fix(native): 按会话记账 + 宿主 keepMounted，原生 tab 状态不再跨切换丢失（#636 + #661） - #776
Open
yanzhaohui1999 wants to merge 5 commits into
Open
yanzhaohui1999 wants to merge 5 commits into
yanzhaohui1999 wants to merge 5 commits into
Conversation
Collaborator
Rebase 提示：
|
`sync()`'s cleanup loop compared the live registrations against the DESCRIPTOR list, which never contains the `files` takeover (its key is `FILES_KIND`, not a descriptor id). Every store/service notification — a Session switch is one — therefore disposed and immediately rebuilt it: [native] DISPOSE type dsh-better-sidebar:files [native] register type dsh-better-sidebar:files Disposal unmounts that kind's body, so the explorer's own component state (its scroll offset) was rebuilt on every pulse even when the plugin-side record survived. Its lifetime really belongs to the editor type's switch, which the loop below already handles; skip it in the cleanup pass. Pinned by "keeps the files takeover registered across a Session switch pulse": reverting the guard turns it red (`expected 2 to be 1` — 2 is the rebuilt registration count). Rebased onto omdsh-dev#777 (`4099700`): the guard itself now lives on `main` — its cleanup pass already skips `FILES_KIND` and releases through `disposeSafely` — so what remains in this commit is the comment above that guard, recording the second reason it is load-bearing (a re-registration on an already inactive context orphaned the id), plus the regression spec below.
Native tab ids restart in every session (`tab1`, `tab2`, …), and DSH 0.1.7 keeps a visited tab body mounted across hiding, tab selection and Session switches, so several sessions' same-named tabs are alive at once. The plugin's registry was a single-keyed map, which made the entering session's body adopt the leaving one's record — and the leaving body's unmount cleanup then delete it. The visible result was omdsh-dev#636: after a conversation switch the explorer stopped responding entirely (no request sent), in both directions. Records are now filed under `sessionId::tabId`, and the seat session is carried explicitly instead of inferred: - `createNativeTabRecords` takes the seat session on `ensure`, and `get/has/update/drop/toggleExpanded/versionOf` all take it too; `retain(sessions)` reclaims the records of sessions that are gone; - the title slot gets its own seat `sessionId` — the chip is drawn BEFORE the body, so it used to read whichever record the live slot happened to hold (another conversation's title on the switch frame); - `SidebarSurface.update/has/activate` take an optional `sessionId`: with one, it names the session; without, the mounted seat answers, a unique match still resolves, and an id that names a tab in several sessions resolves to none rather than guessing (writing into another conversation is worse than refusing); - `BetterSidebarService.updateTab` forwards it, and the six call sites that know their own scope (EditorHost, SideChatView, ChangesTab, tree-mutations) pass theirs. Tests (native-surface.spec.ts, +8): per-seat-session identity (same-named tabs stay apart, a patch/drop hits exactly one session, `retain` reclaims), and surface resolution (explicit session wins, mounted seat is the fallback, an ambiguous bare id is refused, close plus dead-session reclaim). Each is red without this change.
…n end to end DSH 0.1.7 gives a tab TYPE the `keepMounted` switch: a marked body stays mounted through hiding, tab selection and Session switches. That is exactly what the plugin tried to fake with module-level memory, so declare it on every plugin type (the descriptor types and the `files` takeover) and let the host hold the component state — tree scroll, unsaved drafts, chip renames — with no plugin-side memory at all. With the bodies staying up, the per-session record registry from the previous commit is what keeps co-resident sessions apart; both halves are required, because a body may still be remounted by the host (type disposal, session teardown) and the records must outlive that. Verification on a real host (`scripts/e2e-mount.sh`: npm pack -> fresh scratch profile -> real `dsh web` pinned to 0.1.7-rc.1 -> headless Chromium), new lane `tests/e2e/tree-scroll.e2e.ts`: build state in conversation A, switch to B, come back, and compare the whole observable vector (tree expansion, scroll offset, chip title) — plus the negative half, that B does not inherit A's state. That lane also exposed a defect in the old one: it picked its "second conversation" by row index, but the session tree's first row is a WORKSPACE row and a seeded session with no messages takes no row at all, so the "A -> B -> A" round trip was re-selecting the same session and every assertion held vacuously. It now identifies sessions by `data-row-key="session:<id>"` and creates a real second session through the rail. `tests/e2e/mount.e2e.ts` reaches the side-chat tab through its guide entry now: with the record retained, the chip shows the thread's own title, so the old `/Side Chat/` chip match no longer names it. Docs: AGENTS.md records the two facts that must hold together, and docs/plans/2026-09-25-native-tab-state-retention-keepmounted.md replaces the park/archive design (PR omdsh-dev#712) with this one.
…olds "State survived" is not the same claim as "nothing was silently reloaded": the retention lane asserted the visible vector (expansion, scroll, chip) but never looked at the network, so a body that remounted and refetched everything would have passed it. The lane now counts the plugin's own reads. These routes are POSTs, so the path being listed or read travels in the BODY — the counter parses it, which is what lets one level or file be told apart from a fresh read. It then opens a real file in conversation A and requires, on return: - the file is still open as a tab, and selecting it shows the editor without a second `fs.read` of its bytes; - no `fs.read` of that file happened anywhere in the round trip. Measured against the real host (0.1.7-rc.1, headless Chromium): zero plugin reads on returning to A — the seat lands on its previous tab and shows the content it already holds. One ordering fact this lane had to be taught: opening a file in merged mode swaps THAT tab's content to the editor, so the tree component unmounts and its scroll starts over by design (attributed with a probe: hide/show keeps 320, opening a file drops it to 0). The lane therefore re-builds the scroll offset after the file step before measuring the conversation round trip, which is the claim the offset is actually evidence for.
…prevent The comparison with PR omdsh-dev#712 left three of its cases uncovered here: omdsh-dev#712 parked an unsaved editor draft, an unsaved commit message and the chosen worktree in module-level memory and tested each by unmounting. This branch claims the host's `keepMounted` makes that memory unnecessary, but the claim was only ever exercised for tree expansion and scroll — the draft was explicitly recorded as unverified, and it is the one loss a reader cannot re-derive by looking again. All three are now measured on a real host, in the retention lane's A -> B -> A round trip: - an unsaved EDIT typed into the file's editor must still be in the document; - an unsaved COMMIT MESSAGE must still be in the changes tab's box; - the CHOSEN WORKTREE must still be selected, rather than re-derived from the inventory. The lane's workspace becomes a real git repo (idempotently: it persists across local runs) with one staged change — that is what renders a usable commit box — plus a linked checkout for the worktree choice. Two host details the lane has to respect, both recorded in comments: the worktree selector only renders when the inventory holds more than one entry, and paths are compared through `realpathSync` because macOS /var is a symlink to /private/var while git reports the resolved form. Measured: 8/8 on 0.1.7-rc.1 + headless Chromium, and the full unit suite (113 files / 1141 tests) stays green.
yanzhaohui1999
force-pushed
the
fix/native-tab-state-retention
branch
from
September 28, 2026 03:36
007f846 to
77887e2
Compare
Contributor
Author
|
已按提示 rebase 到
验证（都在
其余部分（按会话记账 + 宿主 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概述
原生右侧栏的 tab 插件侧状态在切换 tab / 切换会话时会丢。本 PR 用 DSH 0.1.7 的
keepMounted+ 「按会话记账的记录表」修掉，不引入任何 module-level 记忆层。Closes #636、Closes #661。取代 #712（park/归档方案）。
三个根因
createSurface()→counting(0)），而插件的记录表按裸 id 单键存。进入会话的 body 在 render 阶段接管同号记录，紧接着被离开会话的卸载清理删掉；versionOf(缺失) = 0→ 0→0 快照不变 → 不重渲染 → 此后点击全是静默 no-op（一个请求都不发）。双向失灵。外加一个 #712 未发现的独立根因（本 PR 的第一个 commit 单独修）：
sync()的清理循环拿「描述符清单」比对 live 注册表，而files接管不是描述符（键是FILES_KIND）——于是每一次 store/service 通知（会话切换、改设置、任何 state 变更）都会把它注销再重建：注销即卸载该 kind 的 body —— 所以记录活下来、组件态（滚动/草稿所在的视图）仍会重建。这是「#712 修好了记录、切回会话还是丢」的真答案。
方案
tabId单键sessionId::tabId复合键keepMounted：body 不卸载，组件态免费保留surface.close/ 会话消失；卸载不删files接管寿命update/has/activate新增可选sessionId；缺省按sidebarRight.mounted解析，多会话同名时拒绝而不是猜SidebarSurface只新增可选参数，不删既有签名——在屏会话下与旧行为完全一致，外部插件不传照旧。验证
真机（
scripts/e2e-mount.sh：npm 打包 → 全新 scratch profile → 真实dsh web钉 0.1.7-rc.1 → headless Chromium），tests/e2e/tree-scroll.e2e.ts的 A → B → A 往返，断言完整状态向量：fs.read不再重发——「状态还在」与「其实悄悄重载了一遍」是两回事，所以请求也计数了（路由是 POST，路径在 body 里）；8/8 通过。
单测：
tests/native-surface.spec.ts+8（按会话隔离、按会话 patch 与 drop、retain驱逐、显式 session / 缺省上屏 / 歧义拒绝、同屏双 seat 互不继承、files接管不得被会话脉冲重建）。每条在缺对应修复时都真的红。tsc --noEmit0 错、eslint .干净、全量单测 113 文件 / 1141 例通过、check-consumer-types双档通过。顺带修掉的 e2e 缺陷
#712的 lane 一直没真的切过会话：它按行号点「第二个会话」，而会话树的第一行是 workspace 行，且没有消息的种子会话根本不占行——那次「A → B → A」实际是重选同一个会话，三条断言天然为真。现在按data-row-key="session:<id>"认会话、用 rail 自己的 New session 建出真正的第二个会话，并断言「确实离开了 A」。风险与边界
keepMounted（0.1.7 新增的可选字段）与ISidebarRight.mounted；上游若撤掉，退化为「切 tab 丢组件态」，记录层仍按会话正确。sidebar-right的挂载模型（重新变得只挂当前 tab），组件态会退回需要记忆层——files接管那条修复与复合键记录是抗这种情况的底线。keepMounted的 body 与其记录活到 tab 关闭 / 会话消失，宿主没有 LRU 上限；访问过的会话越多常驻越高。属宿主策略，插件无法单方面设界，记为已知代价。文档
docs/plans/2026-09-25-native-tab-state-retention-keepmounted.md（设计过程与实测记录）AGENTS.md§3.4 新增「原生 tab 的状态保留」段（两条必须同时成立的事实）