dsh-model-switch 0.4.33 → 0.4.35

package/CHANGELOG.md

@@ -1,3 +1,21 @@
+## v0.4.35
+
+「自动审查」不再在每次热重载后掉回「完全权限」：释放 Auto 注册之前先把当时选着 Auto 的会话落盘，重新认领后按 id 恢复。
+
+- **P0 root cause: every release moves the selection, and the in-memory guard dies with the row (`src/auto-review.ts`).** The official gate's teardown moves live `auto` sessions to the fallback preset (`danger-full-access`) by design. This row triggers that teardown on both of its release paths: the `yieldRegistration` seam the takeover runs before a recompose, and its own row disposal during a hot reload. `preserveAuto` could not cover either window — it remembers around an **in-process** action, but a reload loses the selection before anything of ours runs again, and the next mount's `rememberAuto()` returns `[]` because the sessions were already flipped. The user-visible symptom: the session kept reverting from「自动审查」to「完全权限」, reproducibly **at the moment `cordis.patch.yml` was edited** (any profile patch edit forces a recompose).
+- **Fix: durable snapshot + replay (`src/auto-sessions-state.ts`, `src/auto-review.ts`).** `persistAutoSnapshot()` writes the ids of the sessions currently on Auto to `$DSH_HOME/model-switch/auto-sessions.json` **before** each release — from the takeover's `yieldRegistration` seam and from the last-registered teardown effect (the first thing to run on disposal). `claim()` then calls `flushAutoSnapshot()` once the registration is held again, restoring exactly those sessions and consuming the file. A snapshot is rejected after 15 minutes (`AUTO_SESSION_STATE_MAX_AGE_MS`), so a later restart cannot resurrect a preset the user has since chosen; every malformed shape degrades to a no-op, and a read-only `DSH_HOME` never blocks the release.
+- **Tests.** New `tests/auto-sessions-state.spec.ts` pins the round trip, dedupe and dropping of non-string ids, the malformed-shape matrix (missing, empty, invalid JSON, wrong shape, non-numeric `savedAt`), and the 15-minute replay window. `pnpm run check` green (56 test files / **500** tests + pack gate).
+- **Live verification (before tagging).** The build was installed into the real desktop profile and the web UI was driven over CDP from this machine: the session was switched to「自动审查」, a `cordis.patch.yml` edit forced a recompose — the exact action that used to flip the preset — and the preset chip in the live page still read「自动审查」afterwards. The snapshot directory (`$DSH_HOME/model-switch/`) was created and consumed by the same recompose, so the persist → replay path ran end to end.
+
+## v0.4.34
+
+The「停用失败：dsh: warning: 1 entry did not activate auto-review … permission: preset "auto" is already registered」toast is gone: every official-row activation now gets the registration yielded first, and this row releases the slot in the first synchronous step of its own teardown.
+
+- **P0 root cause: the yield was one-shot (`src/auto-review-takeover.ts`).** The `loader/patch-context` listener kept an `handled` set so "a repeated patch-context does not loop" — but that guard also **skipped the yield from the second activation on**. The loader calls `_patchContext([])` right before every `registry.plugin` mount (confirmed in the loader source: `Entry._init()` = import → `_patchContext([])` → `registry.plugin`), and a plugin-manager enable goes through `Entry.update({ disabled: false }) → init()` — so the re-activating official row's `registerAuto` ran while this row still held the slot, failed with `preset "auto" is already registered`, and the manager reported the whole operation as 「停用失败」. `Entry.update({ disabled: true })` never emits `partial-dispose`, so the set was never cleared and the suppression was permanent. The guard is gone: **every** activation of the official row yields synchronously before the loader continues (`releaseForTakeover` is idempotent), and the existing per-entry single-flight reclaim keeps repeated triggers on one cycle, so this cannot loop. Two regression tests pin it: a second activation yields again, and non-official rows stay out of the way.
+- **Release first on our own teardown (`src/auto-review.ts`).** Disposing the gate fiber now releases the Auto registration in its **first synchronous step** (the loader does not await fiber disposal, and the official row's re-activation in the same recompose must find the slot free), and a last-registered effect releases it again as the teardown's first step — belt-and-braces for registrations claimed without a gate fiber (a direct `reclaimRegistration()`).
+- **Live reproduction and verification.** Reproduced on the real desktop Host through the web UI before fixing: enabling「自动授权审查」with `dsh-model-switch` active **failed twice** (the switch silently reverted), and the very same operation succeeded with `dsh-model-switch` disabled — the seam conflict isolated by a single variable. `pnpm run check` green (55 test files / **496** tests + pack gate). Post-fix web-UI regression covers enable/disable cycles with the official row active: no toasts, no warnings, no failed activations.
+- **Diagnosis recorded (the reported "运行失败" was not this plugin).** The turn-28 outage was a **model-request transport failure** (`Connection error.` / `TRANSPORT`, provider `apigoto`, five fast rejections in 8 s) — the same failure class this deployment has seen on provider accounts before (10-13 `ark` ×10, 10-15 `xiaomi` ×29). A web-UI A/B with this plugin **on and off** ran 5/5 model requests clean in both states (and the naming call alongside), and this repo has no code path that touches the request transport (no `llm/stream` listener, no fetch/global patch). The 17:49 recovery after disabling plugins is confounded: the permission preset flipped `auto → danger-full-access` in the same minute, which also removes the per-tool-call review traffic against the same account.
+
 ## v0.4.33
 
 The reviewer takeover now reports its own health to the settings page instead of only writing a Host log line.
@@ -5,6 +23,7 @@ The reviewer takeover now reports its own health to the settings page instead of
 - **Runtime-capability field (`src/runtime-capabilities.ts`, `src/host-runtime.ts`).** `autoReview` joins the capability snapshot: `{ available: true }` while this row holds the official reviewer route, `{ available: false, reason: 'auto-review-takeover' }` once it does not. The value comes from a live record (`autoReviewStatus` in `src/auto-review.ts`) updated on every claim, on the lost-seam path, and by the tool-call safety net — which also counts the calls it allowed unreviewed, so the card can say how much review was skipped.
 - **Settings card warning (`src/client/ModelSwitchSettings.tsx`, `src/client/locales.ts`).** The「自动审查」section renders `autoReviewTakeoverNote` while the route is held and `autoReviewTakeoverWarning` otherwise (zh + en: tool calls keep running, those calls are unreviewed, restart DSH to converge). What used to be a Host-only log line is now visible where the human configures the route.
 - **Verification.** `pnpm run check` green (55 test files / 494 tests + pack gate). The safety net's test now also asserts the allowed-call counter the snapshot reads, so the card's data path is covered end to end (host record → capability snapshot → card text selected by `available`).
+- **Lab reproduction of the reported outage (released artifact).** The published `dsh-model-switch@0.4.33` was installed in the lab profile (`DSH_HOME=~/.dsh-lab`) with the official `auto-review` row **force-enabled** in the profile patch, the Host was booted on 3082, and the web UI was driven over CDP (a dedicated Edge instance, the user's own browser untouched). A first real tool call completed in 8 s with the actual directory listing. **Without restarting the Host**, `dsh plugin --profile lab remove dsh-model-switch` + `add dsh-model-switch@0.4.33` performed the hot swap and the profile patch toggled the official row off and back on (two runtime recompositions); a second real tool call then completed in 7 s with the file list, and the page contained **no** `inactive context` / `cannot get required service` / `body was not executed` / `Auto review of tool` failure. This is the disable → remove → install → enable sequence that used to require a DSH restart.
 
 ## v0.4.32