 1.35 GB 数据）：冷扫描 5002ms → 1438ms，事件循环最大漂移 104ms → 0ms，书签命中重扫 375ms → 150ms。

### 问题修复

- 修复窄面板下工具栏按钮被压扁、文字折行的问题（折叠判据改用实测宽度，见上）。
- 修复快滚列表时偶发露白：窗口化余量由固定 10 行改为按视口高度计算。

### 其他变更

- 客户端 bundle 构建脚本改为**原子写** `lib/client.js`（同目录临时文件 + rename）：宿主按 stat 轮询该文件、一变就重新加载并按内容哈希发版（带一年 immutable 缓存），直接覆写会留出「读到半截 bundle」的窗口。另加 `--out=<path>` 供测量/实验构建写到 `lib/` 之外。
- `scan_discover` 输出条目与 schema 去掉 `messageCount`（面板已不展示；SQLite 摘要读取器同步不再产出该字段）。
- README 增补通过 GUI 导入的界面预览（亮 / 暗各一张），并新增 `screenshots.json` 商店截图清单。
- `package.json` 的 `files` 增补 `docs/*.png`，让 npm 页面上的 README 也能显示预览图。

<h3 id="en-0.19.0">New Features</h3>

- Streaming discovery for "All sources": scan results are appended in discovery order so the first screen no longer waits for the full scan; once the scan finishes the list is re-sorted by most-recent activity.
- Add **Filter: path** (the former workspace filter, now label-style) and **Filter: time** (24 hours / 7 days / 30 days / any time, by last activity or creation) to the import panel.
- Page sizes are now **500 / 2000 / All** (500 by default): "All" drops pagination — with 100k rows the DOM stays at 19 rows / 351 nodes, with 0.6ms hover and selection.
- The page number becomes a "Page x / y" control that opens a **page grid** above the status bar for one-click jumps; the grid scrolls on its own when there are many pages.

### Improvements

- The session list is now **windowed**: fixed row (28px + 1px gap) and group-header (34px) heights make the visible range pure arithmetic, so only the rows within one screen above and below the viewport are mounted and the rest are held by equal-height spacers (scroll height and sticky group headers unchanged).
- Session rows became memo components with ref-stashed callbacks and fully memoised derived data: a hover or a checkbox toggle rebuilds only the affected row; the two per-render O(n) costs on large tiers were removed, cutting a hover from 8ms to 0.6ms at 100k rows.
- Session rows became single-line and now follow the same metrics as the skin's workspace list (matching height, radius, title size and colours): the leading source mark is both the source label and the multi-select checkbox, the relative time sits on the right, and context / branch / import status moved into the hover tooltip, with the per-row import button appearing in the timestamp's place on hover.
- Scan progress and pagination merged into one status line under the list: it reports "N found" while scanning and page / total once done; below one full page the per-page selector is hidden too.
- Toolbar action buttons now collapse based on the **measured width available to the button group** instead of the panel width: a hidden probe measures the width the text form needs, so labels no longer wrap (they fall back to an ellipsis when truly out of room).
- The windowing overscan grew from "±10 rows" to "±one screen", so fast scrolling no longer flashes blank rows.
- **Discovery no longer reads whole SQLite transcripts just to count messages** (the panel no longer shows that field): opencode-family / zcode / hermes sources now use per-session summary readers that only query the session table plus one "latest message time" aggregate, instead of loading every message and part and parsing each cell — on this machine zcode's single blocking read dropped from 185ms to 1ms, one of the causes of panel jank.
- DSH `.zstd` session bodies now decode through **node:zlib's native async zstd** (libuv thread pool, with an automatic fzstd fallback on Node < 22.15): measured on this machine, 60 sessions went from 3.8s of synchronous blocking to 0.2s of async work, so the event loop is no longer stalled.
- Tail reads in discovery (claude / kimi context tokens) now use a chunk-array rolling window: the previous implementation copied the entire accumulated tail on every chunk, and that memcpy was the bulk of the cost on large transcripts.
- Measured on this machine (26 sources, ~1.35 GB of data): cold scan 5002ms → 1438ms, worst event-loop drift 104ms → 0ms, bookmark-hit rescan 375ms → 150ms.

### Bug Fixes

- Fix toolbar buttons being squeezed and their labels wrapping on a narrow panel (the collapse rule now uses a measured width, see above).
- Fix occasional blank rows while fast-scrolling: the windowing overscan is now computed from the viewport height instead of a fixed 10 rows.

### Chores

- The client bundle build script now writes `lib/client.js` **atomically** (same-directory temp file + rename): the host polls that file by stat, reloads whenever it changes and publishes by content hash with a one-year immutable cache, so a plain overwrite leaves a window where a half-written bundle is read. Added `--out=<path>` so measurement / experiment builds land outside `lib/`.
- `scan_discover` entries and schema no longer carry `messageCount` (the panel does not show it, and the SQLite summary readers stop producing it).
- README now includes GUI-import previews (one light, one dark) plus a `screenshots.json` store manifest.
- `package.json` `files` now includes `docs/*.png` so the previews also render on the npm page.

**Full Changelog**: [v0.18.5...v0.19.0](https://github.com/Nwflower/dsh-chat-import/compare/v0.18.5...v0.19.0)

## [0.18.5] - 2026-09-21

- 面板选择区改为一行读完「从 全部来源 导入到 DSH 会话环境」：来源与落点合到同一行、「从」与「导入到」当连接词，工作区另起一行；触发器只留文本（去掉下三角与品牌标，品牌 SVG 标只在下拉弹层行里显示），弹层统一对着这一行定位（行宽 = 弹层宽，窄面板也不会溢出）。
- 工作区下拉里在文件夹名之后用更淡的小字画出绝对路径（宽度不够先截断路径：主标签不参与收缩，文件夹名保持完整，只有它自己超过行宽时才截断；全文留在 title；搜索也匹配路径）；整份选项都没有品牌标时不再保留行首 16px 图标槽位，工作区列表的文字左移贴边。
- 下拉弹层高度改为自适应窗口：列表上限按「视口底部 − 弹层顶端」实测（原来写死 260px），来源列表一屏从 8 行提到近 20 行，短列表仍随内容收缩。
- 选择区只留一行；工作区筛选移到工具栏末位（与动作按钮分组，窄面板下不降级成图标），选择区上下高度提到与其他两层一致（三行统一 8px 12px 内边距）；工具栏去掉「已选 N」（底部主按钮「导入所选 (N)」已经承担）；三个下拉的边框改为与工具栏按钮同款（1px border-l2 + 8px 圆角），内边距收窄。
- 导入面板下拉控件按模型选择器式二级弹层重绘：触发器去掉输入框外观，hover / 展开时浮出一层背景矩形；弹层改为紧凑行（30px 行高、6px 行圆角、行首品牌标、当前项末尾 ✓），搜索框与列表之间加一条分隔线。来源 / 导入到 / 工作区三个下拉统一。
- 导入面板改版：导入按钮移到面板底缘（列表与分页之下），滚动时始终可见；来源 / 导入到 / 工作区三行之间不再画分隔线，三行合并为一组。
- 修复 Antigravity 导入崩溃（`The "path" argument must be of type string …`）：工具层旁读 annotation / 任务回执改为宿主 fs 目标对象契约（先 `resolve` 再 `readText`/`listDir`），单文件与批量导入不再全灭。
- Antigravity 发现迁移到新版存储根 `~/.gemini/antigravity`，旧 CLI 根 `~/.gemini/antigravity-cli` 与 IDE 根 `~/.gemini/antigravity-ide` 继续并扫；`.db`/`.pb` 会话文件按会话 id 去重发现。
- Antigravity 目录批量只收集 canonical `transcript.jsonl`，不再把 `transcript_full.jsonl` 等伴生日志当作独立会话。

## [0.18.4] - 2026-09-20

- 环境变更提示改到首个 `step/start` 之后：旧格式（v0–v2）导入会话不再因宿主 v2→v3 格式迁移被拒载（surface 事件早于首个 step 的形状会被迁移器 fail-closed 拒绝）。
- `verify_session` 新增 `surface-before-first-step` 检查与重导提示：存量旧格式会话被点名，不再等宿主迁移时才暴露。
- 增量续写不再重复注入环境变更提示。
- Kimi Code 缺少 `state.json` 时保留工作区归属。
- 失效旧版 scan-cache，修复 Grok Build 工作区名仍显示 %XX。
- 导入面板图标选中态遮罩按强调色明度选黑/白。

## [0.18.3] - 2026-09-18

- 接入官方原生右侧栏，移除旧版右侧栏与自绘 ShellPanel 回落链。
- 修复 SQLite WAL 盲区与 DB 指纹短路径。
- 修复 `warmProjection` 宿主三参契约调用。
- 用 npm 10 重新生成 lockfile，修复 CI `npm ci` 依赖树漂移。
- Grok Build 工作区列不再显示 %XX 编码乱码。
- 去 AI 化清理：删除未消费层、统一文档计数、移除内部编号。

## [0.18.2] - 2026-09-17

- 修复 ChatGPT 官方导出静默丢弃：兼容缺失 children、占位 root、浮点时间戳。
- `verify_session` 增加非整数时间检查。

## [0.18.1] - 2026-09-17

- 新增 TeleAgent 来源。
- 数据库类批量来源的会话标题统一为「来源 · 话题」。

## [0.18.0] - 2026-09-17

- 面板新增「导入到」下拉：直投 Claude Code / Codex / Kimi Code / opencode。
- 新增 opencode 反向导出。

## [0.17.3] - 2026-09-17

- 修复 Codex 分页链扫描按 thread 过滤。
- 支持 Kimi 新版 `state.json` 的 `workDir` 字段。
- 清理/重导支持带下划线会话 ID。

## [0.17.1] - 2026-09-16

- 修复 Codex 分页 rollout 按 thread 成链导入。

## [0.17.0] - 2026-09-15

- 新增 Crush 来源。

## [0.16.0] - 2026-09-15

- 新增 Zed Agent 来源。

## [0.15.0] - 2026-09-15

- 新增 Goose 来源。

## [0.14.0] - 2026-09-15

- 新增 Cline 来源。

## [0.13.0] - 2026-09-15

- 新增 Continue 来源。

## [0.12.x] - 2026-09-15

- 继续完善来源支持、面板与同步能力；详细历史见 git。

## [0.11.x] - 2026-09-07 ~ 2026-09-15

- 继续新增来源、面板、导出与同步能力；详细历史见 git。

## [0.10.x] - 2026-09-06 ~ 2026-09-07

- 持续完善导入、发现与导出能力；详细历史见 git。

## [0.9.x] - 2026-09-04 ~ 2026-09-06

- 继续完善工具面与文档；详细历史见 git。

## [0.8.x] - 2026-08-26 ~ 2026-09-01

- 新增增量续写、扫描缓存、标题兜底、上下文桥接等能力；详细历史见 git。

## [0.7.x] - 2026-08-23

- 继续完善导入能力与工程基建；详细历史见 git。

## [0.6.x] - 2026-08-17 ~ 2026-08-19

- 继续完善来源支持与互转能力；详细历史见 git。

## [0.5.x] - 2026-08-16

- 继续完善导入与发布流程；详细历史见 git。

## [0.4.0] - 2026-08-16

- 发布规范达标，完善插件元数据与工程配置。

## [0.3.x] - 2026-08-14

- 完善仓库社区健康与工程规范。

## [0.2.0] - 2026-08-14

- 收口版本漂移，补齐 Reasonix/opencode 等能力。

## [0.1.x] - 2026-08-13 ~ 2026-08-14

- 首个发布版本，支持早期外部 Agent 会话导入。