chore: start documentation work

feat: add required_permissions support for Shell sandbox policy

.agents/skills/cursor-prefix-stability/SKILL.md

@@ -0,0 +1,82 @@
+---
+name: cursor-prefix-stability
+description: Implement and review Cursor BYOK conversation projection with append-only provider history and prefix-cache stability. Use when changing runtime prompts, request context, canonical messages, checkpoint hydration, compaction, message identity, or provider history serialization under server.
+---
+
+# Cursor prefix stability
+
+Preserve the longest valid provider prefix across turns. Treat provider-visible history as an append-only log unless an explicit compaction operation replaces it.
+
+## Architecture
+
+Keep the invariant in the provider-independent conversation layers:
+
+```text
+server/
+├── prompt/cursor/*/runtime.md       Per-turn runtime content
+├── src/cursor/request/              Request context and runtime compilation
+├── src/cursor/projection/           Canonical ↔ Cursor checkpoint projection
+├── src/cursor/checkpoint/           Stable roots, turns, and hydration
+├── src/run/                         Provider-independent model history
+└── src/provider/                    Provider-specific serialization only
+```
+
+Do not solve prefix instability independently in each provider adapter. Produce one stable canonical history before dispatching to OpenAI Responses, OpenAI Chat, Anthropic, or another provider.
+
+## Required invariants
+
+- When no compaction occurs, the complete provider-visible history from turn N must be an exact structural prefix of turn N+1. Never edit, remove, merge, reorder, renormalize, or regenerate an earlier message.
+- Keep `PromptSpec.instructions` and the stable tool prefix byte-stable when their inputs have not changed. Deterministic ordering is required; do not use unordered iteration in provider-visible output.
+- Separate conversation/request context from the per-turn runtime message. The runtime message contains current-turn material such as the user query, selected context, open files, action context, mode reminders, and timestamp.
+- Project rules, skills, subagents, environment/Git context, and MCP metadata as a stable `request-context:*` message:
+  - append it on the first applicable turn;
+  - do not append it again when its compiled content is identical to the latest projected request context;
+  - when its content changes, append a new request-context message immediately before the current runtime message;
+  - never represent a context change by rewriting the system prompt, replacing an earlier context message, or mutating a checkpoint root.
+- Give every appended context update a unique event identity. Compare the latest context by content, not only by identifier, so `A → B → A` appends the final `A` again while retries of the same event remain idempotent.
+- Preserve `request-context:*` wire identity through checkpoint encoding and hydration. Deduplication must still work after process restart or conversation resume.
+- Automatic compaction is an explicit prefix reset. Compact obsolete history, retain exactly the latest request-context message, then place the summary and current initial messages in deterministic order. Manual compaction may reproject current context on the next user turn.
+- Background completions and injected runtime events must not manufacture duplicate request context unless they actually start a user turn whose context changed.
+
+## Change workflow
+
+Before editing, trace the whole path that applies:
+
+```text
+AgentRunRequest
+→ request context hydration/compilation
+→ CanonicalMessage identity and persistence
+→ checkpoint encode/decode
+→ projected ModelRequest history
+→ provider serialization
+```
+
+Determine which data is conversation-level and which is turn-level. If a proposed change moves or rewrites an earlier provider-visible value, redesign it as a new append-only event unless the operation is explicitly compaction.
+
+Use TDD for changes in this path. Start with a failing behavioral test, then implement the smallest provider-independent change.
+
+## Verification
+
+Cover the affected behavior with structural assertions, not token-count estimates alone:
+
+- Two turns with identical request context: the first request history is an exact prefix of the second, the system instructions are identical, and only one `request-context:*` message exists.
+- Changed context: one new context message appears at the tail before the new runtime query; all earlier messages remain unchanged.
+- Context reversion `A → B → A`: three distinct context events are retained in order.
+- Retry of one runtime event: no duplicate or conflicting context message is persisted.
+- Checkpoint round-trip: request-context identity and content survive encode/hydrate.
+- Automatic compaction: only the latest context is retained outside the summary.
+- Runtime templates render without embedding conversation-level rules or MCP metadata in every user query.
+
+Run focused tests first, then the relevant server suites:
+
+```bash
+cargo test --lib
+cargo test --test runtime_modes
+cargo test --test prefix_stability
+cargo test --test checkpoint_recovery
+cargo test --test compaction
+cargo clippy --lib -- -D warnings
+cargo fmt --all -- --check
+```
+
+Do not repair unrelated dirty-worktree failures while validating. Report them separately.

.agents/skills/release/SKILL.md

@@ -0,0 +1,77 @@
+---
+name: release
+description: Prepare, authorize, publish, troubleshoot, and verify Cursor BYOK desktop GitHub Releases. Use for version bumps, release tags, GitHub Actions release runs, updater manifests, signing, or release-readiness checks.
+---
+
+# Desktop release
+
+Release through `.github/workflows/release.yml`. Preserve both updater formats: Tauri uses `latest.json`; legacy `v0.0.49` clients use `update.json`.
+
+## Publication authority
+
+- Only the repository author, GitHub user `leookun`, may authorize a live release.
+- Before any live mutation, require an explicit release instruction from the author in the current task and verify `gh api user --jq .login` returns `leookun`.
+- Treat all of these as publication actions: pushing a `v*` tag, rerunning the release workflow, and publishing or editing a GitHub Release. Pushing a release commit to `main` only prepares the release and must never trigger publication by itself.
+- Without that authorization, restrict work to inspection, local edits, validation, and a release-ready commit or branch. Do not infer publication permission from requests such as “prepare”, “check”, or “ready to release”.
+- Never print, commit, or upload `.tauri/cursor-byok.key` anywhere except the repository's `TAURI_SIGNING_PRIVATE_KEY` Actions Secret when the author explicitly requests that secret configuration.
+- Never delete, replace, or move an existing tag or published Release without separate explicit authorization.
+
+## Version and GitHub Release policy
+
+- Do not use GitHub prereleases. Keep `prerelease: false` for every release and publish the completed release as Latest.
+- Use `vMAJOR.MINOR.PATCH` for a stable tag, for example `v0.1.0`.
+- Use standard SemVer `vMAJOR.MINOR.PATCH-beta.N` for a test tag, for example `v0.1.0-beta.1`. A beta is still a normal GitHub Release, not a GitHub prerelease. Make its title or body visibly say Beta.
+- This normal-Release rule is required because both installed update clients resolve assets through GitHub's `/releases/latest/download/` path, which excludes GitHub prereleases.
+- Windows beta builds must use the NSIS bundle. WiX/MSI rejects nonnumeric prerelease identifiers such as `beta.1`; do not weaken the SemVer tag to accommodate MSI.
+- Release only from a `v*` tag whose commit is contained in `origin/main`. The tag must equal `v<version>` from the desktop manifests.
+- Keep ordinary `main` pushes and manual workflow dispatch disabled as release triggers. The author pushes the matching tag only after the release commit is present on `origin/main`.
+- Never republish an already published version. Select a new version instead.
+
+## Release sources
+
+Keep the desktop version identical in the manifests and their locks:
+
+```text
+cursor-byok/
+├── Cargo.lock
+├── apps/desktop/
+│   ├── package.json
+│   ├── package-lock.json
+│   └── src-tauri/
+│       ├── Cargo.toml
+│       └── tauri.conf.json
+├── scripts/cursor-proto/proto/
+│   ├── agent_v1.proto
+│   └── aiserver_v1.proto
+└── .github/workflows/release.yml
+```
+
+The two listed Proto files are required build inputs and must be committed. Keep the other locally extracted Proto files ignored unless the build starts depending on them.
+
+## Prepare and validate
+
+1. Inspect `git status`, fetch `origin/main`, and preserve unrelated user changes. Confirm the release commit is based on the current remote head.
+2. Choose stable or beta numbering explicitly. Update the desktop version in both manifests and lockfiles; do not change the independent `cursor-server` version merely to release the desktop app.
+3. Confirm the updater public key in `tauri.conf.json` matches `.tauri/cursor-byok.key.pub` without exposing the private key.
+4. Confirm `TAURI_SIGNING_PRIVATE_KEY` exists in GitHub Actions. `TAURI_SIGNING_PRIVATE_KEY_PASSWORD` must be absent when the local key has no password.
+5. Confirm neither the intended tag nor Release already exists.
+6. From `apps/desktop`, run:
+
+   ```bash
+   npm run check
+   npm run tauri:build -- --debug --no-bundle
+   ```
+
+7. Validate the workflow YAML and inspect the staged diff. Ensure `.tauri/`, unrelated local files, and unrelated user changes are not staged.
+8. Use the `tauri-action@v1` input `uploadUpdaterJson: true`; `includeUpdaterJson` is not a valid v1 input.
+
+## Publish and verify
+
+After the author explicitly authorizes publication:
+
+1. Commit only the reviewed release set and push it to `main`. Confirm the release commit is present in `origin/main`; this push must not start the release workflow.
+2. Create the matching tag on that commit, for example `v0.1.0-beta.1`, and push only that tag. This tag push is the publication trigger.
+3. Follow the triggered `Release desktop app` run through completion. Report the run URL and stop on failure; diagnose locally before asking the author to authorize another live attempt.
+4. Verify `v<version>` exists, is published rather than draft, has `prerelease: false`, and is the repository's Latest release.
+5. Verify the Release contains signed Tauri updater artifacts plus `latest.json`, and the legacy platform archives plus `update.json`.
+6. For a beta, report clearly that it is a test version even though GitHub represents it as a normal Latest Release.

.agents/skills/release/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Desktop Release"
+  short_description: "Prepare, authorize, publish, and verify desktop releases"
+  default_prompt: "Use $release to prepare and verify a Cursor BYOK desktop release."

.github/workflows/release.yml

@@ -2,11 +2,8 @@ name: Release desktop app
 
 on:
   push:
-    branches:
-      - main
     tags:
       - "v*"
-  workflow_dispatch:
 
 permissions:
   contents: write
@@ -24,6 +21,28 @@ jobs:
       version: ${{ steps.version.outputs.version }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Verify release author
+        env:
+          REPOSITORY_OWNER: ${{ github.repository_owner }}
+        shell: bash
+        run: |
+          if [[ "${GITHUB_ACTOR}" != "${REPOSITORY_OWNER}" ]]; then
+            echo "Only ${REPOSITORY_OWNER} may publish a release" >&2
+            exit 1
+          fi
+
+      - name: Verify updater signing key
+        env:
+          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
+        shell: bash
+        run: |
+          if [[ -z "${TAURI_SIGNING_PRIVATE_KEY}" ]]; then
+            echo "TAURI_SIGNING_PRIVATE_KEY is not configured" >&2
+            exit 1
+          fi
 
       - name: Read and verify app version
         id: version
@@ -34,8 +53,12 @@ jobs:
           cargo_version=$(sed -n '/^version = / { s/version = "\([^"]*\)"/\1/p; q; }' apps/desktop/src-tauri/Cargo.toml)
           test "${version}" = "${package_version}"
           test "${version}" = "${cargo_version}"
-          if [[ "${GITHUB_REF_TYPE}" = "tag" ]]; then
-            test "${GITHUB_REF_NAME}" = "v${version}"
+          test "${GITHUB_REF_TYPE}" = "tag"
+          test "${GITHUB_REF_NAME}" = "v${version}"
+          git fetch origin main:refs/remotes/origin/main
+          if ! git merge-base --is-ancestor "${GITHUB_SHA}" origin/main; then
+            echo "Release tag must point to a commit contained in origin/main" >&2
+            exit 1
           fi
           echo "version=${version}" >> "${GITHUB_OUTPUT}"
 
@@ -68,7 +91,7 @@ jobs:
             target: ""
           - platform: windows-x86_64
             os: windows-latest
-            args: ""
+            args: "--bundles nsis"
             target: ""
           - platform: macos-aarch64
             os: macos-15
@@ -120,10 +143,10 @@ jobs:
           projectPath: apps/desktop
           tagName: v__VERSION__
           releaseName: Cursor BYOK v__VERSION__
-          releaseBody: Download the installer for your platform from the assets below.
+          releaseBody: ${{ contains(needs.prepare.outputs.version, '-') && 'Beta release. Download the installer for your platform from the assets below.' || 'Download the installer for your platform from the assets below.' }}
           releaseDraft: true
           prerelease: false
-          includeUpdaterJson: true
+          uploadUpdaterJson: true
           args: ${{ matrix.args }}
 
       - name: Package legacy Linux updater asset

.gitignore

@@ -3,7 +3,7 @@ claude-server.tar
 dist
 .task
 /local-docs/
-bin
+/bin/
 logs/
 cursor-server.db
 cursor-server.db-shm
@@ -30,4 +30,6 @@ server-go/log/
 .cursor-app-formatted/
 proto/extensions-cursor-app/
 cursor-tab-server/cursor-tab-server-linux-amd64.tar
-/scripts/cursor-proto/proto/
+/scripts/cursor-proto/proto/*
+!/scripts/cursor-proto/proto/agent_v1.proto
+!/scripts/cursor-proto/proto/aiserver_v1.proto

Cargo.lock

@@ -1128,7 +1128,7 @@ checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
 
 [[package]]
 name = "cursor-byok-desktop"
-version = "0.1.0"
+version = "0.1.0-beta.10"
 dependencies = [
  "axum",
  "cursor-server",

Makefile

@@ -1,3 +1,5 @@
+LOCAL_TAURI_SIGNING_KEY := $(CURDIR)/.tauri/cursor-byok.local.key
+
 .PHONY: check dev-web dev-server dev-desktop build-web build-server build-desktop build-docker
 
 check:
@@ -21,8 +23,13 @@ build-web:
 build-server:
 	cargo build --release --package cursor-server --bin cursor-server
 
-build-desktop:
-	npm --prefix apps/desktop run tauri:build
+$(LOCAL_TAURI_SIGNING_KEY):
+	@install -d -m 700 "$(dir $@)"
+	@apps/desktop/node_modules/.bin/tauri signer generate --ci --write-keys "$@" >/dev/null
+	@chmod 600 "$@" "$@.pub"
+
+build-desktop: $(LOCAL_TAURI_SIGNING_KEY)
+	TAURI_SIGNING_PRIVATE_KEY="$(LOCAL_TAURI_SIGNING_KEY)" TAURI_SIGNING_PRIVATE_KEY_PASSWORD="" npm --prefix apps/desktop run tauri:build
 
 build-docker:
 	docker build --tag cursor-byok:local .

apps/desktop/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.10",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "cursor-byok-desktop",
-      "version": "0.1.0",
+      "version": "0.1.0-beta.10",
       "license": "MIT",
       "dependencies": {
         "@floating-ui/dom": "^1.8.0",

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.10",
   "description": "Cursor BYOK desktop management application",
   "type": "module",
   "scripts": {

apps/desktop/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cursor-byok-desktop"
-version = "0.1.0"
+version = "0.1.0-beta.10"
 edition = "2021"
 publish = false
 

apps/desktop/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Cursor BYOK",
-  "version": "0.1.0",
+  "version": "0.1.0-beta.10",
   "identifier": "dev.cursorbyok.desktop",
   "build": {
     "beforeDevCommand": "npm run dev",

apps/desktop/src/App.tsx

@@ -10,8 +10,8 @@ import { CursorSettingsPage } from "./pages/CursorSettingsPage";
 import { HomePage } from "./pages/HomePage";
 import { ProvidersPage } from "./pages/ProvidersPage";
 import { SettingsPage } from "./pages/SettingsPage";
-import { checkForUpdate, hasNativeAppLifecycle } from "./native/appLifecycle";
 import { useAppStore } from "./store/appStore";
+import { updateStore } from "./store/updateStore";
 
 export function App() {
   return (
@@ -47,11 +47,9 @@ function AppMessages() {
   }, [error, showMessage]);
 
   useEffect(() => {
-    if (!hasNativeAppLifecycle()) return;
-    void checkForUpdate().then(async (update) => {
-      if (!update) return;
-      showMessage(t("发现新版本 {version}，可在设置中安装", { version: update.version }), { duration: 6_000 });
-      await update.close();
+    void updateStore.check().then((version) => {
+      if (!version) return;
+      showMessage(t("发现新版本 {version}，可在设置中安装", { version }), { duration: 6_000 });
     }).catch(() => {
       // Startup checks are best-effort; manual checks in Settings report errors.
     });

apps/desktop/src/api.ts

@@ -8,6 +8,7 @@ export interface Provider {
   name: string;
   provider_type: ProviderType;
   base_url: string;
+  api_key?: string;
   has_api_key: boolean;
   custom_headers: Record<string, string | null>;
   extra_params: Record<string, unknown>;
@@ -56,6 +57,15 @@ export interface ModelInput {
   supports_image_generation: boolean;
 }
 
+export interface ModelConnectivityResult {
+  duration_ms: number;
+  first_text_ms: number | null;
+  output_tokens: number;
+  tokens_per_second: number;
+  tokens_estimated: boolean;
+  output: string;
+}
+
 export type CaState = "missing" | "untrusted" | "ready" | "invalid" | "unsupported";
 export type IntegrationState = "disabled" | "enabled" | "degraded";
 export interface CursorHarnessStatus {
@@ -97,6 +107,13 @@ export interface ProxySettingsInput {
   password?: string;
 }
 
+export type TabMode = "public" | "direct" | "custom";
+
+export interface TabSettings {
+  mode: TabMode;
+  address: string;
+}
+
 export interface OverviewMetrics {
   llm_calls: number;
   successful_calls: number;
@@ -214,7 +231,14 @@ async function request<T>(path: string, init?: RequestInit): Promise<T> {
     throw new Error(t("无法连接本地管理服务"), { cause });
   }
   if (!response.ok) {
-    const message = await response.text();
+    const body = await response.text();
+    let message = body;
+    try {
+      const parsed = JSON.parse(body) as { message?: unknown };
+      if (typeof parsed.message === "string") message = parsed.message;
+    } catch {
+      // Plain-text errors are already suitable for display.
+    }
     throw new Error(message || `${response.status} ${response.statusText}`);
   }
   if (response.status === 204) return undefined as T;
@@ -241,6 +265,7 @@ export const api = {
   models: () => request<Model[]>("/models"),
   updateModel: (hash: string, model: ModelInput) => request<Model>(`/models/${hash}`, { method: "PUT", body: JSON.stringify(model) }),
   deleteModel: (hash: string) => request<void>(`/models/${hash}`, { method: "DELETE" }),
+  testModel: (hash: string) => request<ModelConnectivityResult>(`/models/${hash}/test`, { method: "POST" }),
   overview: (filter?: { startMs: number; endMs: number; modelHashes?: string[]; providerIds?: number[] }) => {
     const params = new URLSearchParams();
     if (filter) {
@@ -283,4 +308,6 @@ export const api = {
   clearStatisticsStorage: () => request<StatisticsStorage>("/settings/storage/statistics", { method: "DELETE" }),
   proxySettings: () => request<ProxySettings>("/settings/proxy"),
   setProxySettings: (settings: ProxySettingsInput) => request<ProxySettings>("/settings/proxy", { method: "PUT", body: JSON.stringify(settings) }),
+  tabSettings: () => request<TabSettings>("/settings/tab"),
+  setTabSettings: (settings: TabSettings) => request<TabSettings>("/settings/tab", { method: "PUT", body: JSON.stringify(settings) }),
 };

apps/desktop/src/components/ProviderEditor.tsx

@@ -1,5 +1,5 @@
 import type { ProviderInput, ProviderType } from "../api";
-import { FormField, TextInput } from "./ui/FormControls";
+import { FormField, SecretTextInput, TextInput } from "./ui/FormControls";
 import { JsonEditor } from "./ui/JsonEditor";
 import { Select } from "./ui/Select";
 import { claudeIcon, openAiIcon } from "./ui/icons";
@@ -23,7 +23,7 @@ export function ProviderEditor({ value, headersText, extraText, editing, onChang
       { value: "anthropic", label: "Anthropic", icon: claudeIcon },
     ]} onChange={(provider_type) => patch({ provider_type: provider_type as ProviderType })} /></FormField>
     <FormField label="Base URL" hint={t("模型服务的 API 根地址；修改后会同步更新该上游模型的路由身份。")}><TextInput placeholder="https://api.example.com/v1" value={value.base_url} onChange={(event) => patch({ base_url: event.target.value })} /></FormField>
-    <FormField className={styles.fullWidth} label="API Key" hint={editing ? t("留空表示保留当前 API Key。") : t("访问模型服务所需的密钥。")}><TextInput type="password" autoComplete="off" placeholder={editing ? t("留空以保留当前密钥") : "sk-xxxxxx"} value={value.api_key ?? ""} onChange={(event) => patch({ api_key: event.target.value })} /></FormField>
+    <FormField className={styles.fullWidth} label="API Key" hint={editing ? t("留空表示保留当前 API Key。") : t("访问模型服务所需的密钥。")}><SecretTextInput autoComplete="off" placeholder={editing ? t("留空以保留当前密钥") : "sk-xxxxxx"} value={value.api_key ?? ""} onChange={(event) => patch({ api_key: event.target.value })} /></FormField>
     <FormField className={styles.fullWidth} label={t("自定义 Headers JSON")} hint={t("值必须是字符串；编辑时 null 表示保留对应敏感 Header 的原值。")}><JsonEditor ariaLabel={t("自定义 Headers JSON")} value={headersText} onChange={onHeadersChange} /></FormField>
     <FormField className={styles.fullWidth} label={t("额外参数 JSON")} hint={t("合并到该上游所有模型的请求体。")}><JsonEditor ariaLabel={t("额外参数 JSON")} value={extraText} onChange={onExtraChange} /></FormField>
   </div>;

apps/desktop/src/components/charts/DailyTokenUsageChart.tsx

@@ -77,9 +77,10 @@ export function DailyTokenUsageChart({
   const maximumTotal = data.reduce((maximum, day) => Math.max(maximum, totalTokens(day)), 0);
   const axisMaximum = Math.max(1, maximumTotal / DATA_HEIGHT_RATIO);
   const emptyBarHeight = axisMaximum * EMPTY_BAR_RATIO;
-  const averageLevel = data.length === 0
+  const nonZeroTotals = data.map(totalTokens).filter((total) => total !== 0);
+  const averageLevel = nonZeroTotals.length === 0
     ? 0
-    : data.reduce((sum, day) => sum + totalTokens(day), 0) / data.length;
+    : nonZeroTotals.reduce((sum, total) => sum + total, 0) / nonZeroTotals.length;
 
   const option = useMemo<EChartsCoreOption>(() => ({
     animationDuration: 450,

apps/desktop/src/components/cursor/CursorModelEditor.tsx

@@ -1,5 +1,5 @@
 import type { ModelInput, Provider, ProviderInput, ProviderType } from "../../api";
-import { FormField, TextInput } from "../ui/FormControls";
+import { FormField, SecretTextInput, TextInput } from "../ui/FormControls";
 import { Checkbox } from "../ui/Checkbox";
 import { JsonEditor } from "../ui/JsonEditor";
 import { Combobox, MultiCombobox, Select } from "../ui/Select";
@@ -72,16 +72,19 @@ export function CursorModelEditor({ draft, providers, editing, modelOptions, dis
     <div className={styles.grid}>
       {!editing && draft.providerMode === "new" && <>
         <FormField label="Base URL" hint={t("模型服务的 API 根地址，例如 https://api.openai.com/v1。")}><TextInput placeholder="例如：https://api.openai.com/v1" value={draft.provider.base_url} onChange={(event) => setProvider({ base_url: event.target.value })} /></FormField>
-        <FormField label="API Key" hint={t("访问模型服务所需的密钥。")}><TextInput type="password" placeholder="例如：sk-xxxxxx" autoComplete="off" value={draft.provider.api_key ?? ""} onChange={(event) => setProvider({ api_key: event.target.value })} /></FormField>
+        <FormField label="API Key" hint={t("访问模型服务所需的密钥。")}><SecretTextInput placeholder="例如：sk-xxxxxx" autoComplete="off" value={draft.provider.api_key ?? ""} onChange={(event) => setProvider({ api_key: event.target.value })} /></FormField>
       </>}
       <FormField label={t("端点类型")} hint={t("默认继承上游，可为当前模型单独修改。")}><Select ariaLabel={t("端点类型")} value={draft.model.endpoint_type} options={[
         { value: "openai-responses", label: "OpenAI Responses", icon: openAiIcon }, { value: "openai-chat", label: "OpenAI Chat", icon: openAiIcon }, { value: "anthropic", label: "Anthropic", icon: claudeIcon },
       ]} onChange={(endpointType) => setEndpointType(endpointType as ProviderType)} /></FormField>
       {(editing || draft.modelIds.length <= 1) && <FormField label={t("显示名称")} hint={t("仅用于界面展示，不会改变发送给上游的模型名称。")}><TextInput placeholder="例如：GPT-4.1" value={draft.model.display_name} onChange={(event) => setModel({ display_name: event.target.value })} /></FormField>}
-      <FormField className={styles.fullWidth} label={t("模型名称")} hint={editing ? t("可以直接输入模型标识，也可以从当前上游返回的模型列表中选择。") : t("支持选择或输入多个模型；批量添加时显示名称默认使用对应模型名称。")}>{editing
+      <FormField label={t("模型名称")} hint={editing ? t("可以直接输入模型标识，也可以从当前上游返回的模型列表中选择。") : t("支持选择或输入多个模型；批量添加时显示名称默认使用对应模型名称。")}>{editing
         ? <Combobox value={draft.model.model_id} options={modelOptions} placeholder="例如：gpt-4.1" append={<button type="button" className={controls.secondary} disabled={discovering || !canDiscover} onClick={onDiscover}>{discovering ? t("获取中…") : t("获取模型")}</button>} onChange={(model_id) => setModel({ model_id, display_name: draft.model.display_name || model_id })} />
         : <MultiCombobox value={draft.modelIds} options={modelOptions} placeholder="例如：gpt-4.1" append={<button type="button" className={controls.secondary} disabled={discovering || !canDiscover} onClick={onDiscover}>{discovering ? t("获取中…") : t("获取模型")}</button>} onChange={setModelIds} />
       }</FormField>
+      <FormField label={t("自定义上下文")} hint={t("输入 token 数后，将作为额外选项添加到 Cursor 模型的 Context 列表；只有在 Cursor 中选中该选项时才会生效。")}>
+        <TextInput type="number" min={1} step={1} aria-label={t("自定义上下文 tokens")} placeholder={t("例如：272000")} value={draft.model.context_window_tokens ?? ""} onChange={(event) => setModel({ context_window_tokens: event.target.value === "" ? null : Math.trunc(Number(event.target.value)) })} />
+      </FormField>
       <div className={styles.fullWidth}><Checkbox label={t("自定义请求完整地址")} checked={draft.customRequestUrl} onChange={(customRequestUrl) => onChange({ ...draft, customRequestUrl, model: { ...draft.model, request_url: customRequestUrl ? draft.model.request_url : "" } })} /></div>
       {draft.customRequestUrl && <FormField className={styles.fullWidth} label={t("请求完整地址")} hint={t("支持完整 HTTP(S) 地址或以 / 开头、与上游地址组合的相对路径。")}><TextInput placeholder="例如：https://api.example.com/v1/chat/completions" value={draft.model.request_url} onChange={(event) => setModel({ request_url: event.target.value })} /></FormField>}
       {!editing && draft.providerMode === "new" && <>

apps/desktop/src/components/cursor/CursorModelTestResult.tsx

@@ -0,0 +1,35 @@
+import type { ModelConnectivityResult } from "../../api";
+import { Icon } from "../ui/Icon";
+import { TooltipTrigger } from "../ui/TooltipTrigger";
+import { informationOutlineIcon } from "../ui/icons";
+import styles from "./CursorSettings.module.scss";
+
+export type CursorModelTestState =
+  | { status: "success"; result: ModelConnectivityResult }
+  | { status: "error"; error: string };
+
+export function CursorModelTestResult({ state }: { state: CursorModelTestState }) {
+  const success = state.status === "success";
+  const summary = success
+    ? t("速度：{speed} tokens/s", { speed: formatSpeed(state.result.tokens_per_second) })
+    : t("错误：{error}", { error: state.error });
+  const detail = success
+    ? t("速度 {speed} tokens/s · 首字 {firstText} ms · 总耗时 {duration} ms · 输出 {tokens} tokens{estimated} · 返回：{output}", {
+      speed: formatSpeed(state.result.tokens_per_second),
+      firstText: state.result.first_text_ms ?? "--",
+      duration: state.result.duration_ms,
+      tokens: state.result.output_tokens,
+      estimated: state.result.tokens_estimated ? t("（估算）") : "",
+      output: state.result.output || "--",
+    })
+    : t("测试失败：{error}", { error: state.error });
+
+  return <div className={`${styles.testResult} ${success ? styles.testSuccess : styles.testError}`}>
+    <span className={styles.testResultText}>{summary}</span>
+    <TooltipTrigger label={detail}><span className={styles.testResultHint} tabIndex={0}><Icon icon={informationOutlineIcon} size="1.1em" /></span></TooltipTrigger>
+  </div>;
+}
+
+function formatSpeed(value: number) {
+  return Number.isFinite(value) ? value.toFixed(1) : "0.0";
+}

apps/desktop/src/components/cursor/CursorSettings.module.scss

@@ -1,5 +1,10 @@
 @use "../../styles/typography" as type;
 
+.page {
+  display: grid;
+  gap: 16px;
+}
+
 .gate {
   min-height: 250px;
   display: flex;
@@ -81,11 +86,46 @@
   display: flex;
   gap: 2px;
 }
+.testResult {
+  min-width: 0;
+  max-width: 220px;
+  display: flex;
+  align-items: center;
+  gap: 5px;
+  font-size: type.$font-size-xs;
+}
+.testSuccess {
+  color: var(--vscode-testing-iconPassed, #73c991);
+}
+.testError {
+  color: var(--vscode-errorForeground, #f48771);
+}
+.testResultText {
+  min-width: 0;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+.testResultHint {
+  flex: 0 0 auto;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  color: currentColor;
+  cursor: help;
+  &:focus-visible {
+    outline: 1px solid var(--vscode-focusBorder);
+    outline-offset: 2px;
+  }
+}
 .editor {
   display: flex;
   flex-direction: column;
   gap: 14px;
 }
+.editorTestResult {
+  margin-top: 14px;
+}
 .command {
   margin: 0;
   padding: 10px;

apps/desktop/src/components/cursor/TabSettingsCard.module.scss

@@ -0,0 +1,51 @@
+@use "../../styles/typography" as type;
+
+.title {
+  min-width: 0;
+  display: flex;
+  align-items: center;
+  gap: 5px;
+}
+
+.row {
+  min-height: 62px;
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 20px;
+  padding: 12px 16px;
+
+  & + & {
+    border-top: 1px solid var(--vscode-sideBar-border);
+  }
+}
+
+.description {
+  min-width: 0;
+  display: grid;
+  gap: 5px;
+
+  small {
+    color: var(--vscode-descriptionForeground);
+    font-size: type.$font-size-xs;
+  }
+}
+
+.selectControl,
+.addressInput {
+  width: min(300px, 44%);
+  flex: 0 0 auto;
+}
+
+@media (max-width: 620px) {
+  .row {
+    align-items: stretch;
+    flex-direction: column;
+    gap: 10px;
+  }
+
+  .selectControl,
+  .addressInput {
+    width: 100%;
+  }
+}

apps/desktop/src/components/cursor/TabSettingsCard.tsx

@@ -0,0 +1,53 @@
+import cursorIconUrl from "../../assets/icons/cursor.svg";
+import type { TabMode, TabSettings } from "../../api";
+import { Button } from "../ui/Button";
+import { TextInput } from "../ui/FormControls";
+import { Icon } from "../ui/Icon";
+import { Select } from "../ui/Select";
+import { TitledCard } from "../ui/TitledCard";
+import styles from "./TabSettingsCard.module.scss";
+
+export function TabSettingsCard({ settings, saving, onChange, onSave }: {
+  settings: TabSettings;
+  saving: boolean;
+  onChange: (settings: TabSettings) => void;
+  onSave: () => void;
+}) {
+  return <TitledCard
+    title={<div className={styles.title}><Icon src={cursorIconUrl} size="1.1em" /><span>{t("TAB 设置")}</span></div>}
+    action={<Button size="small" variant="primary" disabled={saving} onClick={onSave}>{saving ? t("保存中…") : t("保存")}</Button>}
+  >
+    <div className={styles.row}>
+      <div className={styles.description}>
+        <strong>{t("TAB 选择")}</strong>
+        <small>{t("控制 Cursor TAB 相关接口的连接方式。")}</small>
+      </div>
+      <div className={styles.selectControl}>
+        <Select
+          value={settings.mode}
+          ariaLabel={t("TAB 选择")}
+          options={[
+            { value: "public", label: t("使用公益服务") },
+            { value: "direct", label: t("直连") },
+            { value: "custom", label: t("自定义") },
+          ]}
+          onChange={(mode) => onChange({ ...settings, mode: mode as TabMode })}
+        />
+      </div>
+    </div>
+    {settings.mode === "custom" && <div className={styles.row}>
+      <div className={styles.description}>
+        <strong>{t("TAB 服务地址")}</strong>
+        <small>{t("原接口路径会追加到此服务地址。")}</small>
+      </div>
+      <TextInput
+        className={styles.addressInput}
+        value={settings.address}
+        placeholder="https://tab.leokun.cn"
+        aria-label={t("TAB 服务地址")}
+        onChange={(event) => onChange({ ...settings, address: event.target.value })}
+        onKeyDown={(event) => { if (event.key === "Enter") onSave(); }}
+      />
+    </div>}
+  </TitledCard>;
+}

apps/desktop/src/components/settings/AppLifecycleSettingsCard.module.scss

@@ -21,3 +21,11 @@
     font-size: type.$font-size-xs;
   }
 }
+
+.updateDot {
+  width: 7px;
+  height: 7px;
+  flex: 0 0 7px;
+  background: #ef4444;
+  border-radius: 50%;
+}

apps/desktop/src/components/settings/AppLifecycleSettingsCard.tsx

@@ -1,13 +1,11 @@
-import { useEffect, useRef, useState } from "react";
-import type { Update } from "@tauri-apps/plugin-updater";
+import { useEffect, useState } from "react";
 import {
-  checkForUpdate,
   currentAppVersion,
   hasNativeAppLifecycle,
-  installUpdate,
   readAutostart,
   writeAutostart,
 } from "../../native/appLifecycle";
+import { updateStore, useUpdateStore } from "../../store/updateStore";
 import { Button } from "../ui/Button";
 import { Switch } from "../ui/Switch";
 import { TitledCard } from "../ui/TitledCard";
@@ -17,13 +15,10 @@ import styles from "./AppLifecycleSettingsCard.module.scss";
 export function AppLifecycleSettingsCard() {
   const message = useMessage();
   const native = hasNativeAppLifecycle();
-  const updateRef = useRef<Update | null>(null);
+  const { availableVersion, checking, installing } = useUpdateStore();
   const [version, setVersion] = useState("…");
   const [autostart, setAutostart] = useState(false);
   const [loadingAutostart, setLoadingAutostart] = useState(native);
-  const [checking, setChecking] = useState(false);
-  const [installing, setInstalling] = useState(false);
-  const [availableVersion, setAvailableVersion] = useState<string | null>(null);
 
   useEffect(() => {
     let disposed = false;
@@ -34,12 +29,7 @@ export function AppLifecycleSettingsCard() {
         .catch((cause) => message(cause instanceof Error ? cause.message : String(cause)))
         .finally(() => { if (!disposed) setLoadingAutostart(false); });
     }
-    return () => {
-      disposed = true;
-      const update = updateRef.current;
-      updateRef.current = null;
-      if (update) void update.close();
-    };
+    return () => { disposed = true; };
   }, [message, native]);
 
   const toggleAutostart = async (enabled: boolean) => {
@@ -57,29 +47,17 @@ export function AppLifecycleSettingsCard() {
 
   const checkUpdate = async () => {
     try {
-      setChecking(true);
-      const previous = updateRef.current;
-      updateRef.current = null;
-      if (previous) await previous.close();
-      const update = await checkForUpdate();
-      updateRef.current = update;
-      setAvailableVersion(update?.version ?? null);
-      message(update ? t("发现新版本 {version}", { version: update.version }) : t("当前已是最新版本"));
+      const nextVersion = await updateStore.check();
+      message(nextVersion ? t("发现新版本 {version}", { version: nextVersion }) : t("当前已是最新版本"));
     } catch (cause) {
       message(cause instanceof Error ? cause.message : String(cause));
-    } finally {
-      setChecking(false);
     }
   };
 
   const updateNow = async () => {
-    const update = updateRef.current;
-    if (!update) return;
     try {
-      setInstalling(true);
-      await installUpdate(update);
+      await updateStore.install();
     } catch (cause) {
-      setInstalling(false);
       message(cause instanceof Error ? cause.message : String(cause));
     }
   };
@@ -107,6 +85,7 @@ export function AppLifecycleSettingsCard() {
       {availableVersion
         ? <Button size="small" variant="primary" disabled={installing} onClick={() => void updateNow()}>
             {installing ? t("安装中…") : t("下载并安装")}
+            <span className={styles.updateDot} aria-hidden="true" />
           </Button>
         : <Button size="small" disabled={!native || checking} onClick={() => void checkUpdate()}>
             {checking ? t("检查中…") : t("检查更新")}

apps/desktop/src/components/ui/Controls.module.scss

@@ -11,8 +11,8 @@
 }
 
 .secondary {
-  min-height: 34px;
-  height: 34px;
+  min-height: 30px;
+  height: 30px;
   display: inline-flex;
   align-items: center;
   gap: 6px;

apps/desktop/src/components/ui/FormControls.module.scss

@@ -48,3 +48,31 @@
   height: 34px;
   padding: 0 10px;
 }
+
+.secret {
+  position: relative;
+  width: 100%;
+
+  input {
+    padding-right: 34px;
+  }
+}
+
+.secretToggle {
+  position: absolute;
+  top: 0;
+  right: 0;
+  width: 30px;
+  height: 34px;
+  display: grid;
+  place-items: center;
+  padding: 0;
+  color: var(--vscode-descriptionForeground);
+  background: transparent;
+  border: 0;
+  cursor: pointer;
+
+  &:hover {
+    color: var(--vscode-foreground);
+  }
+}

apps/desktop/src/components/ui/FormControls.tsx

@@ -1,13 +1,23 @@
-import type { InputHTMLAttributes } from "react";
+import { useState, type InputHTMLAttributes } from "react";
 import { Icon } from "./Icon";
 import { TooltipTrigger } from "./TooltipTrigger";
-import { informationOutlineIcon } from "./icons";
+import { eyeIcon, eyeOffIcon, informationOutlineIcon } from "./icons";
 import styles from "./FormControls.module.scss";
 
 export function TextInput(props: InputHTMLAttributes<HTMLInputElement>) {
   return <input {...props} className={[styles.input, props.className].filter(Boolean).join(" ")} />;
 }
 
+export function SecretTextInput({ className, ...props }: InputHTMLAttributes<HTMLInputElement>) {
+  const [visible, setVisible] = useState(false);
+  return <div className={styles.secret}>
+    <input {...props} type={visible ? "text" : "password"} className={[styles.input, className].filter(Boolean).join(" ")} />
+    <button type="button" className={styles.secretToggle} aria-label={visible ? t("隐藏 API Key") : t("显示 API Key")} onClick={() => setVisible((current) => !current)}>
+      <Icon icon={visible ? eyeOffIcon : eyeIcon} size="1.1em" />
+    </button>
+  </div>;
+}
+
 export function FormField({ label, hint, className, children }: { label: string; hint?: string; className?: string; children: React.ReactNode }) {
   return <label className={[styles.field, className].filter(Boolean).join(" ")}>
     <div className={styles.label}>

apps/desktop/src/components/ui/MessageProvider.module.scss

@@ -2,7 +2,7 @@
 
 .region {
   position: fixed;
-  z-index: 11000;
+  z-index: 16000;
   top: calc(48px + env(safe-area-inset-top));
   right: 0;
   left: 0;

apps/desktop/src/components/ui/Modal.tsx

@@ -4,7 +4,7 @@ import { ScrollableContent } from "../virtual/ScrollableContent";
 import controls from "./Controls.module.scss";
 import styles from "./Modal.module.scss";
 
-export function Modal({ id, open, title, children, busy, wide, onClose, onSubmit, closeLabel = t("取消"), submitLabel = t("保存") }: { id?: string; open: boolean; title: string; children: ReactNode; busy?: boolean; wide?: boolean; onClose: () => void; onSubmit?: () => void; closeLabel?: string; submitLabel?: string }) {
+export function Modal({ id, open, title, children, busy, wide, onClose, onSubmit, secondaryAction, closeLabel = t("取消"), submitLabel = t("保存") }: { id?: string; open: boolean; title: string; children: ReactNode; busy?: boolean; wide?: boolean; onClose: () => void; onSubmit?: () => void; secondaryAction?: ReactNode; closeLabel?: string; submitLabel?: string }) {
   const dialog = useRef<HTMLDivElement>(null);
   const closeRef = useRef(onClose);
   const busyRef = useRef(Boolean(busy));
@@ -29,6 +29,7 @@ export function Modal({ id, open, title, children, busy, wide, onClose, onSubmit
       <ScrollableContent alwaysShowVertical className={styles.body} viewportClassName={styles.bodyViewport} contentClassName={styles.bodyContent}>{children}</ScrollableContent>
       <footer>
         <button type="button" className={controls.primary} disabled={busy} onClick={onClose}>{closeLabel}</button>
+        {secondaryAction}
         {onSubmit && <button type="button" className={controls.primary} disabled={busy} onClick={onSubmit}>{busy ? t("处理中…") : submitLabel}</button>}
       </footer>
     </div>

apps/desktop/src/components/ui/Select.tsx

@@ -178,7 +178,7 @@ export function MultiCombobox({ value, options = [], placeholder, disabled, appe
   return <div className={styles.comboRow}><div ref={root} className={styles.multiCombo} data-open={open || undefined}>
     <div className={styles.multiValues}>
       {value.length > 0 && <span className={styles.multiCount}>{t("已选择 {count} 个", { count: value.length })}</span>}
-      <input ref={input} value={query} placeholder={value.length ? t("继续选择或输入") : placeholder} disabled={disabled} role="combobox" aria-haspopup="listbox" aria-controls={open ? menuId : undefined} aria-expanded={open} aria-autocomplete="list" onFocus={() => { if (options.length) setOpen(true); }} onChange={(event) => { setQuery(event.target.value); setActive(0); if (options.length) setOpen(true); }} onKeyDown={(event) => {
+      <input ref={input} value={query} placeholder={value.length ? t("继续选择或输入") : placeholder} disabled={disabled} role="combobox" aria-haspopup="listbox" aria-controls={open ? menuId : undefined} aria-expanded={open} aria-autocomplete="list" onFocus={() => { if (options.length) setOpen(true); }} onBlur={() => add(query)} onChange={(event) => { setQuery(event.target.value); setActive(0); if (options.length) setOpen(true); }} onKeyDown={(event) => {
         if (event.key === "ArrowDown") { event.preventDefault(); move(1); }
         if (event.key === "ArrowUp") { event.preventDefault(); move(-1); }
         if (event.key === "Enter") {

apps/desktop/src/components/ui/icons.ts

@@ -27,6 +27,7 @@ export const settingsIcon = icon('<path fill="currentColor" fill-rule="evenodd"
 export const addIcon = icon('<path fill="currentColor" d="M19 13h-6v6h-2v-6H5v-2h6V5h2v6h6z"/>'); // mdi:plus
 export const editIcon = icon('<path fill="currentColor" d="m14.06 9l.94.94L5.92 19H5v-.92zm3.6-6c-.25 0-.51.1-.7.29l-1.83 1.83l3.75 3.75l1.83-1.83c.39-.39.39-1.04 0-1.41l-2.34-2.34c-.2-.2-.45-.29-.71-.29m-3.6 3.19L3 17.25V21h3.75L17.81 9.94z"/>'); // mdi:pencil-outline
 export const eyeIcon = icon('<path fill="currentColor" d="M12 9a3 3 0 0 1 3 3a3 3 0 0 1-3 3a3 3 0 0 1-3-3a3 3 0 0 1 3-3m0-4.5c5 0 9.27 3.11 11 7.5c-1.73 4.39-6 7.5-11 7.5S2.73 16.39 1 12c1.73-4.39 6-7.5 11-7.5M3.18 12a9.821 9.821 0 0 0 17.64 0a9.821 9.821 0 0 0-17.64 0"/>'); // mdi:eye-outline
+export const eyeOffIcon = icon('<path fill="currentColor" d="M2 5.27L3.28 4L20 20.72L18.73 22l-3.08-3.08c-1.15.38-2.37.58-3.65.58c-5 0-9.27-3.11-11-7.5c.69-1.76 1.79-3.31 3.19-4.54zM12 9a3 3 0 0 1 3 3a3 3 0 0 1-.17 1L11 9.17A3 3 0 0 1 12 9m0-4.5c5 0 9.27 3.11 11 7.5a11.8 11.8 0 0 1-4 5.19l-1.42-1.43A9.86 9.86 0 0 0 20.82 12A9.82 9.82 0 0 0 12 6.5c-1.09 0-2.16.18-3.16.5L7.3 5.47c1.44-.62 3.03-.97 4.7-.97M3.18 12A9.82 9.82 0 0 0 12 17.5c.69 0 1.37-.07 2-.21L11.72 15A3.064 3.064 0 0 1 9 12.28L5.6 8.87c-.99.85-1.82 1.91-2.42 3.13"/>'); // mdi:eye-off-outline
 export const informationOutlineIcon = icon('<path fill="currentColor" d="M11 9h2V7h-2m1 13c-4.41 0-8-3.59-8-8s3.59-8 8-8s8 3.59 8 8s-3.59 8-8 8m0-18A10 10 0 0 0 2 12a10 10 0 0 0 10 10a10 10 0 0 0 10-10A10 10 0 0 0 12 2m-1 15h2v-6h-2z"/>'); // mdi:information-outline
 export const cilBadgeIcon = icon('<path fill="currentColor" d="m328.375 384l3.698 74.999l-75.862-52.719l-76.287 52.769L183.625 384h-32.039l-5.522 112h36.692l73.413-50.78L329.242 496h36.694l-5.522-112zm87.034-229.086l-2.194-48.054L372.7 80.933l-25.932-40.519l-48.055-2.2L256 16.093l-42.713 22.126l-48.055 2.2L139.3 80.933L98.785 106.86l-2.194 48.054l-22.127 42.714l22.127 42.715l2.2 48.053l40.509 25.927l25.928 40.52l48.055 2.195L256 379.164l42.713-22.126l48.055-2.195l25.928-40.52l40.518-25.923l2.195-48.053l22.127-42.715Zm-31.646 76.949L382 270.377l-32.475 20.78l-20.78 32.475l-38.515 1.76L256 343.125l-34.234-17.733l-38.515-1.76l-20.78-32.475L130 270.377l-1.759-38.514l-17.741-34.235l17.737-34.228L130 124.88l32.471-20.78l20.78-32.474l38.515-1.76L256 52.132l34.234 17.733l38.515 1.76l20.78 32.474L382 124.88l1.759 38.515l17.741 34.233Z"/>', 512, 512); // cil:badge
 export const refreshIcon = icon('<path fill="currentColor" d="M17.65 6.35A7.96 7.96 0 0 0 12 4a8 8 0 0 0-8 8a8 8 0 0 0 8 8c3.73 0 6.84-2.55 7.73-6h-2.08A5.99 5.99 0 0 1 12 18a6 6 0 0 1-6-6a6 6 0 0 1 6-6c1.66 0 3.14.69 4.22 1.78L13 11h7V4z"/>'); // mdi:refresh

apps/desktop/src/components/virtual/VirtualListEngine.tsx

@@ -174,6 +174,7 @@ export function VirtualList<TItem>(props: VirtualListProps<TItem>) {
   const shouldResetScrollRef = useRef(false)
   const scrollApiRef = useRef<ScrollAreaApi | null>(null)
   const scrollStateRef = useRef<ScrollAreaState | null>(null)
+  const contentElementRef = useRef<HTMLDivElement | null>(null)
   const spacerRef = useRef<HTMLDivElement | null>(null)
   const [contentInsets, setContentInsets] = useState<ContentInsets>({
     top: 0,
@@ -192,7 +193,8 @@ export function VirtualList<TItem>(props: VirtualListProps<TItem>) {
   })
   const [, forceUpdate] = useState(0)
 
-  const setContentRef = useCallback((node: HTMLDivElement | null) => {
+  const readContentInsets = useCallback(() => {
+    const node = contentElementRef.current
     const styles = node ? getComputedStyle(node) : null
     const nextInsets = {
       top: styles ? Number.parseFloat(styles.paddingTop) || 0 : 0,
@@ -205,6 +207,26 @@ export function VirtualList<TItem>(props: VirtualListProps<TItem>) {
     )
   }, [])
 
+  const setContentRef = useCallback((node: HTMLDivElement | null) => {
+    contentElementRef.current = node
+    readContentInsets()
+  }, [readContentInsets])
+
+  useLayoutEffect(() => {
+    const node = contentElementRef.current
+    if (!node) return
+
+    readContentInsets()
+    const resizeObserver = new ResizeObserver(readContentInsets)
+    resizeObserver.observe(node)
+    const frame = requestAnimationFrame(readContentInsets)
+
+    return () => {
+      cancelAnimationFrame(frame)
+      resizeObserver.disconnect()
+    }
+  }, [readContentInsets])
+
   const contentInsetTop = contentInsets.top
 
   if (!scrollStateRef.current) {

apps/desktop/src/i18n/locales/en-US.json

@@ -5,6 +5,7 @@
   "028a4de61bff743d": "Regular input: {tokens} × ${price}/1M = {cost}",
   "03ff62ab4b818492": "Cache write: {tokens} × ${price}/1M = {cost}",
   "051836569928a9f9": "Edit",
+  "05468af47054d488": "Connectivity test for {model} succeeded ({duration} ms)",
   "0580e0a99a6f1afc": "Artifacts",
   "05912a17829faacc": "All call records and detailed traces will be deleted. Providers, models, CA, and application settings are unaffected. This action cannot be undone.",
   "076832c1b2de22c3": "Cache write: {tokens}",
@@ -22,9 +23,11 @@
   "0d5e2bdb15579fc4": "Messages",
   "0e41f8e3d59ec47b": "Storage management",
   "0ec1e85b0c3cfa65": "Call details",
+  "105a9082c346f958": "Testing…",
   "124be3f86f197802": "Token usage",
   "12ae77e6202d063e": "Custom Headers",
   "12b86e89701b0241": "Enter a model name and press Enter to add it",
+  "133340e53175128a": "Test all",
   "13a9ac7a68c5fd96": "The CA is stored only on this device and is used to securely inspect Cursor HTTPS requests.",
   "13b61c5f697b6700": "Cache hit rate",
   "146da2e2a991493e": "Fetching…",
@@ -56,6 +59,7 @@
   "2eb2bf7c6597ab9a": "Detailed records",
   "2f1b67cdaa23351a": "Custom full request URL",
   "2f4a361f878176d1": "{label} must be valid JSON",
+  "2f4a9609285d8f49": "TAB settings saved",
   "2f5f1d6fbfb061ed": "Not set",
   "2f6416a2c424856b": "Final request URL",
   "2f7dec3be28d7597": "{count} selected",
@@ -75,13 +79,14 @@
   "3cfae5728b92b334": "Token usage: {tokens}",
   "3d13868593ae4eeb": "Display language",
   "3da0bf1610ff5db5": "Recommended",
+  "3f6c25aa329163a4": "The original endpoint path is appended to this service address.",
   "3fd47edce45b3603": "Close",
   "4125fc7ba333524c": "Default light",
   "42655ed8e4108ae2": "Input (non-cached)",
   "42a1d9e5b037c210": "Bytes",
   "42aa8e01e98c0d8c": "Total duration",
   "43cb41d62de2d179": "Proxy requires authentication",
-  "4510b63fdcf7451f": "Cursor Assistant v0.1.0",
+  "461d6a57900c2ed7": "Connectivity test failed: {error}",
   "470049252e54de6a": "Success rate: {rate}",
   "48b970b568a7f8f9": "Proxy settings",
   "48d8db17bae06246": "{count} total",
@@ -92,14 +97,21 @@
   "4d0680f9efaef147": "Unread",
   "4eafa9e925b30bcd": "Custom",
   "51d04bc3d286f018": "Last calendar day",
+  "51de3bcec137ab1b": "Connectivity tests succeeded for all {count} models",
   "5228358a6db59fe7": "For example: now, 2026-08-23 18:00",
+  "5401344227e49e2f": "TAB settings",
   "54644705e9c61009": "Port settings",
   "54c53e5fe791d1f3": "Initialize CA",
   "550eddc3c7fefa99": "Sponsored",
   "555737734a6371e6": "Delete provider",
   "56432ba297009bdc": "Initialize the CA first",
+  "59346e82b3dd2998": "TAB service address",
+  "5a3bd99fa69a40c1": "Use public service",
   "5ae715656ffbc35d": "Merge into the request body for every model from this provider.",
+  "5b17f59d33bde39e": "Error: {error}",
   "5c55a67935af8f45": "All",
+  "5cca0b7972a11f3a": "Custom context must be an integer greater than 0",
+  "5d59857bf039cac9": "Cursor Assistant v{version}",
   "5f8d556a9c47da3c": "Launch at login disabled",
   "5f9acfb945229062": "Are you sure you no longer want to see this ad?",
   "5fd2ec5a6e9b654c": "Total: {cost}",
@@ -111,6 +123,7 @@
   "63c73c4730f4473e": "Apply",
   "6478a5f1218c484e": "Use the desktop app to copy to the system clipboard",
   "651f274470153a05": "Software updates",
+  "652ec5d40c29fd6a": "Speed {speed} tokens/s · first token {firstText} ms · total {duration} ms · output {tokens} tokens{estimated} · response: {output}",
   "653b123c956d3bcb": "Calls",
   "656ab25e264cc4e4": "No models are available to Cursor yet",
   "65a6318e07ec1e07": "Tools",
@@ -118,6 +131,7 @@
   "68152165b3348852": "Select the request protocol used by this provider.",
   "68ea5dd4d7af20e6": "System settings",
   "6a9906c79f26c0ba": "Start time",
+  "6aa8f49cc992dfd7": "Test",
   "6ae80538c2b2572d": "Minimize window",
   "6d1876364ac6457d": "Proxy mode",
   "6e86570183c3cdd0": "You're up to date",
@@ -132,6 +146,7 @@
   "7a3cec4ca715de80": "Call statistics",
   "7ba2d6728fe2531b": "Confirm clear",
   "7d9f043f8f7ab45c": "Version {version} is available in Settings",
+  "7e0891860c9e6374": "TAB service address is required",
   "7e1845870b528392": "Enable model",
   "7e1f06318e80c3af": "Statistics cleared",
   "7e9ab9ada2cbf2cb": "Name and Base URL are required",
@@ -142,6 +157,7 @@
   "83fcfb4c1f2c1641": "Fetch models",
   "842b9f11cdd96bda": "Launch at login",
   "84924374710e03bd": "Base URL must be a valid URL",
+  "86b7355ec3bd55ef": "Hide API Key",
   "8716e1344b0daddb": "Cursor official",
   "878a8ab176429a86": "View instructions",
   "883cc47637fe70f3": "Custom request headers appended to every request for this provider. Values must be strings.",
@@ -149,6 +165,7 @@
   "8b6ff498515bcc2f": "Time",
   "8d0c47eb9eac2d34": "Call type",
   "8df48894086d6fbd": "Reason (optional)",
+  "8f9b0d6cc477d334": "Choose how Cursor connects to TAB endpoints.",
   "90800c48a1dd0655": "{label} must be a JSON object",
   "919cb0ce0c8db4e7": "Leave blank to keep the current password",
   "91aaf184cfc17ffd": "Overview",
@@ -158,11 +175,13 @@
   "94803f35c825e47a": "Full request URL",
   "94df1e7f04815daf": "Used only for display; does not change the model name sent to the provider.",
   "95f76d30c25d5eda": "CA installation is not supported on this system",
+  "966498853d801a52": "TAB connection",
   "9850ed41a5bfbb0c": "{count} selected",
   "997ec8201c2adeda": "Open terminal to install CA",
   "9a026819dd1af5c5": "Enter a model identifier directly or select one returned by the current provider.",
   "9ac0ac940982895d": "Select provider",
   "9c41b3a9e12ac994": "Reasoning effort",
+  "9d8f2ef4e85ea665": "Custom context tokens",
   "9e46da6923836182": "For example: 2026-08-23 09:00, 1 hour ago",
   "9f6fee1aba17a565": "Language",
   "9fb48101d237ff96": "Last week",
@@ -175,6 +194,7 @@
   "a3ab741ceb188e9e": "Request content was not recorded. Enable detailed records and try again.",
   "a49ffd73bc85333d": "Average",
   "a621ab606db2a11f": "Password",
+  "a693d69af48bfe48": "Save and test",
   "a748cc074f78de00": "View details",
   "a98585871c5313ff": "Display name",
   "a9ab292ea9feecdc": "Provider",
@@ -186,17 +206,20 @@
   "ad6a60ee93d3ba3e": "Loading call details…",
   "ae2d0b7f79cea4a3": "Model output: {tokens} × ${price}/1M = {cost}",
   "aee88743413144a2": "Refresh",
+  "b06325c5660f0c29": "Direct",
   "b4411558b932266f": "Provider type",
   "b502b1d414664337": "Prompt: {tokens}",
   "b5141d3d19e9a048": "Yes",
   "b710ec36ad312918": "The model service API root URL, for example https://api.openai.com/v1.",
   "b75a46aad3e7c132": "Non-cached input: {tokens}",
   "b79354009c614ae9": "Statistics",
+  "b86967982067d295": " (estimated)",
   "b89a0e4584f27ab5": "Open terminal",
   "b8c9b486c83b5778": "Hide ad",
   "b9670c85a4ab939e": "Route",
   "b97ad406809572e1": "Enable reasoning",
   "b9af2de88d903be7": "Proxy address",
+  "baff6c144180b185": "Connectivity tests completed: {successful} succeeded, {failed} failed",
   "bb2b7736433ae867": "Cursor tracing",
   "bb7efdcb6af6e805": "Default dark",
   "bda62ce1d5e4ace9": "Tell us why",
@@ -206,6 +229,7 @@
   "c1e98892a77f7a19": "{count} per page",
   "c3760858cdb6d9f4": "Request body",
   "c62a58459251b02c": "Image generation",
+  "c6cc835023617457": "Custom context",
   "c7ea2c9bc43134bd": "Edit model",
   "c8df3c14a003bfcd": "Unable to load call details",
   "c98e118e0a43f078": "Model",
@@ -218,6 +242,7 @@
   "cfe999e50be8ef54": "Whether the model declares image-generation support.",
   "d0bfccc77315d887": "Last month",
   "d15a909c3490a7e0": "Endpoint type",
+  "d27db596b73a0a66": "For example: 272000",
   "d2d648bd1c94b7f9": "Authentication",
   "d30d35c5ec4a888d": "Select or enter at least one model",
   "d34335433395cd3a": "Start Cursor BYOK automatically after signing in.",
@@ -236,16 +261,19 @@
   "df8b71c74d9b8478": "Response stream",
   "dfb802238b38fbd4": "Enabled",
   "e025f1ff71996425": "Set",
+  "e0fae77446a389a3": "Speed: {speed} tokens/s",
   "e1295adecbb77755": "Close ad",
   "e14115de7f7c5795": "Token usage over the past year",
   "e14a5eee9b0b0f9f": "Whether the model declares reasoning support.",
   "e14f20d572c02611": "Provider call sequence",
   "e18516550b9a5105": "No usage",
+  "e24ebe4a866d69bf": "Test failed: {error}",
   "e25bf3f419bb68f0": "Call history",
   "e3fee05f688708b4": "LLM calls",
   "e5043c7a2b408271": "Last 10 minutes",
   "e59ae97924d62f01": "First page",
   "e5b9961a0d5242e3": "Port settings saved. Restart the app to apply them.",
+  "e671f8c7598139ef": "After entering a token count, it is added as an extra option to the Cursor model Context list. It only takes effect when selected in Cursor.",
   "e6ca887f22288cde": "Model ID and display name are required",
   "e77e3d58b0dcffaa": "Duration",
   "e828bd3a0151edc2": "The local CA must be trusted by the system",
@@ -259,6 +287,7 @@
   "ee239f3943293f87": "Sunday",
   "ee6b89a6a740a4c4": "If a port is occupied, a new random port is selected and saved automatically. Restart the app after changing these settings.",
   "f04c91a6bc3a6926": "Extra parameters",
+  "f2bdc88464c51c2e": "Show API Key",
   "f396118b8afd2a21": "Cursor interception is active. Add a provider and its model configuration to use BYOK models.",
   "f3a76d896853c1df": "Miss",
   "f4694c46b1e19602": "Final request type",

apps/desktop/src/i18n/locales/zh-CN.json

@@ -5,6 +5,7 @@
   "028a4de61bff743d": "普通输入：{tokens} × ${price}/1M = {cost}",
   "03ff62ab4b818492": "缓存写入：{tokens} × ${price}/1M = {cost}",
   "051836569928a9f9": "编辑",
+  "05468af47054d488": "模型 {model} 连通性测试成功（{duration} ms）",
   "0580e0a99a6f1afc": "工件数",
   "05912a17829faacc": "所有调用记录和详细追踪数据都会被删除。供应商、模型、CA 和应用设置不会受到影响，此操作无法撤销。",
   "076832c1b2de22c3": "缓存写入：{tokens}",
@@ -22,9 +23,11 @@
   "0d5e2bdb15579fc4": "消息数",
   "0e41f8e3d59ec47b": "存储管理",
   "0ec1e85b0c3cfa65": "调用详情",
+  "105a9082c346f958": "测试中…",
   "124be3f86f197802": "Token 消耗",
   "12ae77e6202d063e": "自定义 Headers",
   "12b86e89701b0241": "输入模型名称后按 Enter 添加",
+  "133340e53175128a": "一键测试",
   "13a9ac7a68c5fd96": "CA 仅保存在本机，用于安全解析 Cursor 的 HTTPS 请求。",
   "13b61c5f697b6700": "缓存命中率",
   "146da2e2a991493e": "获取中…",
@@ -56,6 +59,7 @@
   "2eb2bf7c6597ab9a": "详细记录",
   "2f1b67cdaa23351a": "自定义请求完整地址",
   "2f4a361f878176d1": "{label} 必须是有效 JSON",
+  "2f4a9609285d8f49": "TAB 设置已保存",
   "2f5f1d6fbfb061ed": "未设置",
   "2f6416a2c424856b": "最终请求地址",
   "2f7dec3be28d7597": "已选择 {count} 个",
@@ -75,13 +79,14 @@
   "3cfae5728b92b334": "Token 用量：{tokens}",
   "3d13868593ae4eeb": "界面语言",
   "3da0bf1610ff5db5": "推荐内容",
+  "3f6c25aa329163a4": "原接口路径会追加到此服务地址。",
   "3fd47edce45b3603": "关闭",
   "4125fc7ba333524c": "默认亮色",
   "42655ed8e4108ae2": "输入（非缓存）",
   "42a1d9e5b037c210": "字节数",
   "42aa8e01e98c0d8c": "总耗时",
   "43cb41d62de2d179": "代理需要认证",
-  "4510b63fdcf7451f": "Cursor 助手 v0.1.0",
+  "461d6a57900c2ed7": "连通性测试失败：{error}",
   "470049252e54de6a": "成功占比：{rate}",
   "48b970b568a7f8f9": "代理设置",
   "48d8db17bae06246": "共 {count} 条",
@@ -92,14 +97,21 @@
   "4d0680f9efaef147": "未读",
   "4eafa9e925b30bcd": "自定义",
   "51d04bc3d286f018": "近1自然日",
+  "51de3bcec137ab1b": "全部 {count} 个模型连通性测试成功",
   "5228358a6db59fe7": "如：现在、2026-08-23 18:00",
+  "5401344227e49e2f": "TAB 设置",
   "54644705e9c61009": "端口设置",
   "54c53e5fe791d1f3": "初始化 CA",
   "550eddc3c7fefa99": "推广",
   "555737734a6371e6": "删除上游",
   "56432ba297009bdc": "请先初始化 CA",
+  "59346e82b3dd2998": "TAB 服务地址",
+  "5a3bd99fa69a40c1": "使用公益服务",
   "5ae715656ffbc35d": "合并到该上游所有模型的请求体。",
+  "5b17f59d33bde39e": "错误：{error}",
   "5c55a67935af8f45": "全部",
+  "5cca0b7972a11f3a": "自定义上下文必须是大于 0 的整数",
+  "5d59857bf039cac9": "Cursor 助手 v{version}",
   "5f8d556a9c47da3c": "已关闭开机启动",
   "5f9acfb945229062": "你确认不想再看到此广告吗？",
   "5fd2ec5a6e9b654c": "合计：{cost}",
@@ -111,6 +123,7 @@
   "63c73c4730f4473e": "应用",
   "6478a5f1218c484e": "请在桌面应用中复制到系统剪贴板",
   "651f274470153a05": "软件更新",
+  "652ec5d40c29fd6a": "速度 {speed} tokens/s · 首字 {firstText} ms · 总耗时 {duration} ms · 输出 {tokens} tokens{estimated} · 返回：{output}",
   "653b123c956d3bcb": "调用",
   "656ab25e264cc4e4": "还没有可供 Cursor 使用的模型",
   "65a6318e07ec1e07": "工具数",
@@ -118,6 +131,7 @@
   "68152165b3348852": "选择上游服务使用的请求协议。",
   "68ea5dd4d7af20e6": "系统设置",
   "6a9906c79f26c0ba": "开始时间",
+  "6aa8f49cc992dfd7": "测试",
   "6ae80538c2b2572d": "最小化窗口",
   "6d1876364ac6457d": "代理方式",
   "6e86570183c3cdd0": "当前已是最新版本",
@@ -132,6 +146,7 @@
   "7a3cec4ca715de80": "调用统计",
   "7ba2d6728fe2531b": "确认清理",
   "7d9f043f8f7ab45c": "发现新版本 {version}，可在设置中安装",
+  "7e0891860c9e6374": "TAB 服务地址不能为空",
   "7e1845870b528392": "启用模型",
   "7e1f06318e80c3af": "统计数据已清理",
   "7e9ab9ada2cbf2cb": "名称和 Base URL 不能为空",
@@ -142,6 +157,7 @@
   "83fcfb4c1f2c1641": "获取模型",
   "842b9f11cdd96bda": "开机启动",
   "84924374710e03bd": "Base URL 必须是有效地址",
+  "86b7355ec3bd55ef": "隐藏 API Key",
   "8716e1344b0daddb": "Cursor 官方",
   "878a8ab176429a86": "查看说明",
   "883cc47637fe70f3": "附加到该上游所有请求的自定义请求头，值必须是字符串。",
@@ -149,6 +165,7 @@
   "8b6ff498515bcc2f": "时间",
   "8d0c47eb9eac2d34": "调用类型",
   "8df48894086d6fbd": "原因（可选）",
+  "8f9b0d6cc477d334": "控制 Cursor TAB 相关接口的连接方式。",
   "90800c48a1dd0655": "{label} 必须是 JSON 对象",
   "919cb0ce0c8db4e7": "留空表示保留当前密码",
   "91aaf184cfc17ffd": "数据概览",
@@ -158,11 +175,13 @@
   "94803f35c825e47a": "请求完整地址",
   "94df1e7f04815daf": "仅用于界面展示，不会改变发送给上游的模型名称。",
   "95f76d30c25d5eda": "当前系统暂不支持安装 CA",
+  "966498853d801a52": "TAB 选择",
   "9850ed41a5bfbb0c": "已选 {count} 项",
   "997ec8201c2adeda": "打开终端安装 CA",
   "9a026819dd1af5c5": "可以直接输入模型标识，也可以从当前上游返回的模型列表中选择。",
   "9ac0ac940982895d": "选择上游",
   "9c41b3a9e12ac994": "思考强度",
+  "9d8f2ef4e85ea665": "自定义上下文 tokens",
   "9e46da6923836182": "如：2026-08-23 09:00、1小时前",
   "9f6fee1aba17a565": "语言",
   "9fb48101d237ff96": "近一周",
@@ -175,6 +194,7 @@
   "a3ab741ceb188e9e": "未记录请求内容，请开启详细记录后重试。",
   "a49ffd73bc85333d": "平均",
   "a621ab606db2a11f": "密码",
+  "a693d69af48bfe48": "保存并测试",
   "a748cc074f78de00": "查看详情",
   "a98585871c5313ff": "显示名称",
   "a9ab292ea9feecdc": "上游",
@@ -186,17 +206,20 @@
   "ad6a60ee93d3ba3e": "正在加载调用详情…",
   "ae2d0b7f79cea4a3": "模型输出：{tokens} × ${price}/1M = {cost}",
   "aee88743413144a2": "刷新",
+  "b06325c5660f0c29": "直连",
   "b4411558b932266f": "上游类型",
   "b502b1d414664337": "提示词：{tokens}",
   "b5141d3d19e9a048": "是",
   "b710ec36ad312918": "模型服务的 API 根地址，例如 https://api.openai.com/v1。",
   "b75a46aad3e7c132": "非缓存输入：{tokens}",
   "b79354009c614ae9": "统计数据",
+  "b86967982067d295": "（估算）",
   "b89a0e4584f27ab5": "打开终端",
   "b8c9b486c83b5778": "不再显示广告",
   "b9670c85a4ab939e": "路由",
   "b97ad406809572e1": "启用推理",
   "b9af2de88d903be7": "代理地址",
+  "baff6c144180b185": "连通性测试完成：成功 {successful}，失败 {failed}",
   "bb2b7736433ae867": "Cursor 追踪",
   "bb7efdcb6af6e805": "默认暗色",
   "bda62ce1d5e4ace9": "可以告诉我们原因",
@@ -206,6 +229,7 @@
   "c1e98892a77f7a19": "{count} 条/页",
   "c3760858cdb6d9f4": "请求体",
   "c62a58459251b02c": "图片生成",
+  "c6cc835023617457": "自定义上下文",
   "c7ea2c9bc43134bd": "编辑模型",
   "c8df3c14a003bfcd": "无法加载调用详情",
   "c98e118e0a43f078": "模型",
@@ -218,6 +242,7 @@
   "cfe999e50be8ef54": "是否声明模型支持图片生成。",
   "d0bfccc77315d887": "近一个月",
   "d15a909c3490a7e0": "端点类型",
+  "d27db596b73a0a66": "例如：272000",
   "d2d648bd1c94b7f9": "认证",
   "d30d35c5ec4a888d": "请至少选择或输入一个模型",
   "d34335433395cd3a": "登录系统后自动启动 Cursor BYOK。",
@@ -236,16 +261,19 @@
   "df8b71c74d9b8478": "响应流",
   "dfb802238b38fbd4": "已启用",
   "e025f1ff71996425": "已设置",
+  "e0fae77446a389a3": "速度：{speed} tokens/s",
   "e1295adecbb77755": "关闭广告",
   "e14115de7f7c5795": "过去一年的 Token 用量",
   "e14a5eee9b0b0f9f": "是否声明模型支持推理能力。",
   "e14f20d572c02611": "上游调用序号",
   "e18516550b9a5105": "无用量",
+  "e24ebe4a866d69bf": "测试失败：{error}",
   "e25bf3f419bb68f0": "调用详细",
   "e3fee05f688708b4": "LLM 调用",
   "e5043c7a2b408271": "近10分钟",
   "e59ae97924d62f01": "第一页",
   "e5b9961a0d5242e3": "端口设置已保存，重启软件后生效",
+  "e671f8c7598139ef": "输入 token 数后，将作为额外选项添加到 Cursor 模型的 Context 列表；只有在 Cursor 中选中该选项时才会生效。",
   "e6ca887f22288cde": "Model ID 和显示名称不能为空",
   "e77e3d58b0dcffaa": "耗时",
   "e828bd3a0151edc2": "需要在系统中信任本地 CA",
@@ -259,6 +287,7 @@
   "ee239f3943293f87": "周日",
   "ee6b89a6a740a4c4": "端口被占用时会自动选择新的随机端口并保存。修改后需要重启软件才会生效。",
   "f04c91a6bc3a6926": "额外参数",
+  "f2bdc88464c51c2e": "显示 API Key",
   "f396118b8afd2a21": "Cursor 接管已生效；添加上游及其模型配置后即可使用 BYOK 模型。",
   "f3a76d896853c1df": "未命中",
   "f4694c46b1e19602": "最终请求类型",

apps/desktop/src/layouts/AppHeader.tsx

@@ -1,4 +1,6 @@
+import { useEffect, useState } from "react";
 import appIcon from "../../src-tauri/icons/32x32.png";
+import { currentAppVersion } from "../native/appLifecycle";
 import type { DesktopPlatform } from "./AppFrame";
 import { WindowControls } from "./WindowControls";
 import styles from "./AppHeader.module.scss";
@@ -10,6 +12,15 @@ type AppHeaderProps = {
 
 export function AppHeader({ platform, nativeDesktop }: AppHeaderProps) {
   const showNativeUi = nativeDesktop && platform !== "macos";
+  const [version, setVersion] = useState("…");
+
+  useEffect(() => {
+    let disposed = false;
+    void currentAppVersion().then((next) => {
+      if (!disposed) setVersion(next);
+    });
+    return () => { disposed = true; };
+  }, []);
 
   return <header className={styles.root}>
     <div className={styles.dragLayer} data-tauri-drag-region aria-hidden="true" />
@@ -17,7 +28,7 @@ export function AppHeader({ platform, nativeDesktop }: AppHeaderProps) {
       {showNativeUi && <>
         <div className={styles.identity} aria-label="Cursor BYOK">
           <img src={appIcon} alt="" />
-          <span>{t("Cursor 助手 v0.1.0")}</span>
+          <span>{t("Cursor 助手 v{version}", { version })}</span>
         </div>
         <WindowControls />
       </>}

apps/desktop/src/layouts/AppLayout.module.scss

@@ -83,7 +83,7 @@
   }
 }
 
-.menuUnreadDot {
+.menuIndicatorDot {
   width: 7px;
   height: 7px;
   flex: 0 0 7px;

apps/desktop/src/layouts/AppLayout.tsx

@@ -18,6 +18,7 @@ import { useMessage } from "../components/ui/message";
 import { VirtualList } from "../components/virtual/VirtualList";
 import { useI18n } from "../i18n/store";
 import { appStore, useAppStore } from "../store/appStore";
+import { useUpdateStore } from "../store/updateStore";
 import styles from "./AppLayout.module.scss";
 import { PageActionsTarget } from "./PageActions";
 
@@ -43,10 +44,12 @@ function loadStoredAdIds(key: string): Set<string> {
 
 export function AppLayout() {
   const { busy } = useAppStore();
+  const { availableVersion } = useUpdateStore();
   const { locale } = useI18n();
   const message = useMessage();
   const location = useLocation();
-  const [actionTarget, setActionTarget] = useState<HTMLDivElement | null>(null);
+  const [leftActionTarget, setLeftActionTarget] = useState<HTMLDivElement | null>(null);
+  const [rightActionTarget, setRightActionTarget] = useState<HTMLDivElement | null>(null);
   const [ads, setAds] = useState<AdSlot[]>([]);
   const [activeAd, setActiveAd] = useState<AdSlot | null>(null);
   const [dismissCandidate, setDismissCandidate] = useState<AdSlot | null>(null);
@@ -198,7 +201,7 @@ export function AppLayout() {
                 ? <Icon src={item.icon} size="1.3em" />
                 : <Icon icon={item.icon} size="1.3em" />}
               <span>{item.label}</span>
-              {!tutorialRead && <span className={styles.menuUnreadDot} aria-hidden="true" />}
+              {!tutorialRead && <span className={styles.menuIndicatorDot} aria-hidden="true" />}
             </button>
           </div>
           : <div className={styles.navigationRow} key={item.path}>
@@ -207,6 +210,7 @@ export function AppLayout() {
                 ? <Icon src={item.icon} size="1.3em" />
                 : <Icon icon={item.icon} size="1.3em" />}
               <span>{item.label}</span>
+              {item.path === "/settings" && availableVersion && <span className={styles.menuIndicatorDot} aria-hidden="true" />}
             </NavLink>
           </div>}
         </VirtualList>
@@ -238,13 +242,14 @@ export function AppLayout() {
     <main className={styles.content}>
       <div className={styles.actionRegion}>
         <Card className={styles.actions}>
+          <div ref={setLeftActionTarget} className={styles.pageActions} />
           {location.pathname !== "/" && <TooltipTrigger label={t("刷新")}><button className={controls.iconButton} aria-label={t("刷新")} disabled={busy} onClick={() => void appStore.refresh()}>
             <Icon className={busy ? controls.spin : ""} icon={refreshIcon} size="1.1em" />
           </button></TooltipTrigger>}
-          <div ref={setActionTarget} className={styles.pageActions} />
+          <div ref={setRightActionTarget} className={styles.pageActions} />
         </Card>
       </div>
-      <PageActionsTarget.Provider value={actionTarget}>
+      <PageActionsTarget.Provider value={{ left: leftActionTarget, right: rightActionTarget }}>
         <KeepAliveRouteOutlet
           activeCacheKey={location.pathname}
           include={keptAlivePages}

apps/desktop/src/layouts/PageActions.tsx

@@ -2,10 +2,16 @@ import { createContext, useContext, type ReactNode } from "react";
 import { createPortal } from "react-dom";
 import { useKeepAliveContext } from "keepalive-for-react";
 
-export const PageActionsTarget = createContext<HTMLElement | null>(null);
+type PageActionsTargets = {
+  left: HTMLElement | null;
+  right: HTMLElement | null;
+};
 
-export function PageActions({ children }: { children: ReactNode }) {
-  const target = useContext(PageActionsTarget);
+export const PageActionsTarget = createContext<PageActionsTargets>({ left: null, right: null });
+
+export function PageActions({ children, position = "right" }: { children: ReactNode; position?: "left" | "right" }) {
+  const targets = useContext(PageActionsTarget);
   const { active } = useKeepAliveContext();
+  const target = targets[position];
   return active && target ? createPortal(children, target) : null;
 }

apps/desktop/src/pages/CursorSettingsPage.tsx

@@ -1,7 +1,9 @@
 import { useEffect, useMemo, useState } from "react";
-import { api, type Model, type ProviderSelection } from "../api";
+import { api, type Model, type ProviderSelection, type TabSettings } from "../api";
 import { CursorCaGate, CursorCaProvider, CursorModelGate, CursorModelProvider } from "../components/cursor/CursorGates";
 import { CursorModelEditor, emptyCursorModelDraft, type CursorModelDraft } from "../components/cursor/CursorModelEditor";
+import { CursorModelTestResult, type CursorModelTestState } from "../components/cursor/CursorModelTestResult";
+import { TabSettingsCard } from "../components/cursor/TabSettingsCard";
 import styles from "../components/cursor/CursorSettings.module.scss";
 import { PageContent } from "../components/layout/PageContent";
 import controls from "../components/ui/Controls.module.scss";
@@ -24,12 +26,23 @@ export function CursorSettingsPage() {
   const [caCommand, setCaCommand] = useState<string | null>(null);
   const [waitingForCaRefresh, setWaitingForCaRefresh] = useState(false);
   const [deleting, setDeleting] = useState<Model | null>(null);
+  const [tabDraft, setTabDraft] = useState<TabSettings | null>(null);
+  const [savingTab, setSavingTab] = useState(false);
+  const [testingModelHashes, setTestingModelHashes] = useState<Set<string>>(() => new Set());
+  const [modelTestResults, setModelTestResults] = useState<Map<string, CursorModelTestState>>(() => new Map());
+  const [savingAndTesting, setSavingAndTesting] = useState(false);
+  const [batchTesting, setBatchTesting] = useState(false);
   const grouped = useMemo(() => providers.map((provider) => ({ provider, models: models.filter((model) => model.provider_id === provider.provider_id) })).filter((group) => group.models.length > 0), [providers, models]);
   const caReady = cursorHarness?.ca === "ready";
   useEffect(() => {
     if (!caCommand) return;
     void api.copyCursorText(caCommand);
   }, [caCommand]);
+  useEffect(() => {
+    void api.tabSettings()
+      .then(setTabDraft)
+      .catch((cause) => message(cause instanceof Error ? cause.message : String(cause)));
+  }, [message]);
   const initializeCa = async () => {
     const status = await appStore.initializeCursorCa();
     if (status?.ca === "untrusted" && status.ca_install_command) setCaCommand(status.ca_install_command);
@@ -42,7 +55,7 @@ export function CursorSettingsPage() {
     next.model = {
       model_id: model.model_id, display_name: model.display_name, enabled: model.enabled, sort_order: model.sort_order,
       endpoint_type: model.endpoint_type, request_url: model.request_url,
-      context_window_tokens: null, max_output_tokens: null,
+      context_window_tokens: model.context_window_tokens, max_output_tokens: null,
       reasoning_enabled: model.reasoning_enabled, reasoning_effort: null,
       supports_image_generation: model.supports_image_generation,
     };
@@ -71,29 +84,79 @@ export function CursorSettingsPage() {
   const save = async () => {
     if (!draft) return;
     try {
-      const modelIds = editing
-        ? [draft.model.model_id.trim()]
-        : [...new Set(draft.modelIds.map((modelId) => modelId.trim()).filter(Boolean))];
-      if (!modelIds.length) throw new Error(t("请至少选择或输入一个模型"));
-      if (editing && !draft.model.display_name.trim()) throw new Error(t("Model ID 和显示名称不能为空"));
-      if (draft.customRequestUrl && !draft.model.request_url.trim()) throw new Error(t("请求完整地址不能为空"));
-      const modelInputs = modelIds.map((modelId, index) => ({
-        ...draft.model,
-        model_id: modelId,
-        display_name: modelIds.length === 1 ? draft.model.display_name.trim() || modelId : modelId,
-        sort_order: draft.model.sort_order + index,
-      }));
-      const ok = editing
+      const modelInputs = cursorModelInputs(draft, editing !== null);
+      const saved = editing
         ? await appStore.updateCursorModel(editing.model_hash, modelInputs[0])
         : await appStore.createCursorModels(providerSelection(draft), modelInputs);
-      if (ok) { setDraft(null); setEditing(null); }
+      if (saved) { setDraft(null); setEditing(null); }
     } catch (cause) { message(cause instanceof Error ? cause.message : String(cause)); }
   };
+  const testModel = async (model: Model, notify = true) => {
+    setTestingModelHashes((current) => new Set(current).add(model.model_hash));
+    try {
+      const result = await api.testModel(model.model_hash);
+      setModelTestResults((current) => new Map(current).set(model.model_hash, { status: "success", result }));
+      if (notify) message(t("模型 {model} 连通性测试成功（{duration} ms）", { model: model.display_name, duration: result.duration_ms }));
+      return true;
+    } catch (cause) {
+      const error = cause instanceof Error ? cause.message : String(cause);
+      setModelTestResults((current) => new Map(current).set(model.model_hash, { status: "error", error }));
+      if (notify) message(t("连通性测试失败：{error}", { error }), { duration: 5000 });
+      return false;
+    } finally {
+      setTestingModelHashes((current) => {
+        const next = new Set(current);
+        next.delete(model.model_hash);
+        return next;
+      });
+    }
+  };
+  const testSingleModel = async (model: Model) => {
+    await testModel(model);
+    await appStore.refresh();
+  };
+  const saveAndTest = async () => {
+    if (!draft || !editing) return;
+    setSavingAndTesting(true);
+    try {
+      const [input] = cursorModelInputs(draft, true);
+      const saved = await appStore.updateCursorModel(editing.model_hash, input);
+      if (!saved) {
+        const error = appStore.getSnapshot().error;
+        if (error) message(error);
+        return;
+      }
+      setEditing(saved);
+      await testSingleModel(saved);
+    } catch (cause) {
+      message(cause instanceof Error ? cause.message : String(cause));
+    } finally {
+      setSavingAndTesting(false);
+    }
+  };
+  const testAllModels = async () => {
+    if (!models.length || batchTesting) return;
+    const targets = [...models];
+    setBatchTesting(true);
+    try {
+      const results = await Promise.all(targets.map((model) => testModel(model, false)));
+      await appStore.refresh();
+      const successful = results.filter(Boolean).length;
+      const failed = targets.length - successful;
+      message(failed === 0
+        ? t("全部 {count} 个模型连通性测试成功", { count: targets.length })
+        : t("连通性测试完成：成功 {successful}，失败 {failed}", { successful, failed }), { duration: failed === 0 ? 2400 : 5000 });
+    } finally {
+      setBatchTesting(false);
+    }
+  };
   const list = <div className={styles.groups}>{grouped.map(({ provider, models: childModels }) => <TitledCard key={provider.provider_id} title={<div className={styles.providerTitle}><Icon icon={provider.provider_type === "anthropic" ? claudeIcon : openAiIcon} /><span>{provider.name}</span></div>}>
     <div className={styles.models}>{childModels.map((model) => <div className={styles.modelRow} key={model.model_hash}>
       <div className={styles.modelName}><strong>{model.display_name}</strong><small>{model.model_id} · {model.model_hash}</small></div>
       {/* <span className={styles.badge}>{model.enabled ? t("已启用") : t("已停用")}</span> */}
+      {modelTestResults.get(model.model_hash) && <CursorModelTestResult state={modelTestResults.get(model.model_hash)!} />}
       <div className={styles.rowActions}>
+        <button type="button" className={`${controls.secondary} ${controls.small}`} disabled={testingModelHashes.size > 0 || cursorBusy || batchTesting} onClick={() => void testSingleModel(model)}>{testingModelHashes.has(model.model_hash) ? t("测试中…") : t("测试")}</button>
         <TooltipTrigger label={t("编辑模型")}><button className={controls.iconButton} aria-label={t("编辑模型")} onClick={() => openEdit(model)}><Icon icon={editIcon} size="1.1em" /></button></TooltipTrigger>
         <TooltipTrigger label={t("删除模型")}><button className={`${controls.iconButton} ${controls.danger}`} aria-label={t("删除模型")} onClick={() => setDeleting(model)}><Icon icon={trashIcon} size="1.1em" /></button></TooltipTrigger>
       </div>
@@ -111,17 +174,39 @@ export function CursorSettingsPage() {
     setCaCommand(null);
     setWaitingForCaRefresh(true);
   };
+  const saveTab = async () => {
+    if (!tabDraft) return;
+    try {
+      if (tabDraft.mode === "custom" && !tabDraft.address.trim()) throw new Error(t("TAB 服务地址不能为空"));
+      setSavingTab(true);
+      setTabDraft(await api.setTabSettings(tabDraft));
+      message(t("TAB 设置已保存"));
+    } catch (cause) {
+      message(cause instanceof Error ? cause.message : String(cause));
+    } finally {
+      setSavingTab(false);
+    }
+  };
   const content = <CursorCaProvider><CursorCaGate busy={cursorBusy} waitingForRefresh={waitingForCaRefresh} onInitialize={() => void initializeCa()} onRefresh={() => void refreshCa()}>
-    <CursorModelProvider><CursorModelGate onAdd={openNew}>{list}</CursorModelGate></CursorModelProvider>
+    <div className={styles.page}>
+      {tabDraft && <TabSettingsCard settings={tabDraft} saving={savingTab} onChange={setTabDraft} onSave={() => void saveTab()} />}
+      <CursorModelProvider><CursorModelGate onAdd={openNew}>{list}</CursorModelGate></CursorModelProvider>
+    </div>
   </CursorCaGate></CursorCaProvider>;
 
   return <>
+    {models.length > 0 && <PageActions position="left">
+      <button type="button" className={controls.secondary} disabled={cursorBusy || testingModelHashes.size > 0 || batchTesting} onClick={() => void testAllModels()}>{batchTesting ? t("测试中…") : t("一键测试")}</button>
+    </PageActions>}
     <PageActions>
       <TooltipTrigger label={caReady ? t("添加模型") : t("请先初始化 CA")}><button className={controls.iconButton} aria-label={t("添加模型")} disabled={!caReady || cursorBusy} onClick={openNew}><Icon icon={addIcon} size="1.1em" /></button></TooltipTrigger>
     </PageActions>
-    <PageContent title={t("Cursor 设置")} sections={[{ key: "cursor-settings", estimatedHeight: Math.max(280, models.length * 55 + grouped.length * 62), content }]} />
-    <Modal open={draft !== null} title={editing ? t("编辑模型") : t("添加模型")} busy={cursorBusy} onClose={() => setDraft(null)} onSubmit={() => void save()}>
-      {draft && <CursorModelEditor draft={draft} providers={providers} editing={editing !== null} modelOptions={modelOptions} discovering={discovering} onChange={setDraft} onDiscover={() => void discover()} />}
+    <PageContent title={t("Cursor 设置")} sections={[{ key: "cursor-settings", estimatedHeight: Math.max(430, models.length * 55 + grouped.length * 62 + 145), content }]} />
+    <Modal open={draft !== null} title={editing ? t("编辑模型") : t("添加模型")} busy={cursorBusy || savingAndTesting} onClose={() => setDraft(null)} onSubmit={() => void save()} secondaryAction={editing ? <button type="button" className={controls.secondary} disabled={cursorBusy || savingAndTesting} onClick={() => void saveAndTest()}>{savingAndTesting ? t("测试中…") : t("保存并测试")}</button> : undefined}>
+      {draft && <>
+        <CursorModelEditor draft={draft} providers={providers} editing={editing !== null} modelOptions={modelOptions} discovering={discovering} onChange={setDraft} onDiscover={() => void discover()} />
+        {editing && modelTestResults.get(editing.model_hash) && <div className={styles.editorTestResult}><CursorModelTestResult state={modelTestResults.get(editing.model_hash)!} /></div>}
+      </>}
     </Modal>
     <Modal open={caCommand !== null} title={t("安装本地 CA")} closeLabel={t("关闭")} submitLabel={t("打开终端")} onClose={() => setCaCommand(null)} onSubmit={openCaTerminal}>
       <div className={styles.editor}>
@@ -143,6 +228,22 @@ function parseObject(text: string, label: string): Record<string, unknown> {
   return parsed as Record<string, unknown>;
 }
 
+function cursorModelInputs(draft: CursorModelDraft, editing: boolean) {
+  const modelIds = editing
+    ? [draft.model.model_id.trim()]
+    : [...new Set(draft.modelIds.map((modelId) => modelId.trim()).filter(Boolean))];
+  if (!modelIds.length) throw new Error(t("请至少选择或输入一个模型"));
+  if (editing && !draft.model.display_name.trim()) throw new Error(t("Model ID 和显示名称不能为空"));
+  if (draft.customRequestUrl && !draft.model.request_url.trim()) throw new Error(t("请求完整地址不能为空"));
+  if (draft.model.context_window_tokens !== null && (!Number.isSafeInteger(draft.model.context_window_tokens) || draft.model.context_window_tokens <= 0)) throw new Error(t("自定义上下文必须是大于 0 的整数"));
+  return modelIds.map((modelId, index) => ({
+    ...draft.model,
+    model_id: modelId,
+    display_name: modelIds.length === 1 ? draft.model.display_name.trim() || modelId : modelId,
+    sort_order: draft.model.sort_order + index,
+  }));
+}
+
 function providerName(baseUrl: string): string {
   try {
     return new URL(baseUrl.trim()).hostname;

apps/desktop/src/pages/ProvidersPage.tsx

@@ -40,7 +40,7 @@ export function ProvidersPage() {
   };
   const openEdit = (provider: Provider) => {
     setEditing(provider);
-    setDraft({ name: provider.name, provider_type: provider.provider_type, base_url: provider.base_url, api_key: "", custom_headers: provider.custom_headers, extra_params: provider.extra_params });
+    setDraft({ name: provider.name, provider_type: provider.provider_type, base_url: provider.base_url, api_key: provider.api_key ?? "", custom_headers: provider.custom_headers, extra_params: provider.extra_params });
     setHeadersText(JSON.stringify(provider.custom_headers, null, 2));
     setExtraText(JSON.stringify(provider.extra_params, null, 2));
   };

apps/desktop/src/store/appStore.ts

@@ -198,12 +198,12 @@ export const appStore = {
   async updateCursorModel(hash: string, model: ModelInput) {
     update({ cursorBusy: true, error: null });
     try {
-      await api.updateModel(hash, model);
+      const updated = await api.updateModel(hash, model);
       await appStore.refresh();
-      return true;
+      return updated;
     } catch (cause) {
       update({ error: cause instanceof Error ? cause.message : String(cause) });
-      return false;
+      return null;
     } finally { update({ cursorBusy: false }); }
   },
 

apps/desktop/src/store/updateStore.ts

@@ -0,0 +1,74 @@
+import { useSyncExternalStore } from "react";
+import type { Update } from "@tauri-apps/plugin-updater";
+import {
+  checkForUpdate,
+  hasNativeAppLifecycle,
+  installUpdate,
+} from "../native/appLifecycle";
+
+export type UpdateSnapshot = {
+  availableVersion: string | null;
+  checking: boolean;
+  installing: boolean;
+};
+
+let snapshot: UpdateSnapshot = {
+  availableVersion: null,
+  checking: false,
+  installing: false,
+};
+let availableUpdate: Update | null = null;
+let pendingCheck: Promise<string | null> | null = null;
+const listeners = new Set<() => void>();
+
+function update(patch: Partial<UpdateSnapshot>) {
+  snapshot = { ...snapshot, ...patch };
+  listeners.forEach((listener) => listener());
+}
+
+async function replaceAvailableUpdate(next: Update | null) {
+  const previous = availableUpdate;
+  availableUpdate = next;
+  update({ availableVersion: next?.version ?? null });
+  if (previous && previous !== next) await previous.close();
+}
+
+export const updateStore = {
+  subscribe(listener: () => void) {
+    listeners.add(listener);
+    return () => listeners.delete(listener);
+  },
+  getSnapshot: () => snapshot,
+
+  async check(): Promise<string | null> {
+    if (!hasNativeAppLifecycle()) return null;
+    if (pendingCheck) return pendingCheck;
+    update({ checking: true });
+    pendingCheck = (async () => {
+      const next = await checkForUpdate();
+      await replaceAvailableUpdate(next);
+      return next?.version ?? null;
+    })();
+    try {
+      return await pendingCheck;
+    } finally {
+      pendingCheck = null;
+      update({ checking: false });
+    }
+  },
+
+  async install(): Promise<void> {
+    const current = availableUpdate;
+    if (!current) return;
+    update({ installing: true });
+    try {
+      await installUpdate(current);
+    } finally {
+      update({ installing: false });
+    }
+  },
+};
+
+export function useUpdateStore(): UpdateSnapshot {
+  return useSyncExternalStore(updateStore.subscribe, updateStore.getSnapshot);
+}

cursor-proto/proto/git_forge_v1.proto

@@ -0,0 +1,929 @@
+syntax = "proto3";
+
+package git_forge.v1;
+
+option go_package = "github.com/leookun/cursor-byok/cursor-proto/gen/git_forge/v1;git_forgev1";
+
+// Copied from: local:git_forge.v1.BatchGetBlobOidsRequest (var: Gn)
+message BatchGetBlobOidsRequest {
+  string repo_uuid = 1;
+  string revision = 2;
+  repeated string paths = 3;
+}
+
+// Copied from: local:git_forge.v1.BatchGetBlobOidsResponse (var: Vn)
+message BatchGetBlobOidsResponse {
+  repeated BlobOidResult results = 1;
+  string resolved_commit_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.BatchGetCommitMetadataRequest (var: Ut)
+message BatchGetCommitMetadataRequest {
+  string repo_uuid = 1;
+  repeated string revisions = 2;
+}
+
+// Copied from: local:git_forge.v1.BatchGetCommitMetadataResponse (var: Gt)
+message BatchGetCommitMetadataResponse {
+  repeated BatchGetCommitMetadataResult results = 1;
+}
+
+// Copied from: local:git_forge.v1.BatchGetCommitMetadataResult (var: Qt)
+message BatchGetCommitMetadataResult {
+  string revision = 1;
+  CommitMetadata metadata = 2;
+}
+
+// Copied from: local:git_forge.v1.BatchGetRepoContentRequest (var: Un)
+message BatchGetRepoContentRequest {
+  string repo_uuid = 1;
+  string revision = 2;
+  repeated string paths = 3;
+  optional uint64 max_file_size = 4;
+}
+
+// Copied from: local:git_forge.v1.BatchGetRepoContentResponse (var: Qn)
+message BatchGetRepoContentResponse {
+  repeated BatchRepoContentResult results = 1;
+  string resolved_commit_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.BatchRepoContentResult (var: xn)
+message BatchRepoContentResult {
+  string path = 1;
+  bool found = 2;
+  oneof content {
+    FileContent file_content = 3;
+    DirectoryContent directory_content = 4;
+  }
+}
+
+// Copied from: local:git_forge.v1.BlameChunk (var: GSv)
+message BlameChunk {
+  repeated ShortCommit commits = 1;
+  repeated BlameLineRange line_ranges = 2;
+}
+
+// Copied from: local:git_forge.v1.BlameLineRange (var: VSv)
+message BlameLineRange {
+  bytes commit_sha = 1;
+  uint32 start_in_blamed_file = 2;
+  uint32 len = 3;
+}
+
+// Copied from: local:git_forge.v1.BlobOidResult (var: Hn)
+message BlobOidResult {
+  string path = 1;
+  bool found = 2;
+  string blob_oid = 3;
+}
+
+// Copied from: local:git_forge.v1.BranchDeletion (var: Or)
+message BranchDeletion {
+  string target_ref = 1;
+  string expected_head_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.CanMergeRequest (var: on)
+message CanMergeRequest {
+  string repo_uuid = 1;
+  string ours = 2;
+  string theirs = 3;
+  uint64 change_number = 4;
+  MergeMode mode = 7;
+  bool use_git_conflict_semantics = 8;
+}
+
+// Copied from: local:git_forge.v1.CanMergeResponse (var: LSv)
+message CanMergeResponse {
+  bool can_merge_without_conflicts = 1;
+  optional bytes merged_tree_sha = 2;
+  repeated string conflicted_paths = 3;
+}
+
+// Copied from: local:git_forge.v1.ChangeKind (var: TQd)
+enum ChangeKind {
+  CHANGE_KIND_UNSPECIFIED = 0;
+  CHANGE_KIND_ADDED = 1;
+  CHANGE_KIND_DELETED = 2;
+  CHANGE_KIND_MODIFIED = 3;
+  CHANGE_KIND_RENAMED = 4;
+  CHANGE_KIND_COPIED = 5;
+}
+
+// Copied from: local:git_forge.v1.Commit (var: azi)
+message Commit {
+  string sha = 1;
+  string message = 2;
+  Signature author = 3;
+  Signature committer = 4;
+  repeated string parent_shas = 5;
+  string tree_sha = 7;
+  optional string change_id = 8;
+}
+
+// Copied from: local:git_forge.v1.CommitDiffChunk (var: rwv)
+message CommitDiffChunk {
+  optional CommitDiffHeader header = 1;
+  repeated DiffEntry entries = 3;
+}
+
+// Copied from: local:git_forge.v1.CommitDiffHeader (var: iwv)
+message CommitDiffHeader {
+  Commit commit = 1;
+  optional string base_commit_sha = 2;
+  CommitDiffStats stats = 3;
+  bool has_more = 5;
+  optional string next_page_cursor = 6;
+  repeated DiffEntry entries = 7;
+}
+
+// Copied from: local:git_forge.v1.CommitDiffStats (var: owv)
+message CommitDiffStats {
+  uint32 files_changed = 1;
+  int32 additions = 2;
+  int32 deletions = 3;
+}
+
+// Copied from: local:git_forge.v1.CommitFileDelete (var: OSv)
+message CommitFileDelete {
+}
+
+// Copied from: local:git_forge.v1.CommitFileMode (var: kSv)
+enum CommitFileMode {
+  COMMIT_FILE_MODE_UNSPECIFIED = 0;
+  COMMIT_FILE_MODE_REGULAR = 1;
+  COMMIT_FILE_MODE_EXECUTABLE = 2;
+  COMMIT_FILE_MODE_SYMLINK = 3;
+}
+
+// Copied from: local:git_forge.v1.CommitFileOperation (var: FSv)
+message CommitFileOperation {
+  string path = 1;
+  oneof operation {
+    CommitFileUpsert upsert = 2;
+    CommitFileDelete delete = 3;
+  }
+}
+
+// Copied from: local:git_forge.v1.CommitFileUpsert (var: NSv)
+message CommitFileUpsert {
+  bytes content = 1;
+  CommitFileMode mode = 2;
+}
+
+// Copied from: local:git_forge.v1.CommitMetadata (var: xt)
+message CommitMetadata {
+  string sha = 1;
+  Signature author = 2;
+  Signature committer = 3;
+  string message = 4;
+}
+
+// Copied from: local:git_forge.v1.CompareCommitsRequest (var: rn)
+message CompareCommitsRequest {
+  string repo_uuid = 1;
+  string base_revision = 2;
+  string head_revision = 3;
+}
+
+// Copied from: local:git_forge.v1.CompareCommitsResponse (var: PSv)
+message CompareCommitsResponse {
+  CompareCommitsStatus status = 1;
+  int32 ahead_by = 2;
+  int32 behind_by = 3;
+  string base_commit_sha = 4;
+  string head_commit_sha = 5;
+  string merge_base_commit_sha = 6;
+}
+
+// Copied from: local:git_forge.v1.CompareCommitsStatus (var: wSv)
+enum CompareCommitsStatus {
+  COMPARE_COMMITS_STATUS_UNSPECIFIED = 0;
+  COMPARE_COMMITS_STATUS_IDENTICAL = 1;
+  COMPARE_COMMITS_STATUS_AHEAD = 2;
+  COMPARE_COMMITS_STATUS_BEHIND = 3;
+  COMPARE_COMMITS_STATUS_DIVERGED = 4;
+}
+
+// Copied from: local:git_forge.v1.ComputeMergeCommitRequest (var: mn)
+message ComputeMergeCommitRequest {
+  string repo_uuid = 1;
+  string ours_sha = 2;
+  string theirs_sha = 3;
+  string message = 4;
+  Signature author = 5;
+  Signature committer = 6;
+  MergeMode mode = 7;
+}
+
+// Copied from: local:git_forge.v1.ComputeMergeCommitResponse (var: cn)
+message ComputeMergeCommitResponse {
+  string merge_commit_sha = 1;
+  bytes packfile = 2;
+}
+
+// Copied from: local:git_forge.v1.CreateBranchRequest (var: Qr)
+message CreateBranchRequest {
+  string repo_uuid = 1;
+  string target_ref = 2;
+  string new_head_sha = 3;
+}
+
+// Copied from: local:git_forge.v1.CreateBranchResponse (var: Gr)
+message CreateBranchResponse {
+  string new_head_sha = 1;
+  string wal_entry_key = 2;
+  bool unchanged = 3;
+}
+
+// Copied from: local:git_forge.v1.CreateCommitFromFilesRequest (var: En)
+message CreateCommitFromFilesRequest {
+  string repo_uuid = 1;
+  string target_ref = 2;
+  optional string expected_head_sha = 3;
+  string message = 4;
+  Signature author = 5;
+  optional Signature committer = 6;
+  repeated CommitFileOperation files = 7;
+}
+
+// Copied from: local:git_forge.v1.CreateCommitFromFilesResponse (var: Sn)
+message CreateCommitFromFilesResponse {
+  string commit_sha = 1;
+  string tree_sha = 2;
+  string old_head_sha = 3;
+  string wal_entry_key = 4;
+}
+
+// Copied from: local:git_forge.v1.CreateMergeCommitRequest (var: ln)
+message CreateMergeCommitRequest {
+  string repo_uuid = 1;
+  optional string ours_sha = 2;
+  optional string theirs_sha = 3;
+  string ours_ref = 4;
+  string theirs_ref = 5;
+  string message = 6;
+  Signature author = 7;
+  Signature committer = 8;
+  uint64 change_number = 9;
+  MergeMode mode = 10;
+}
+
+// Copied from: local:git_forge.v1.CreateMergeCommitResponse (var: un)
+message CreateMergeCommitResponse {
+  string merge_commit_sha = 1;
+  string wal_entry_key = 2;
+}
+
+// Copied from: local:git_forge.v1.CreateRepoRequest (var: kr)
+message CreateRepoRequest {
+  string repo_uuid = 1;
+}
+
+// Copied from: local:git_forge.v1.CreateRepoResponse (var: wr)
+message CreateRepoResponse {
+}
+
+// Copied from: local:git_forge.v1.CreateRevertBranchConflict (var: Jn)
+message CreateRevertBranchConflict {
+  repeated string conflicted_paths = 1;
+}
+
+// Copied from: local:git_forge.v1.CreateRevertBranchRequest (var: In)
+message CreateRevertBranchRequest {
+  string repo_uuid = 1;
+  string commit_sha = 2;
+  string base_ref = 3;
+  string branch_ref = 4;
+  string message = 5;
+  Signature author = 6;
+  optional Signature committer = 7;
+}
+
+// Copied from: local:git_forge.v1.CreateRevertBranchResponse (var: vn)
+message CreateRevertBranchResponse {
+  oneof result {
+    CreateRevertBranchSuccess success = 1;
+    CreateRevertBranchConflict conflict = 2;
+  }
+}
+
+// Copied from: local:git_forge.v1.CreateRevertBranchSuccess (var: Bn)
+message CreateRevertBranchSuccess {
+  string revert_commit_sha = 1;
+  string base_sha = 2;
+  string wal_entry_key = 3;
+}
+
+// Copied from: local:git_forge.v1.DeleteBranchesRequest (var: Ur)
+message DeleteBranchesRequest {
+  string repo_uuid = 1;
+  repeated BranchDeletion branches = 2;
+}
+
+// Copied from: local:git_forge.v1.DeleteBranchesResponse (var: xr)
+message DeleteBranchesResponse {
+  uint32 deleted_count = 1;
+  uint32 missing_count = 2;
+  string wal_entry_key = 3;
+  bool unchanged = 4;
+}
+
+// Copied from: local:git_forge.v1.DiffEntry (var: czi)
+message DiffEntry {
+  string path = 1;
+  optional string old_path = 2;
+  ChangeKind change_kind = 3;
+  optional FileMode old_mode = 4;
+  optional FileMode new_mode = 5;
+  optional string old_sha = 6;
+  optional string new_sha = 7;
+  optional FileStats stats = 8;
+  optional string patch = 9;
+  optional bool is_generated = 10;
+}
+
+// Copied from: local:git_forge.v1.DiffHeader (var: swv)
+message DiffHeader {
+  string merge_base_commit_sha = 1;
+  bool has_more = 4;
+  optional string next_page_cursor = 5;
+  repeated DiffEntry entries = 6;
+}
+
+// Copied from: local:git_forge.v1.DirectoryContent (var: BSv)
+message DirectoryContent {
+  repeated RepoContentEntry entries = 1;
+  string sha = 2;
+}
+
+// Copied from: local:git_forge.v1.FastForwardRefRequest (var: Fr)
+message FastForwardRefRequest {
+  string repo_uuid = 1;
+  string target_ref = 2;
+  string expected_head_sha = 3;
+  string new_head_sha = 4;
+}
+
+// Copied from: local:git_forge.v1.FastForwardRefResponse (var: Dr)
+message FastForwardRefResponse {
+  string old_head_sha = 1;
+  string new_head_sha = 2;
+  string wal_entry_key = 3;
+  bool unchanged = 4;
+}
+
+// Copied from: local:git_forge.v1.FileContent (var: IQd)
+message FileContent {
+  string size = 1;
+  string encoding = 2;
+  string content = 3;
+  string sha = 4;
+}
+
+// Copied from: local:git_forge.v1.FileHistoryCommitEntry (var: HSv)
+message FileHistoryCommitEntry {
+  ShortCommit commit = 1;
+  optional string diff_base_commit_sha = 2;
+  optional FileStats path_stats = 8;
+}
+
+// Copied from: local:git_forge.v1.FileHistoryWithDiffStatsChunk (var: er)
+message FileHistoryWithDiffStatsChunk {
+  repeated FileHistoryCommitEntry entries = 1;
+  bool exhausted = 2;
+  bool scan_truncated = 3;
+  optional string resume_commit_sha = 4;
+}
+
+// Copied from: local:git_forge.v1.FileMode (var: xQd)
+enum FileMode {
+  FILE_MODE_UNSPECIFIED = 0;
+  FILE_MODE_REGULAR = 1;
+  FILE_MODE_EXECUTABLE = 2;
+  FILE_MODE_SYMLINK = 3;
+  FILE_MODE_GITLINK = 4;
+}
+
+// Copied from: local:git_forge.v1.FileStats (var: DQd)
+message FileStats {
+  int32 additions = 1;
+  int32 deletions = 2;
+  bool is_binary = 3;
+}
+
+// Copied from: local:git_forge.v1.GetBlameRequest (var: nr)
+message GetBlameRequest {
+  string repo_uuid = 1;
+  string start_commit_sha = 2;
+  string path = 3;
+}
+
+// Copied from: local:git_forge.v1.GetBlobRequest (var: Ht)
+message GetBlobRequest {
+  string repo_uuid = 1;
+  string blob_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.GetBlobResponse (var: xSv)
+message GetBlobResponse {
+  FileContent blob = 1;
+}
+
+// Copied from: local:git_forge.v1.GetCommitDiffRequest (var: Ir)
+message GetCommitDiffRequest {
+  string repo_uuid = 1;
+  string commit_sha = 2;
+  optional string base_commit_sha = 3;
+  bool include_patches = 4;
+  repeated string paths = 5;
+  optional uint32 page_size = 6;
+  optional string page_cursor = 7;
+}
+
+// Copied from: local:git_forge.v1.GetCommitRequest (var: Dt)
+message GetCommitRequest {
+  string repo_uuid = 1;
+  string commit_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.GetCommitResponse (var: ESv)
+message GetCommitResponse {
+  Commit commit = 1;
+}
+
+// Copied from: local:git_forge.v1.GetDefaultBranchRequest (var: qn)
+message GetDefaultBranchRequest {
+  string repo_uuid = 1;
+}
+
+// Copied from: local:git_forge.v1.GetDefaultBranchResponse (var: Fn)
+message GetDefaultBranchResponse {
+  string default_branch = 1;
+}
+
+// Copied from: local:git_forge.v1.GetFileHistoryPageWithDiffStatsResponse (var: zSv)
+message GetFileHistoryPageWithDiffStatsResponse {
+  repeated FileHistoryCommitEntry entries = 1;
+  bool has_more = 2;
+  optional string next_cursor = 3;
+}
+
+// Copied from: local:git_forge.v1.GetFileHistoryRequest (var: zn)
+message GetFileHistoryRequest {
+  string repo_uuid = 1;
+  string start_commit_sha = 2;
+  optional string path = 3;
+  uint32 max_commits = 4;
+}
+
+// Copied from: local:git_forge.v1.GetFileHistoryResponse (var: WSv)
+message GetFileHistoryResponse {
+  repeated ShortCommit commits = 1;
+}
+
+// Copied from: local:git_forge.v1.GetFileHistoryWithDiffStatsRequest (var: Xn)
+message GetFileHistoryWithDiffStatsRequest {
+  string repo_uuid = 1;
+  string start_commit_sha = 2;
+  optional string path = 3;
+  uint32 max_commits = 4;
+  optional string next_cursor = 5;
+  bool include_diff_stats = 6;
+  optional uint32 max_commits_scanned = 7;
+}
+
+// Copied from: local:git_forge.v1.GetFileHistoryWithDiffStatsResponse (var: Zn)
+message GetFileHistoryWithDiffStatsResponse {
+  repeated FileHistoryCommitEntry entries = 1;
+  bool has_more = 2;
+  optional string next_cursor = 3;
+}
+
+// Copied from: local:git_forge.v1.GetFuzzyPathsRequest (var: lr)
+message GetFuzzyPathsRequest {
+  string repo_uuid = 1;
+  string commit_sha = 2;
+  string query = 3;
+  uint32 limit = 4;
+}
+
+// Copied from: local:git_forge.v1.GetFuzzyPathsResponse (var: QSv)
+message GetFuzzyPathsResponse {
+  repeated string paths = 1;
+  bool has_more = 2;
+}
+
+// Copied from: local:git_forge.v1.GetLocalDevInfoRequest (var: qt)
+message GetLocalDevInfoRequest {
+}
+
+// Copied from: local:git_forge.v1.GetLocalDevInfoResponse (var: Ft)
+message GetLocalDevInfoResponse {
+  string repo_uuid = 1;
+  string git_forge_root_dir = 2;
+}
+
+// Copied from: local:git_forge.v1.GetPullRequestDiffRequest (var: Er)
+message GetPullRequestDiffRequest {
+  string repo_uuid = 1;
+  string head_commit_sha = 2;
+  string base_commit_sha = 3;
+  optional bool include_patches = 4;
+  optional uint32 page_size = 5;
+  optional string page_cursor = 6;
+  optional bool include_file_stats = 7;
+  optional bool include_generated = 8;
+}
+
+// Copied from: local:git_forge.v1.GetRepoContentDetailsRequest (var: Kn)
+message GetRepoContentDetailsRequest {
+  string repo_uuid = 1;
+  PathIdentifier path_identifier = 2;
+}
+
+// Copied from: local:git_forge.v1.GetRepoContentDetailsResponse (var: jSv)
+message GetRepoContentDetailsResponse {
+  optional RepoContentDetails details = 1;
+  PathIdentifier path_identifier = 2;
+  string resolved_commit_sha = 3;
+}
+
+// Copied from: local:git_forge.v1.GetRepoContentRequest (var: Dn)
+message GetRepoContentRequest {
+  string repo_uuid = 1;
+  optional uint64 max_file_size = 4;
+  oneof id {
+    PathIdentifier path_identifier = 2;
+    string ref_and_path = 3;
+  }
+}
+
+// Copied from: local:git_forge.v1.GetRepoContentResponse (var: RQd)
+message GetRepoContentResponse {
+  PathIdentifier path_identifier = 3;
+  string resolved_commit_sha = 4;
+  oneof content {
+    FileContent file_content = 1;
+    DirectoryContent directory_content = 2;
+  }
+}
+
+// Copied from: local:git_forge.v1.GetTagRequest (var: Yt)
+message GetTagRequest {
+  string repo_uuid = 1;
+  string tag_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.GetTagResponse (var: TSv)
+message GetTagResponse {
+  Tag tag = 1;
+}
+
+// Copied from: local:git_forge.v1.GetTreeBlameRequest (var: or)
+message GetTreeBlameRequest {
+  string repo_uuid = 1;
+  string start_commit_sha = 2;
+  string path = 3;
+}
+
+// Copied from: local:git_forge.v1.GetTreeBlameResponse (var: KSv)
+message GetTreeBlameResponse {
+  repeated TreeEntryBlame entries = 1;
+}
+
+// Copied from: local:git_forge.v1.GetTreeRequest (var: Wt)
+message GetTreeRequest {
+  string repo_uuid = 1;
+  string tree_sha = 2;
+  bool recursive = 3;
+}
+
+// Copied from: local:git_forge.v1.GetTreeResponse (var: ASv)
+message GetTreeResponse {
+  Tree tree = 1;
+}
+
+// Copied from: local:git_forge.v1.GrepLineKind (var: CSv)
+enum GrepLineKind {
+  GREP_LINE_KIND_UNSPECIFIED = 0;
+  GREP_LINE_KIND_MATCH = 1;
+  GREP_LINE_KIND_CONTEXT = 2;
+}
+
+// Copied from: local:git_forge.v1.GrepMatch (var: JSv)
+message GrepMatch {
+  string path = 1;
+  string lines = 2;
+  uint32 line_number = 3;
+  uint64 absolute_offset = 4;
+  repeated GrepSubmatch submatches = 5;
+  GrepLineKind kind = 6;
+}
+
+// Copied from: local:git_forge.v1.GrepRepoChunk (var: ewv)
+message GrepRepoChunk {
+  repeated GrepMatch matches = 1;
+  bool limit_hit = 2;
+}
+
+// Copied from: local:git_forge.v1.GrepRepoRequest (var: pr)
+message GrepRepoRequest {
+  string repo_uuid = 1;
+  string revision = 2;
+  string query = 3;
+  GrepSearchOptions options = 4;
+  uint32 max_results = 5;
+}
+
+// Copied from: local:git_forge.v1.GrepSearchOptions (var: ZSv)
+message GrepSearchOptions {
+  bool literal = 1;
+  bool case_insensitive = 2;
+  bool whole_word = 3;
+  uint32 context_before = 4;
+  uint32 context_after = 5;
+  optional uint64 max_lines = 6;
+  optional string filter_path = 7;
+  repeated string includes = 8;
+  repeated string excludes = 9;
+}
+
+// Copied from: local:git_forge.v1.GrepSubmatch (var: XSv)
+message GrepSubmatch {
+  uint32 start = 1;
+  uint32 end = 2;
+}
+
+// Copied from: local:git_forge.v1.ListCommitsInRangeRequest (var: tn)
+message ListCommitsInRangeRequest {
+  string repo_uuid = 1;
+  string base_revision = 2;
+  string head_revision = 3;
+  optional int32 max_commits = 4;
+  bool oldest_first = 5;
+  ListCommitsSort sort = 6;
+}
+
+// Copied from: local:git_forge.v1.ListCommitsInRangeResponse (var: MSv)
+message ListCommitsInRangeResponse {
+  repeated Commit commits = 1;
+  string base_commit_sha = 2;
+  string head_commit_sha = 3;
+  string merge_base_commit_sha = 4;
+  bool truncated = 5;
+}
+
+// Copied from: local:git_forge.v1.ListCommitsRequest (var: Zt)
+message ListCommitsRequest {
+  string repo_uuid = 1;
+  string revision = 2;
+  int32 page = 3;
+  int32 per_page = 4;
+  ListCommitsSort sort = 5;
+}
+
+// Copied from: local:git_forge.v1.ListCommitsResponse (var: en)
+message ListCommitsResponse {
+  repeated Commit commits = 1;
+  optional int32 next_page = 2;
+}
+
+// Copied from: local:git_forge.v1.ListCommitsSort (var: k)
+enum ListCommitsSort {
+  LIST_COMMITS_SORT_UNSPECIFIED = 0;
+  LIST_COMMITS_SORT_COMMIT_TIME = 1;
+  LIST_COMMITS_SORT_TOPOLOGICAL = 2;
+}
+
+// Copied from: local:git_forge.v1.ListRefsFilter (var: AQd)
+enum ListRefsFilter {
+  LIST_REFS_FILTER_UNSPECIFIED = 0;
+  LIST_REFS_FILTER_ALL = 1;
+  LIST_REFS_FILTER_BRANCHES = 2;
+  LIST_REFS_FILTER_TAGS = 3;
+}
+
+// Copied from: local:git_forge.v1.ListRefsRequest (var: Pr)
+message ListRefsRequest {
+  string repo_uuid = 1;
+  ListRefsFilter filter = 2;
+  bool names_only = 3;
+  string prefix = 4;
+  optional uint32 limit = 5;
+  optional string page_token = 7;
+}
+
+// Copied from: local:git_forge.v1.ListRefsResponse (var: awv)
+message ListRefsResponse {
+  repeated string refs = 1;
+  repeated RefInfo ref_infos = 2;
+  optional string next_page_token = 5;
+}
+
+// Copied from: local:git_forge.v1.ListTreePathsRequest (var: mr)
+message ListTreePathsRequest {
+  string repo_uuid = 1;
+  string revision = 2;
+  repeated string includes = 3;
+  repeated string excludes = 4;
+  uint32 limit = 5;
+}
+
+// Copied from: local:git_forge.v1.ListTreePathsResponse (var: YSv)
+message ListTreePathsResponse {
+  repeated string paths = 1;
+  bool has_more = 2;
+}
+
+// Copied from: local:git_forge.v1.MergeMode (var: EQd)
+enum MergeMode {
+  MERGE_MODE_UNSPECIFIED = 0;
+  MERGE_MODE_MERGE_COMMIT = 1;
+  MERGE_MODE_SQUASH = 2;
+}
+
+// Copied from: local:git_forge.v1.NotifyRepoPushedRequest (var: Ar)
+message NotifyRepoPushedRequest {
+  string repo_uuid = 1;
+}
+
+// Copied from: local:git_forge.v1.NotifyRepoPushedResponse (var: _r)
+message NotifyRepoPushedResponse {
+}
+
+// Copied from: local:git_forge.v1.PathIdentifier (var: B$t)
+message PathIdentifier {
+  string revision = 1;
+  string path = 2;
+}
+
+// Copied from: local:git_forge.v1.PrepareChangeMergeRequest (var: dn)
+message PrepareChangeMergeRequest {
+  string repo_uuid = 1;
+  string base_ref = 2;
+  string head_ref = 3;
+  uint64 change_number = 4;
+  optional string expected_base_sha = 5;
+  optional string expected_head_sha = 6;
+  MergeMode mode = 7;
+}
+
+// Copied from: local:git_forge.v1.PrepareChangeMergeResponse (var: pn)
+message PrepareChangeMergeResponse {
+  bool mergeable = 1;
+  optional string merged_tree_sha = 2;
+  optional string change_merge_ref = 3;
+  optional string dummy_commit_sha = 4;
+}
+
+// Copied from: local:git_forge.v1.PullRequestDiffChunk (var: nwv)
+message PullRequestDiffChunk {
+  optional DiffHeader header = 1;
+  repeated DiffEntry entries = 3;
+}
+
+// Copied from: local:git_forge.v1.RebaseStackBranch (var: gn)
+message RebaseStackBranch {
+  string head_ref = 1;
+  string expected_old_oid = 2;
+  optional string replay_base_oid = 3;
+}
+
+// Copied from: local:git_forge.v1.RebaseStackBranchUpdate (var: _n)
+message RebaseStackBranchUpdate {
+  string head_ref = 1;
+  string old_oid = 2;
+  string new_oid = 3;
+}
+
+// Copied from: local:git_forge.v1.RebaseStackConflict (var: Tn)
+message RebaseStackConflict {
+  string conflicted_head_ref = 1;
+  repeated string conflicted_paths = 2;
+}
+
+// Copied from: local:git_forge.v1.RebaseStackRequest (var: fn)
+message RebaseStackRequest {
+  string repo_uuid = 1;
+  string onto_ref = 2;
+  optional string expected_onto_oid = 3;
+  repeated RebaseStackBranch branches = 4;
+  bool dry_run = 5;
+}
+
+// Copied from: local:git_forge.v1.RebaseStackResponse (var: hn)
+message RebaseStackResponse {
+  oneof result {
+    RebaseStackSuccess success = 1;
+    RebaseStackConflict conflict = 2;
+  }
+}
+
+// Copied from: local:git_forge.v1.RebaseStackSuccess (var: An)
+message RebaseStackSuccess {
+  string wal_entry_key = 1;
+  repeated RebaseStackBranchUpdate updates = 2;
+}
+
+// Copied from: local:git_forge.v1.RefInfo (var: MQd)
+message RefInfo {
+  string name = 1;
+  string target_sha = 2;
+  string object_sha = 3;
+  string object_type = 4;
+}
+
+// Copied from: local:git_forge.v1.RepoContentDetails (var: $Sv)
+message RepoContentDetails {
+  string type = 1;
+  optional uint64 size = 2;
+  bool is_binary = 3;
+  bool too_large_to_introspect = 4;
+}
+
+// Copied from: local:git_forge.v1.RepoContentEntry (var: twv)
+message RepoContentEntry {
+  string type = 1;
+  string name = 2;
+  string path = 3;
+  string sha = 4;
+  optional uint64 size = 5;
+}
+
+// Copied from: local:git_forge.v1.ResolveRefPathRequest (var: Mn)
+message ResolveRefPathRequest {
+  string repo_uuid = 1;
+  string ref_path = 2;
+}
+
+// Copied from: local:git_forge.v1.ResolveRefPathResponse (var: USv)
+message ResolveRefPathResponse {
+  PathIdentifier path_identifier = 1;
+  string resolved_commit_sha = 2;
+}
+
+// Copied from: local:git_forge.v1.SearchRefsRequest (var: Lr)
+message SearchRefsRequest {
+  string repo_uuid = 1;
+  ListRefsFilter filter = 2;
+  string query = 3;
+  optional uint32 limit = 4;
+}
+
+// Copied from: local:git_forge.v1.SearchRefsResponse (var: cwv)
+message SearchRefsResponse {
+  repeated RefInfo ref_infos = 1;
+  bool exact_match = 2;
+}
+
+// Copied from: local:git_forge.v1.ShortCommit (var: fMn)
+message ShortCommit {
+  bytes sha = 1;
+  string summary = 2;
+  string author_name = 3;
+  string author_email = 4;
+  int64 timestamp = 5;
+}
+
+// Copied from: local:git_forge.v1.Signature (var: o_t)
+message Signature {
+  string name = 1;
+  string email = 2;
+  int64 timestamp = 3;
+  int32 timezone_offset = 4;
+}
+
+// Copied from: local:git_forge.v1.Tag (var: DSv)
+message Tag {
+  string sha = 1;
+  string name = 2;
+  string message = 3;
+  Signature tagger = 4;
+  string object_sha = 5;
+  string object_type = 6;
+}
+
+// Copied from: local:git_forge.v1.Tree (var: ISv)
+message Tree {
+  string sha = 1;
+  repeated TreeEntry tree = 2;
+  bool truncated = 3;
+}
+
+// Copied from: local:git_forge.v1.TreeEntry (var: RSv)
+message TreeEntry {
+  string path = 1;
+  string mode = 2;
+  string type = 3;
+  string sha = 4;
+  optional uint64 size = 5;
+}
+
+// Copied from: local:git_forge.v1.TreeEntryBlame (var: qSv)
+message TreeEntryBlame {
+  string name = 1;
+  ShortCommit last_commit = 2;
+}
+

cursor-proto/proto/internapi_v1.proto

@@ -0,0 +1,41 @@
+syntax = "proto3";
+
+package internapi.v1;
+
+option go_package = "github.com/leookun/cursor-byok/cursor-proto/gen/internapi/v1;internapiv1";
+
+// Copied from: local:internapi.v1.BlobData (var: RLs)
+message BlobData {
+  BlobType blob_type = 1;
+  bytes blob_id = 2;
+  int32 index = 3;
+  oneof type_specific_data {
+    ImageBlobData image_data = 4;
+  }
+}
+
+// Copied from: local:internapi.v1.BlobDataPerMessage (var: ocu)
+message BlobDataPerMessage {
+  repeated BlobData blob_data = 1;
+}
+
+// Copied from: local:internapi.v1.BlobType (var: rcu)
+enum BlobType {
+  BLOB_TYPE_UNSPECIFIED = 0;
+  BLOB_TYPE_IMAGE = 1;
+  BLOB_TYPE_INVOCATION_CONTEXT = 2;
+  BLOB_TYPE_EXTRA_CONTEXT = 3;
+  BLOB_TYPE_GIT_PR_DIFF_SELECTION = 4;
+  BLOB_TYPE_SELECTED_PULL_REQUEST = 5;
+  BLOB_TYPE_TEXT = 6;
+  BLOB_TYPE_RICH_TEXT = 7;
+  BLOB_TYPE_EXTERNAL_LINK_PDF = 8;
+  BLOB_TYPE_DOCUMENT = 9;
+  BLOB_TYPE_VIDEO = 10;
+}
+
+// Copied from: local:internapi.v1.ImageBlobData (var: scu)
+message ImageBlobData {
+  string mime_type = 1;
+}
+

server/prompt/cursor/agent/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Agent mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 <system_reminder>

server/prompt/cursor/ask/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Ask mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/compaction/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<timestamp>{{TIMESTAMP}}</timestamp>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<timestamp>{{TIMESTAMP}}</timestamp>
 <user_query>
 {{USER_QUERY}}
 </user_query>

server/prompt/cursor/debug/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Debug mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/multitask/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Multitask mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/plan/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are now in Plan mode. You have EXITED your previous mode. Continue with the task in the new mode.
 </system_reminder>
 

server/prompt/cursor/subagent/runtime.md

@@ -1,4 +1,4 @@
-{{REQUEST_CONTEXT}}{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
+{{OPEN_FILES}}{{SELECTED_CONTEXT}}{{ACTION_CONTEXT}}<system_reminder>
 You are currently working inside a Task subagent. Your parent agent has delegated a clearly bounded assignment to you. Complete that assignment directly with the tools available in this session. The Task tool is unavailable inside subagents, so delegation cannot be nested.
 </system_reminder>
 <timestamp>{{TIMESTAMP}}</timestamp>

server/prompt/cursor/tools.json

@@ -580,7 +580,7 @@
       "type": "function",
       "function": {
         "name": "Shell",
-        "description": "Executes a given command in a shell session, waiting for output for `block_until_ms` millis.\nYou can monitor commands by configuring `notify_on_output`. You will be notified at the end of your turn whenever stdout/stderr output matches the regex `pattern`. Output redirected only to a file will not trigger it. Configure a 5-or-fewer-word `reason` explaining what you are watching for, and optionally configure `debounce_ms`.",
+        "description": "Executes a given command in a shell session with optional foreground timeout.\n\nIMPORTANT: This tool is for terminal operations like git, npm, docker, etc. DO NOT use it for file operations (reading, writing, editing, searching, finding files, sleeping) - use the specialized tools for this instead.\n\nYou can monitor commands by configuring `notify_on_output`. You will be notified at the end of your turn whenever stdout/stderr output matches the regex `pattern`. Output redirected only to a file will not trigger it. Configure a 5-or-fewer-word `reason` explaining what you are watching for, and optionally configure `debounce_ms`.\n\n<sandboxing>\nBy default, your commands will run in a sandbox. The sandbox allows most writes to the workspace and reads to the rest of the filesystem. Some other syscalls are also disallowed like access to USB devices.\n\nThe sandbox includes network access for common package managers and version control providers (e.g. npm, pypi, crates.io, Maven Central, GitHub, etc.). Standard operations like package installs and fetching dependencies will work without requesting additional permissions.\n\nFor broader network access beyond the allowed domains, you may still need to request 'full_network' permissions.\n\nThe required_permissions argument is used to request additional permissions. If you know you will need a permission, request it. Requesting permissions will slow down the command execution as it will ask the user for approval. Do not hesitate to request permissions if you are certain you need them. For commands you know will need unrestricted network access, request the full_network permission rather than waiting for the command to fail and asking for it later.\n\nThe following permissions are supported:\n\n- full_network: Grants unrestricted network access. This is useful for any commands that need to contact the outside internet, outside of the allowed domains.\n- all: Disables the sandbox entirely. If all is requested the command will run outside of the sandbox.\n\nIf you think a command failed due to sandbox restrictions, run the command again with the required_permissions argument to request what you need.\n</sandboxing>",
         "parameters": {
           "type": "object",
           "properties": {
@@ -629,6 +629,14 @@
             "working_directory": {
               "description": "The absolute path to the working directory to execute the command in (defaults to current directory)",
               "type": "string"
+            },
+            "required_permissions": {
+              "description": "Optional list of permissions to request if the command needs them. Use \"full_network\" for unrestricted network access beyond the sandbox allowlist, or \"all\" to disable the sandbox entirely.",
+              "type": "array",
+              "items": {
+                "type": "string",
+                "enum": ["full_network", "all"]
+              }
             }
           },
           "required": [

server/src/app.rs

@@ -41,8 +41,9 @@ impl App {
             config.provider_request_timeout,
         ));
         let run_registry = RunRegistry::default();
-        let registry = CursorSessionRegistry::new(store.clone(), provider, compiler, run_registry);
-        let control = control::ControlService::new(store.clone())?;
+        let registry =
+            CursorSessionRegistry::new(store.clone(), provider.clone(), compiler, run_registry);
+        let control = control::ControlService::new(store.clone(), provider)?;
         let harness = control.cursor_harness().clone();
         let mut router = handlers::router(registry.clone())?;
         router = match &config.console {

server/src/bin/cursor-server.rs

@@ -0,0 +1,15 @@
+use cursor_server::{App, Config, Result};
+use tracing_subscriber::prelude::*;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::registry()
+        .with(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| "cursor_server=info".into()),
+        )
+        .with(tracing_subscriber::fmt::layer())
+        .init();
+
+    App::new(Config::from_env()?).await?.serve().await
+}

server/src/control/mod.rs

@@ -22,7 +22,8 @@ use tower_http::{
 use url::{Host, Url};
 
 pub use service::{
-    CallDetail, CallSummary, ControlService, DiscoveredModels, ObservabilitySettings,
+    CallDetail, CallSummary, ControlService, DiscoveredModels, ModelConnectivityResult,
+    ObservabilitySettings,
 };
 
 pub fn web_router(service: ControlService, assets: impl AsRef<std::path::Path>) -> Router {
@@ -137,6 +138,10 @@ pub fn api_router(service: ControlService) -> Router {
             "/__byok-api__/api/models/{model_hash}",
             put(models::update).delete(models::remove),
         )
+        .route(
+            "/__byok-api__/api/models/{model_hash}/test",
+            post(models::test),
+        )
         .route("/__byok-api__/api/llm-calls", get(calls::list))
         .route("/__byok-api__/api/llm-calls/{call_id}", get(calls::detail))
         .route(
@@ -155,6 +160,10 @@ pub fn api_router(service: ControlService) -> Router {
             "/__byok-api__/api/settings/proxy",
             get(settings::get_proxy).put(settings::update_proxy),
         )
+        .route(
+            "/__byok-api__/api/settings/tab",
+            get(settings::get_tab).put(settings::update_tab),
+        )
         .route(
             "/__byok-api__/api/harness/cursor/status",
             get(harness::status),
@@ -242,7 +251,11 @@ mod tests {
         ))
         .await
         .unwrap();
-        let router = api_router(ControlService::new(store).unwrap());
+        let provider = std::sync::Arc::new(crate::provider::ProviderRouter::new(
+            store.clone(),
+            std::time::Duration::from_secs(300),
+        ));
+        let router = api_router(ControlService::new(store, provider).unwrap());
 
         let response = router
             .clone()

server/src/control/models.rs

@@ -10,7 +10,7 @@ use crate::{
     Result,
 };
 
-use super::{ControlService, DiscoveredModels};
+use super::{ControlService, DiscoveredModels, ModelConnectivityResult};
 
 #[derive(Deserialize)]
 pub struct SaveModels {
@@ -48,6 +48,13 @@ pub async fn update(
     Ok(Json(service.update_model(&model_hash, &input).await?))
 }
 
+pub async fn test(
+    State(service): State<ControlService>,
+    Path(model_hash): Path<String>,
+) -> Result<Json<ModelConnectivityResult>> {
+    Ok(Json(service.test_model(&model_hash).await?))
+}
+
 pub async fn discover(
     State(service): State<ControlService>,
     Path(provider_id): Path<i64>,

server/src/control/service.rs

@@ -1,8 +1,10 @@
-use std::collections::BTreeSet;
+use std::{collections::BTreeSet, sync::Arc, time::Instant};
 
 use base64::{engine::general_purpose::STANDARD, Engine};
+use futures_util::StreamExt;
 use reqwest::header::{HeaderName, HeaderValue};
 use serde::{Deserialize, Serialize};
+use tokio_util::sync::CancellationToken;
 use url::Url;
 
 use super::ads::{
@@ -13,11 +15,15 @@ use super::ads::{
 use crate::{
     harness::CursorHarness,
     model::{
-        CursorRunTraceArtifact, CursorRunTraceSummary, LlmCallRequest, LlmCallResponseChunk,
-        LlmCallSummary, Overview, ProviderEndpoint, ProviderEndpointInput, ProviderEndpointSecret,
-        ProviderModel, ProviderModelInput, ProviderType,
+        ContentPart, CursorRunTraceArtifact, CursorRunTraceSummary, LlmCallRequest,
+        LlmCallResponseChunk, LlmCallSummary, ModelInvocation, ModelRequest, ModelSpec, Overview,
+        ProjectedContent, ProjectedMessage, PromptSpec, ProviderEndpoint, ProviderEndpointInput,
+        ProviderModel, ProviderModelInput, ProviderType, Role,
+    },
+    provider::{ModelEvent, Provider},
+    store::{
+        PortSettings, ProxySettings, ProxySettingsInput, StatisticsStorage, Store, TabSettings,
     },
-    store::{PortSettings, ProxySettings, ProxySettingsInput, StatisticsStorage, Store},
     Error, Result,
 };
 
@@ -25,6 +31,7 @@ use crate::{
 pub struct ControlService {
     store: Store,
     cursor_harness: CursorHarness,
+    provider: Arc<dyn Provider>,
 }
 
 #[derive(Clone, Debug, Serialize)]
@@ -32,6 +39,16 @@ pub struct DiscoveredModels {
     pub models: Vec<String>,
 }
 
+#[derive(Clone, Debug, Serialize)]
+pub struct ModelConnectivityResult {
+    pub duration_ms: u64,
+    pub first_text_ms: Option<u64>,
+    pub output_tokens: u64,
+    pub tokens_per_second: f64,
+    pub tokens_estimated: bool,
+    pub output: String,
+}
+
 #[derive(Clone, Debug, Serialize)]
 pub struct CallDetail {
     pub call: CallSummary,
@@ -72,10 +89,11 @@ pub struct ObservabilitySettings {
 }
 
 impl ControlService {
-    pub fn new(store: Store) -> Result<Self> {
+    pub fn new(store: Store, provider: Arc<dyn Provider>) -> Result<Self> {
         Ok(Self {
             cursor_harness: CursorHarness::new(store.clone())?,
             store,
+            provider,
         })
     }
 
@@ -200,6 +218,125 @@ impl ControlService {
         self.store.update_provider_model(model_hash, input).await
     }
 
+    pub async fn test_model(&self, model_hash: &str) -> Result<ModelConnectivityResult> {
+        const TEST_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(45);
+        const TEST_PROMPT: &str = "Output the numbers 1 through 120 separated by a single space. No commas, no newlines, no explanation.";
+
+        let configured = self
+            .store
+            .provider_model(model_hash)
+            .await?
+            .ok_or_else(|| Error::RunNotFound(format!("model {model_hash}")))?;
+        let mut model = ModelSpec::new(model_hash);
+        if configured.reasoning_enabled {
+            model.reasoning.enabled = true;
+            model.reasoning.effort = Some(
+                configured
+                    .reasoning_effort
+                    .filter(|effort| !effort.trim().is_empty())
+                    .unwrap_or_else(|| "medium".into()),
+            );
+        }
+        let test_id = format!("model-test-{}", uuid::Uuid::new_v4());
+        let call_id = test_id.clone();
+        let invocation = ModelInvocation {
+            call_id: test_id.clone(),
+            run_id: test_id.clone(),
+            conversation_id: test_id,
+            provider_call_index: 0,
+            request: ModelRequest {
+                prompt: PromptSpec {
+                    instructions: String::new(),
+                    tools: Vec::new(),
+                },
+                model,
+                history: vec![ProjectedMessage {
+                    message_id: "connectivity-test".into(),
+                    role: Role::User,
+                    content: ProjectedContent::Parts(vec![ContentPart::Text {
+                        text: TEST_PROMPT.into(),
+                    }]),
+                }],
+            },
+        };
+        let cancellation = CancellationToken::new();
+        let started = Instant::now();
+        let mut first_text_at = None;
+        let mut output_tokens = None;
+        let mut output = String::new();
+        let stream = self.provider.stream(invocation, cancellation.clone());
+        let completed = tokio::time::timeout(TEST_TIMEOUT, async {
+            futures_util::pin_mut!(stream);
+            let mut finished = false;
+            while let Some(event) = stream.next().await {
+                match event? {
+                    ModelEvent::TextDelta(delta) => {
+                        if first_text_at.is_none() && !delta.trim().is_empty() {
+                            first_text_at = Some(Instant::now());
+                        }
+                        output.push_str(&delta);
+                    }
+                    ModelEvent::Usage(usage) => {
+                        if let Some(tokens) = usage.output_tokens.filter(|tokens| *tokens > 0) {
+                            output_tokens = Some(
+                                output_tokens.map_or(tokens, |current: u64| current.max(tokens)),
+                            );
+                        }
+                    }
+                    ModelEvent::Done(_) => finished = true,
+                    _ => {}
+                }
+            }
+            if !finished {
+                return Err(Error::Protocol(
+                    "provider stream ended without Done during connectivity test".into(),
+                ));
+            }
+            Ok(())
+        })
+        .await;
+        match completed {
+            Ok(result) => result?,
+            Err(_) => {
+                cancellation.cancel();
+                self.store
+                    .finish_llm_call(
+                        &call_id,
+                        "error",
+                        None,
+                        started.elapsed().as_millis().min(i64::MAX as u128) as i64,
+                        Some("timeout"),
+                        Some("model connectivity test timed out after 45 seconds"),
+                    )
+                    .await?;
+                return Err(Error::Provider(
+                    "model connectivity test timed out after 45 seconds".into(),
+                ));
+            }
+        }
+        let elapsed = started.elapsed();
+        let output = output.trim().to_string();
+        let tokens_estimated = output_tokens.is_none();
+        let output_tokens = output_tokens.unwrap_or_else(|| estimate_output_tokens(&output));
+        Ok(ModelConnectivityResult {
+            duration_ms: elapsed.as_millis().min(u128::from(u64::MAX)) as u64,
+            first_text_ms: first_text_at.map(|first| {
+                first
+                    .duration_since(started)
+                    .as_millis()
+                    .min(u128::from(u64::MAX)) as u64
+            }),
+            output_tokens,
+            tokens_per_second: if elapsed.is_zero() {
+                0.0
+            } else {
+                output_tokens as f64 / elapsed.as_secs_f64()
+            },
+            tokens_estimated,
+            output,
+        })
+    }
+
     pub async fn create_provider_with_models(
         &self,
         provider: &ProviderEndpointInput,
@@ -212,34 +349,15 @@ impl ControlService {
 
     pub async fn discover_input(&self, input: &ProviderEndpointInput) -> Result<DiscoveredModels> {
         let client = crate::network::client(&self.store).await?;
-        let endpoint = ProviderEndpoint {
-            provider_id: 0,
-            name: input.name.clone(),
-            provider_type: input.provider_type,
-            base_url: crate::model::normalize_base_url(&input.base_url)?,
-            has_api_key: input
-                .api_key
-                .as_deref()
-                .is_some_and(|value| !value.is_empty()),
-            custom_headers: input.custom_headers.clone(),
-            extra_params: input.extra_params.clone(),
-            created_at_ms: 0,
-            updated_at_ms: 0,
-        };
-        let secret = ProviderEndpointSecret {
-            endpoint,
-            api_key: input.api_key.clone().unwrap_or_default(),
-            custom_headers: input.custom_headers.clone(),
-        };
-        let mut models = match input.provider_type {
-            ProviderType::OpenAiChat | ProviderType::OpenAiResponses => {
-                openai_models(&client, &secret).await?
-            }
-            ProviderType::Anthropic => anthropic_models(&client, &secret).await?,
-        };
-        models.sort();
-        models.dedup();
-        Ok(DiscoveredModels { models })
+        let base_url = crate::model::normalize_base_url(&input.base_url)?;
+        discover_provider_models(
+            &client,
+            input.provider_type,
+            &base_url,
+            input.api_key.as_deref().unwrap_or_default(),
+            &input.custom_headers,
+        )
+        .await
     }
 
     pub async fn discover_models(&self, provider_id: i64) -> Result<DiscoveredModels> {
@@ -249,15 +367,14 @@ impl ControlService {
             .provider(provider_id)
             .await?
             .ok_or_else(|| Error::RunNotFound(format!("provider {provider_id}")))?;
-        let mut models = match provider.endpoint.provider_type {
-            ProviderType::OpenAiChat | ProviderType::OpenAiResponses => {
-                openai_models(&client, &provider).await?
-            }
-            ProviderType::Anthropic => anthropic_models(&client, &provider).await?,
-        };
-        models.sort();
-        models.dedup();
-        Ok(DiscoveredModels { models })
+        discover_provider_models(
+            &client,
+            provider.endpoint.provider_type,
+            &provider.endpoint.base_url,
+            provider.endpoint.api_key.as_deref().unwrap_or_default(),
+            &provider.custom_headers,
+        )
+        .await
     }
 
     pub async fn calls(&self, limit: i64) -> Result<Vec<CallSummary>> {
@@ -372,6 +489,14 @@ impl ControlService {
     pub async fn set_proxy_settings(&self, settings: ProxySettingsInput) -> Result<ProxySettings> {
         self.store.set_proxy_settings(settings).await
     }
+
+    pub async fn tab_settings(&self) -> Result<TabSettings> {
+        self.store.tab_settings().await
+    }
+
+    pub async fn set_tab_settings(&self, settings: TabSettings) -> Result<TabSettings> {
+        self.cursor_harness.set_tab_settings(settings).await
+    }
 }
 
 fn official_call(trace: CursorRunTraceSummary) -> CallSummary {
@@ -461,15 +586,51 @@ fn readable_utf8(data: &[u8]) -> Option<&str> {
         .then_some(value)
 }
 
+async fn discover_provider_models(
+    client: &reqwest::Client,
+    provider_type: ProviderType,
+    base_url: &str,
+    api_key: &str,
+    custom_headers: &serde_json::Value,
+) -> Result<DiscoveredModels> {
+    let mut models = match provider_type {
+        ProviderType::OpenAiChat | ProviderType::OpenAiResponses => {
+            openai_models(client, base_url, api_key, custom_headers).await?
+        }
+        ProviderType::Anthropic => {
+            anthropic_models(client, base_url, api_key, custom_headers).await?
+        }
+    };
+    models.sort();
+    models.dedup();
+    Ok(DiscoveredModels { models })
+}
+
+fn model_discovery_url(base_url: &str) -> Result<Url> {
+    let mut url = Url::parse(base_url)
+        .map_err(|error| Error::Config(format!("invalid provider base URL: {error}")))?;
+    if url.host_str().is_none() {
+        return Err(Error::Config(
+            "provider base URL must contain a host".into(),
+        ));
+    }
+    url.set_path("/v1/models");
+    url.set_query(None);
+    url.set_fragment(None);
+    Ok(url)
+}
+
 async fn openai_models(
     client: &reqwest::Client,
-    provider: &ProviderEndpointSecret,
+    base_url: &str,
+    api_key: &str,
+    custom_headers: &serde_json::Value,
 ) -> Result<Vec<String>> {
-    let mut request = client.get(format!("{}/models", provider.endpoint.base_url));
-    if !provider.api_key.is_empty() {
-        request = request.bearer_auth(&provider.api_key);
+    let mut request = client.get(model_discovery_url(base_url)?);
+    if !api_key.is_empty() {
+        request = request.bearer_auth(api_key);
     }
-    let response = apply_custom_headers(request, &provider.custom_headers)?
+    let response = apply_discovery_headers(request, custom_headers)?
         .send()
         .await?;
     let status = response.status();
@@ -484,22 +645,24 @@ async fn openai_models(
 
 async fn anthropic_models(
     client: &reqwest::Client,
-    provider: &ProviderEndpointSecret,
+    base_url: &str,
+    api_key: &str,
+    custom_headers: &serde_json::Value,
 ) -> Result<Vec<String>> {
     let mut after_id = None::<String>;
     let mut found = BTreeSet::new();
     loop {
         let mut request = client
-            .get(format!("{}/models", provider.endpoint.base_url))
+            .get(model_discovery_url(base_url)?)
             .query(&[("limit", "100")])
             .header("anthropic-version", "2023-06-01");
-        if !provider.api_key.is_empty() {
-            request = request.header("x-api-key", &provider.api_key);
+        if !api_key.is_empty() {
+            request = request.header("x-api-key", api_key);
         }
         if let Some(after_id) = &after_id {
             request = request.query(&[("after_id", after_id)]);
         }
-        let response = apply_custom_headers(request, &provider.custom_headers)?
+        let response = apply_discovery_headers(request, custom_headers)?
             .send()
             .await?;
         let status = response.status();
@@ -543,7 +706,18 @@ fn model_ids(value: &serde_json::Value) -> Vec<String> {
         .collect()
 }
 
-fn apply_custom_headers(
+fn estimate_output_tokens(output: &str) -> u64 {
+    let words = output.split_whitespace().count() as u64;
+    if words > 0 {
+        words
+    } else if output.is_empty() {
+        0
+    } else {
+        (output.chars().count() as u64).div_ceil(4)
+    }
+}
+
+fn apply_discovery_headers(
     mut request: reqwest::RequestBuilder,
     headers: &serde_json::Value,
 ) -> Result<reqwest::RequestBuilder> {
@@ -551,6 +725,9 @@ fn apply_custom_headers(
         .as_object()
         .ok_or_else(|| Error::Config("custom headers must be an object".into()))?;
     for (name, value) in object {
+        if name.eq_ignore_ascii_case("user-agent") {
+            continue;
+        }
         let value = value
             .as_str()
             .ok_or_else(|| Error::Config(format!("custom header {name} must be a string")))?;
@@ -562,3 +739,208 @@ fn apply_custom_headers(
     }
     Ok(request)
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::{Arc, Mutex};
+
+    use tokio_util::sync::CancellationToken;
+
+    use crate::{
+        model::{
+            ModelInvocation, ProjectedContent, ProviderEndpointInput, ProviderModelInput,
+            ProviderType,
+        },
+        provider::{FinishReason, ModelEvent, Provider, ProviderStream},
+        store::Store,
+    };
+
+    use super::ControlService;
+
+    struct TestProvider {
+        invocation: Arc<Mutex<Option<ModelInvocation>>>,
+    }
+
+    impl Provider for TestProvider {
+        fn stream(
+            &self,
+            invocation: ModelInvocation,
+            _cancellation: CancellationToken,
+        ) -> ProviderStream {
+            *self.invocation.lock().unwrap() = Some(invocation);
+            Box::pin(futures_util::stream::iter([
+                Ok(ModelEvent::Start {
+                    model_call_id: "test-call".into(),
+                }),
+                Ok(ModelEvent::TextStart),
+                Ok(ModelEvent::TextDelta("OK".into())),
+                Ok(ModelEvent::TextEnd),
+                Ok(ModelEvent::Usage(crate::model::Usage {
+                    output_tokens: Some(2),
+                    ..Default::default()
+                })),
+                Ok(ModelEvent::Done(FinishReason::Stop)),
+            ]))
+        }
+    }
+
+    #[tokio::test]
+    async fn connectivity_test_uses_the_configured_llm_provider() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("control.db").display()
+        ))
+        .await
+        .unwrap();
+        let invocation = Arc::new(Mutex::new(None));
+        let provider = store
+            .create_provider(&ProviderEndpointInput {
+                name: "Test".into(),
+                provider_type: ProviderType::OpenAiResponses,
+                base_url: "https://example.com/v1".into(),
+                api_key: None,
+                custom_headers: serde_json::json!({}),
+                extra_params: serde_json::json!({}),
+            })
+            .await
+            .unwrap();
+        let model = store
+            .save_provider_model(
+                provider.provider_id,
+                &ProviderModelInput {
+                    model_id: "reasoning-model".into(),
+                    display_name: "Reasoning Model".into(),
+                    endpoint_type: ProviderType::OpenAiResponses,
+                    request_url: String::new(),
+                    enabled: true,
+                    sort_order: 0,
+                    context_window_tokens: None,
+                    max_output_tokens: None,
+                    reasoning_enabled: true,
+                    reasoning_effort: None,
+                    supports_image_generation: false,
+                },
+            )
+            .await
+            .unwrap();
+        let service = ControlService::new(
+            store,
+            Arc::new(TestProvider {
+                invocation: invocation.clone(),
+            }),
+        )
+        .unwrap();
+
+        let result = service.test_model(&model.model_hash).await.unwrap();
+
+        assert_eq!(result.output, "OK");
+        assert_eq!(result.output_tokens, 2);
+        assert!(!result.tokens_estimated);
+        assert!(result.tokens_per_second > 0.0);
+        let invocation = invocation.lock().unwrap().clone().unwrap();
+        assert_eq!(invocation.request.model.model_id, model.model_hash);
+        assert!(invocation.request.model.reasoning.enabled);
+        assert_eq!(
+            invocation.request.model.reasoning.effort.as_deref(),
+            Some("medium")
+        );
+        assert!(invocation.request.prompt.tools.is_empty());
+        assert_eq!(invocation.request.history.len(), 1);
+        assert!(matches!(
+            &invocation.request.history[0].content,
+            ProjectedContent::Parts(parts)
+                if matches!(&parts[..], [crate::model::ContentPart::Text { text }] if text == "Output the numbers 1 through 120 separated by a single space. No commas, no newlines, no explanation.")
+        ));
+    }
+
+    #[test]
+    fn connectivity_output_token_estimate_handles_words_and_empty_text() {
+        assert_eq!(super::estimate_output_tokens("1 2 3"), 3);
+        assert_eq!(super::estimate_output_tokens(""), 0);
+    }
+
+    #[test]
+    fn model_discovery_url_uses_only_the_provider_origin() {
+        assert_eq!(
+            super::model_discovery_url("https://example.com:8443/arbitrary/v1/chat/completions")
+                .unwrap()
+                .as_str(),
+            "https://example.com:8443/v1/models"
+        );
+    }
+
+    #[tokio::test]
+    async fn model_discovery_does_not_inherit_user_agent_or_request_body_settings() {
+        type CapturedRequest = (
+            axum::http::Method,
+            axum::http::Uri,
+            axum::http::HeaderMap,
+            bytes::Bytes,
+        );
+
+        async fn models(
+            axum::extract::State(sender): axum::extract::State<
+                tokio::sync::mpsc::UnboundedSender<CapturedRequest>,
+            >,
+            request: axum::extract::Request,
+        ) -> axum::Json<serde_json::Value> {
+            let (parts, body) = request.into_parts();
+            let body = axum::body::to_bytes(body, usize::MAX).await.unwrap();
+            sender
+                .send((parts.method, parts.uri, parts.headers, body))
+                .unwrap();
+            axum::Json(serde_json::json!({ "data": [{ "id": "model-a" }] }))
+        }
+
+        let (sender, mut requests) = tokio::sync::mpsc::unbounded_channel();
+        let app = axum::Router::new()
+            .route("/v1/models", axum::routing::get(models))
+            .with_state(sender);
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        let server = tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("discovery.db").display()
+        ))
+        .await
+        .unwrap();
+        let service = ControlService::new(
+            store,
+            Arc::new(TestProvider {
+                invocation: Arc::new(Mutex::new(None)),
+            }),
+        )
+        .unwrap();
+        let result = service
+            .discover_input(&ProviderEndpointInput {
+                name: "Test".into(),
+                provider_type: ProviderType::OpenAiResponses,
+                base_url: format!("http://{address}/custom/responses"),
+                api_key: Some("secret".into()),
+                custom_headers: serde_json::json!({
+                    "uSeR-aGeNt": "inherited-user-agent",
+                    "x-tenant": "tenant-a"
+                }),
+                extra_params: serde_json::json!({ "temperature": 0.7 }),
+            })
+            .await
+            .unwrap();
+
+        assert_eq!(result.models, vec!["model-a"]);
+        let (method, uri, headers, body) = requests.recv().await.unwrap();
+        assert_eq!(method, axum::http::Method::GET);
+        assert_eq!(uri.path(), "/v1/models");
+        assert!(body.is_empty());
+        assert!(headers.get(axum::http::header::USER_AGENT).is_none());
+        assert_eq!(headers.get("x-tenant").unwrap(), "tenant-a");
+        assert_eq!(
+            headers.get(axum::http::header::AUTHORIZATION).unwrap(),
+            "Bearer secret"
+        );
+        server.abort();
+    }
+}

server/src/control/settings.rs

@@ -1,7 +1,9 @@
 use crate::Result;
 use axum::{extract::State, Json};
 
-use crate::store::{PortSettings, ProxySettings, ProxySettingsInput, StatisticsStorage};
+use crate::store::{
+    PortSettings, ProxySettings, ProxySettingsInput, StatisticsStorage, TabSettings,
+};
 
 use super::{ControlService, ObservabilitySettings};
 
@@ -47,3 +49,14 @@ pub async fn update_proxy(
 ) -> Result<Json<ProxySettings>> {
     Ok(Json(service.set_proxy_settings(settings).await?))
 }
+
+pub async fn get_tab(State(service): State<ControlService>) -> Result<Json<TabSettings>> {
+    Ok(Json(service.tab_settings().await?))
+}
+
+pub async fn update_tab(
+    State(service): State<ControlService>,
+    Json(settings): Json<TabSettings>,
+) -> Result<Json<TabSettings>> {
+    Ok(Json(service.set_tab_settings(settings).await?))
+}

server/src/cursor/bidi_append.rs

@@ -1,6 +1,7 @@
 use prost::Message;
 
 use crate::{
+    cursor::interaction,
     cursor::proto::{agent::v1 as agent, aiserver::v1 as ai},
     cursor::{CursorCommand, CursorParent, CursorSessionRegistry},
     Error, Result,
@@ -166,6 +167,12 @@ pub async fn append(
     if let Some(parent) = parent {
         handle.set_parent(parent)?;
     }
+    if matches!(
+        request.message.message.as_ref(),
+        Some(agent::agent_client_message::Message::ClientHeartbeat(_))
+    ) {
+        handle.emit(&interaction::heartbeat())?;
+    }
     handle
         .command(CursorCommand::Append {
             seqno: request.seqno,

server/src/cursor/handlers.rs

@@ -13,7 +13,7 @@ use crate::{
         observability::CursorTraceRecorder,
         proto::{agent::v1 as agent, aiserver::v1 as ai},
         proxy::{self, CursorProxy},
-        run_sse,
+        run_sse, tab,
     },
     cursor::{CursorParent, CursorSessionRegistry},
     Result,
@@ -70,6 +70,7 @@ fn router_with_proxy(registry: CursorSessionRegistry, proxy: CursorProxy) -> Rou
             post(analytics::bootstrap_statsig),
         )
         .route("/auth/full_stripe_profile", get(account::stripe_profile))
+        .merge(tab::router())
         .route_layer(DefaultBodyLimit::disable())
         .route_layer(RequestDecompressionLayer::new())
         .fallback(proxy::forward)

server/src/cursor/interaction/mod.rs

@@ -66,6 +66,12 @@ pub fn thinking_completed(elapsed: Duration) -> pb::AgentServerMessage {
     ))
 }
 
+pub fn heartbeat() -> pb::AgentServerMessage {
+    server_interaction(pb::interaction_update::Message::Heartbeat(
+        pb::HeartbeatUpdate {},
+    ))
+}
+
 pub fn arguments_delta(call: &ToolCall, delta: &str) -> Result<pb::AgentServerMessage> {
     Ok(server_interaction(
         pb::interaction_update::Message::PartialToolCall(pb::PartialToolCallUpdate {

server/src/cursor/mod.rs

@@ -22,6 +22,7 @@ pub mod request;
 pub mod run_sse;
 pub mod session;
 pub mod sessions;
+pub(crate) mod tab;
 pub mod tools;
 mod usage;
 

server/src/cursor/model_catalog.rs

@@ -14,7 +14,7 @@ use crate::{
         proxy::{self, CursorProxy},
         CursorSessionRegistry,
     },
-    model::ProviderModel,
+    model::{format_token_count, parse_token_count, ProviderModel},
     Error, Result,
 };
 
@@ -205,6 +205,23 @@ const EFFORTS: [(&str, &str); 5] = [
 ];
 const DEFAULT_CONTEXT: &str = "200k";
 
+fn context_options(model: &ProviderModel) -> Vec<(String, String)> {
+    let mut contexts = CONTEXTS
+        .into_iter()
+        .map(|(value, display_name)| (value.to_owned(), display_name.to_owned()))
+        .collect::<Vec<_>>();
+    if let Some(tokens) = model.context_window_tokens {
+        let value = tokens.to_string();
+        let duplicate = contexts
+            .iter()
+            .any(|(existing, _)| parse_token_count(existing) == Some(tokens));
+        if !duplicate {
+            contexts.push((value, format!("{} (Custom)", format_token_count(tokens))));
+        }
+    }
+    contexts
+}
+
 pub async fn available_models(
     State(registry): State<CursorSessionRegistry>,
     Extension(proxy): Extension<CursorProxy>,
@@ -324,7 +341,8 @@ fn unary_payload(body: &Bytes) -> Result<(bool, &[u8])> {
 }
 
 fn available_model(model: &ProviderModel, provider_name: &str) -> AvailableModel {
-    let variants = model_variants(model);
+    let contexts = context_options(model);
+    let variants = model_variants(model, &contexts);
     let legacy_slugs = variants
         .iter()
         .filter_map(|variant| variant.legacy_slug.clone())
@@ -348,7 +366,7 @@ fn available_model(model: &ProviderModel, provider_name: &str) -> AvailableModel
         inputbox_short_model_name: Some(model.display_name.clone()),
         supports_sandboxing: Some(true),
         supports_cmd_k: Some(false),
-        parameter_definitions: model_parameters(),
+        parameter_definitions: model_parameters(&contexts),
         variants,
         legacy_slugs,
         named_model_section_index: Some(1),
@@ -365,7 +383,7 @@ fn available_model(model: &ProviderModel, provider_name: &str) -> AvailableModel
     }
 }
 
-fn model_parameters() -> Vec<ModelParameterDefinition> {
+fn model_parameters(contexts: &[(String, String)]) -> Vec<ModelParameterDefinition> {
     vec![
         ModelParameterDefinition {
             id: "context".into(),
@@ -374,11 +392,11 @@ fn model_parameters() -> Vec<ModelParameterDefinition> {
             parameter_type: Some(ModelParameterType {
                 boolean_parameter: None,
                 enum_parameter: Some(EnumParameter {
-                    values: CONTEXTS
-                        .into_iter()
+                    values: contexts
+                        .iter()
                         .map(|(value, display_name)| EnumParameterValue {
-                            value: value.into(),
-                            display_name: Some(display_name.into()),
+                            value: value.clone(),
+                            display_name: Some(display_name.clone()),
                         })
                         .collect(),
                 }),
@@ -429,9 +447,9 @@ fn model_parameters() -> Vec<ModelParameterDefinition> {
     ]
 }
 
-fn model_variants(model: &ProviderModel) -> Vec<ModelVariant> {
-    let mut variants = Vec::with_capacity(CONTEXTS.len() * EFFORTS.len() * 2);
-    for (context, context_name) in CONTEXTS {
+fn model_variants(model: &ProviderModel, contexts: &[(String, String)]) -> Vec<ModelVariant> {
+    let mut variants = Vec::with_capacity(contexts.len() * EFFORTS.len() * 2);
+    for (context, context_name) in contexts {
         for (effort, effort_name) in EFFORTS {
             for fast in [false, true] {
                 variants.push(model_variant(
@@ -546,7 +564,7 @@ mod tests {
             request_url: String::new(),
             enabled: true,
             sort_order: 0,
-            context_window_tokens: Some(200_000),
+            context_window_tokens: Some(272_000),
             max_output_tokens: None,
             reasoning_enabled: false,
             reasoning_effort: None,
@@ -591,7 +609,22 @@ mod tests {
             .iter()
             .map(|value| value.value.as_str())
             .collect::<Vec<_>>();
-        assert_eq!(context_values, ["200k", "356k", "800k", "1m"]);
+        assert_eq!(context_values, ["200k", "356k", "800k", "1m", "272000"]);
+        let custom_context = context
+            .parameter_type
+            .as_ref()
+            .unwrap()
+            .enum_parameter
+            .as_ref()
+            .unwrap()
+            .values
+            .iter()
+            .find(|value| value.value == "272000")
+            .unwrap();
+        assert_eq!(
+            custom_context.display_name.as_deref(),
+            Some("272K (Custom)")
+        );
         let effort = mapped
             .parameter_definitions
             .iter()
@@ -607,8 +640,8 @@ mod tests {
             .values
             .iter()
             .any(|value| value.value == "max"));
-        assert_eq!(mapped.variants.len(), 40);
-        assert_eq!(mapped.legacy_slugs.len(), 40);
+        assert_eq!(mapped.variants.len(), 50);
+        assert_eq!(mapped.legacy_slugs.len(), 50);
         assert_eq!(mapped.model_picker_badges.len(), 1);
         assert_eq!(mapped.model_picker_badges[0].label, "OpenRouter");
         assert!(!mapped.model_picker_badges[0].dismiss_on_selection);

server/src/cursor/projection/decode.rs

@@ -30,17 +30,15 @@ pub fn decode(data: &[u8], internal_id: String) -> Result<CanonicalMessage> {
         .and_then(Value::as_str)
         .unwrap_or_default()
         .to_string();
+    let is_request_context = role == Role::User && wire_id.starts_with("request-context:");
+    let is_prompt_context =
+        is_request_context || role == Role::User && wire_id.starts_with("selected-context:");
     let origin = match role {
         Role::System => Origin::Prompt,
         Role::Assistant => Origin::Assistant,
         Role::Tool => Origin::Tool,
         Role::User if wire_id.starts_with("runtime:") => Origin::Runtime,
-        Role::User
-            if wire_id.starts_with("request-context:")
-                || wire_id.starts_with("selected-context:") =>
-        {
-            Origin::Prompt
-        }
+        Role::User if is_prompt_context => Origin::Prompt,
         Role::User => Origin::User,
     };
     let runtime_event_id = wire_id.strip_prefix("runtime:").map(str::to_string);
@@ -49,7 +47,7 @@ pub fn decode(data: &[u8], internal_id: String) -> Result<CanonicalMessage> {
         Role::Tool => MessageContent::ToolResult(decode_tool_result(&value)?),
         _ => decode_text(&value)?,
     };
-    let message_id = if runtime_event_id.is_some() {
+    let message_id = if runtime_event_id.is_some() || is_request_context {
         wire_id
     } else {
         internal_id

server/src/cursor/projection/tests.rs

@@ -131,6 +131,23 @@ fn runtime_wire_identity_survives_checkpoint_hydration() {
     );
 }
 
+#[test]
+fn request_context_identity_survives_checkpoint_hydration() {
+    let wire = json!({
+        "role": "user",
+        "id": "request-context:digest",
+        "content": "<rules>current rules</rules>",
+    });
+    let message = decode(
+        serde_json::to_vec(&wire).unwrap().as_slice(),
+        "cursor-root:blob-id:20".into(),
+    )
+    .unwrap();
+
+    assert_eq!(message.message_id, "request-context:digest");
+    assert_eq!(message.origin, crate::model::Origin::Prompt);
+}
+
 #[test]
 fn cursor_user_image_uses_image_field() {
     let wire = json!({

server/src/cursor/prompting/assets.rs

@@ -134,7 +134,6 @@ impl PromptAssets {
 }
 
 const RUNTIME_VARIABLES: &[&str] = &[
-    "REQUEST_CONTEXT",
     "OPEN_FILES",
     "SELECTED_CONTEXT",
     "ACTION_CONTEXT",

server/src/cursor/prompting/derived_state.rs

@@ -27,7 +27,9 @@ pub fn fold_derived_state(messages: &[CanonicalMessage]) -> DerivedState {
                     continue;
                 };
                 match normalize(&name).as_str() {
-                    "todowrite" | "updatetodos" => state.todos = Some(input),
+                    "todowrite" | "updatetodos" => {
+                        state.todos = Some(apply_todo_write(state.todos.take(), input));
+                    }
                     "createplan" | "updateplan" | "writeplan" => state.plan = Some(input),
                     _ => {}
                 }
@@ -38,6 +40,39 @@ pub fn fold_derived_state(messages: &[CanonicalMessage]) -> DerivedState {
     state
 }
 
+fn apply_todo_write(current: Option<Value>, mut input: Value) -> Value {
+    if !input.get("merge").and_then(Value::as_bool).unwrap_or(false) {
+        return input;
+    }
+    let mut todos = current
+        .as_ref()
+        .and_then(|value| value.get("todos"))
+        .and_then(Value::as_array)
+        .cloned()
+        .unwrap_or_default();
+    let patches = input
+        .get("todos")
+        .and_then(Value::as_array)
+        .cloned()
+        .unwrap_or_default();
+    for patch in patches {
+        let existing = patch.get("id").and_then(Value::as_str).and_then(|id| {
+            todos
+                .iter_mut()
+                .find(|todo| todo.get("id").and_then(Value::as_str) == Some(id))
+        });
+        match (existing, patch) {
+            (Some(Value::Object(todo)), Value::Object(patch)) => todo.extend(patch),
+            (_, patch) => todos.push(patch),
+        }
+    }
+    if let Some(object) = input.as_object_mut() {
+        object.insert("merge".into(), Value::Bool(false));
+        object.insert("todos".into(), Value::Array(todos));
+    }
+    input
+}
+
 fn normalize(value: &str) -> String {
     value
         .chars()
@@ -45,3 +80,87 @@ fn normalize(value: &str) -> String {
         .flat_map(char::to_lowercase)
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::model::{Origin, Role, ToolCallContent, ToolResultContent};
+
+    #[test]
+    fn todo_write_merge_materializes_complete_existing_items_and_appends_new_ids() {
+        let messages = vec![
+            assistant_call(
+                "create",
+                serde_json::json!({
+                    "merge": false,
+                    "todos": [
+                        {"id": "first", "content": "First", "status": "in_progress"},
+                        {"id": "second", "content": "Second", "status": "pending"}
+                    ]
+                }),
+            ),
+            successful_result("create"),
+            assistant_call(
+                "merge",
+                serde_json::json!({
+                    "merge": true,
+                    "todos": [
+                        {"id": "first", "status": "completed"},
+                        {"id": "second", "content": "Second updated"},
+                        {"id": "third", "content": "Third", "status": "cancelled"}
+                    ]
+                }),
+            ),
+            successful_result("merge"),
+        ];
+
+        let state = fold_derived_state(&messages);
+
+        assert_eq!(
+            state.todos.unwrap()["todos"],
+            serde_json::json!([
+                {"id": "first", "content": "First", "status": "completed"},
+                {"id": "second", "content": "Second updated", "status": "pending"},
+                {"id": "third", "content": "Third", "status": "cancelled"}
+            ])
+        );
+    }
+
+    fn assistant_call(call_id: &str, arguments: Value) -> CanonicalMessage {
+        CanonicalMessage {
+            message_id: format!("assistant-{call_id}"),
+            role: Role::Assistant,
+            origin: Origin::Assistant,
+            content: MessageContent::Assistant {
+                text: String::new(),
+                thinking: String::new(),
+                tool_round_id: Some(format!("round-{call_id}").into()),
+                replay_state: None,
+                tool_calls: vec![ToolCallContent {
+                    index: 0,
+                    call_id: call_id.into(),
+                    name: "TodoWrite".into(),
+                    arguments,
+                }],
+            },
+            runtime_event_id: None,
+        }
+    }
+
+    fn successful_result(call_id: &str) -> CanonicalMessage {
+        CanonicalMessage {
+            message_id: format!("result-{call_id}"),
+            role: Role::Tool,
+            origin: Origin::Tool,
+            content: MessageContent::ToolResult(ToolResultContent {
+                call_id: call_id.into(),
+                name: "TodoWrite".into(),
+                content: "{}".into(),
+                is_error: false,
+                image: None,
+                provider_parts: Vec::new(),
+            }),
+            runtime_event_id: None,
+        }
+    }
+}

server/src/cursor/proxy.rs

@@ -82,14 +82,34 @@ impl CursorProxy {
 pub async fn forward(
     Extension(proxy): Extension<CursorProxy>,
     request: Request<Body>,
+) -> Result<Response<Body>> {
+    forward_request(&proxy, request, None).await
+}
+
+pub(crate) async fn forward_to_service(
+    proxy: &CursorProxy,
+    request: Request<Body>,
+    service_url: &str,
+) -> Result<Response<Body>> {
+    forward_request(proxy, request, Some(service_url)).await
+}
+
+async fn forward_request(
+    proxy: &CursorProxy,
+    request: Request<Body>,
+    service_url: Option<&str>,
 ) -> Result<Response<Body>> {
     let started = Instant::now();
     let (parts, body) = request.into_parts();
     let path = parts
         .uri
         .path_and_query()
-        .map_or("/", |value| value.as_str());
-    let url = upstream_url(&parts.headers, &proxy.upstream, path)?;
+        .map_or("/", |value| value.as_str())
+        .to_owned();
+    let url = match service_url {
+        Some(service_url) => format!("{}{}", service_url.trim_end_matches('/'), path),
+        None => upstream_url(&parts.headers, &proxy.upstream, &path)?,
+    };
 
     let mut headers = parts.headers;
     headers.remove(UPSTREAM_URL_HEADER);
@@ -239,7 +259,7 @@ mod tests {
     };
     use tower::ServiceExt;
 
-    use super::{forward, CursorProxy};
+    use super::{forward, forward_to_service, CursorProxy};
 
     #[tokio::test]
     async fn preserves_request_and_response() {
@@ -289,4 +309,34 @@ mod tests {
         );
         server.abort();
     }
+
+    #[tokio::test]
+    async fn tab_service_keeps_its_base_path_and_the_original_query() {
+        let upstream = Router::new().route(
+            "/base/aiserver.v1.AiService/StreamCpp",
+            any(|request: Request<Body>| async move {
+                request.uri().path_and_query().unwrap().as_str().to_owned()
+            }),
+        );
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        let server = tokio::spawn(async move { axum::serve(listener, upstream).await.unwrap() });
+        let proxy = CursorProxy::for_upstream("http://unused.invalid").unwrap();
+
+        let response = forward_to_service(
+            &proxy,
+            Request::post("/aiserver.v1.AiService/StreamCpp?client=cursor")
+                .body(Body::empty())
+                .unwrap(),
+            &format!("http://{address}/base"),
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(
+            to_bytes(response.into_body(), usize::MAX).await.unwrap(),
+            "/base/aiserver.v1.AiService/StreamCpp?client=cursor"
+        );
+        server.abort();
+    }
 }

server/src/cursor/request/context.rs

@@ -459,24 +459,74 @@ pub fn dynamic_mcp(
                 Error::Protocol(format!("MCP tool {} is missing input schema", wire.name))
             })?),
         };
+        let parameters = normalize_mcp_parameters(&wire.name, parameters)?;
+        let name = model_tool_name(&wire.name);
         let definition = ToolDefinition {
-            name: wire.name.clone(),
+            name: name.clone(),
             description: wire.description.clone(),
             parameters,
         };
         if output
-            .insert(wire.name.clone(), (wire.clone(), definition))
+            .insert(name.clone(), (wire.clone(), definition))
             .is_some()
         {
             return Err(Error::Protocol(format!(
-                "duplicate MCP tool definition: {}",
-                wire.name
+                "duplicate MCP tool name after normalization: {name}"
             )));
         }
     }
     Ok(output)
 }
 
+fn normalize_mcp_parameters(tool_name: &str, mut parameters: Value) -> Result<Value> {
+    let schema = parameters
+        .as_object_mut()
+        .ok_or_else(|| invalid_mcp_parameters(tool_name))?;
+    match schema.get("type") {
+        Some(Value::String(schema_type)) if schema_type == "object" => return Ok(parameters),
+        Some(_) => return Err(invalid_mcp_parameters(tool_name)),
+        None => {}
+    }
+    let object_only_union = ["anyOf", "oneOf"].into_iter().any(|keyword| {
+        schema
+            .get(keyword)
+            .and_then(Value::as_array)
+            .is_some_and(|branches| {
+                !branches.is_empty()
+                    && branches.iter().all(|branch| {
+                        branch
+                            .as_object()
+                            .and_then(|branch| branch.get("type"))
+                            .and_then(Value::as_str)
+                            == Some("object")
+                    })
+            })
+    });
+    if !object_only_union {
+        return Err(invalid_mcp_parameters(tool_name));
+    }
+    schema.insert("type".into(), Value::String("object".into()));
+    Ok(parameters)
+}
+
+fn invalid_mcp_parameters(tool_name: &str) -> Error {
+    Error::Protocol(format!(
+        "MCP tool {tool_name} input schema must describe an object"
+    ))
+}
+
+fn model_tool_name(name: &str) -> String {
+    name.chars()
+        .map(|character| {
+            if character.is_ascii_alphanumeric() || matches!(character, '_' | '-') {
+                character
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
 fn prost_value(value: &prost_types::Value) -> Value {
     use prost_types::value::Kind;
     match value.kind.as_ref() {
@@ -511,6 +561,167 @@ fn xml(value: &str) -> String {
 mod tests {
     use super::*;
 
+    fn direct_mcp_tool(name: &str) -> pb::McpToolDefinition {
+        pb::McpToolDefinition {
+            name: name.into(),
+            provider_identifier: "extension-GitKraken".into(),
+            tool_name: "git_status".into(),
+            description: "Get repository status".into(),
+            input_schema_json: Some(r#"{"type":"object"}"#.into()),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn dynamic_mcp_normalizes_extension_identifier_for_model_tool_names() {
+        let original = "user-eamodio.gitlens-extension-GitKraken-git_status";
+        let request = pb::AgentRunRequest {
+            mcp_tools: Some(pb::McpTools {
+                mcp_tools: vec![direct_mcp_tool(original)],
+            }),
+            ..Default::default()
+        };
+
+        let tools = dynamic_mcp(&request, &pb::RequestContext::default()).unwrap();
+        let normalized = "user-eamodio_gitlens-extension-GitKraken-git_status";
+        let (wire, definition) = tools.get(normalized).unwrap();
+
+        assert_eq!(definition.name, normalized);
+        assert_eq!(wire.name, original);
+        assert_eq!(wire.provider_identifier, "extension-GitKraken");
+        assert_eq!(wire.tool_name, "git_status");
+        assert!(normalized
+            .chars()
+            .all(|character| character.is_ascii_alphanumeric() || matches!(character, '_' | '-')));
+    }
+
+    #[test]
+    fn dynamic_mcp_rejects_names_that_collide_after_normalization() {
+        let request = pb::AgentRunRequest {
+            mcp_tools: Some(pb::McpTools {
+                mcp_tools: vec![
+                    direct_mcp_tool("server.name-tool"),
+                    direct_mcp_tool("server_name-tool"),
+                ],
+            }),
+            ..Default::default()
+        };
+
+        let error = dynamic_mcp(&request, &pb::RequestContext::default()).unwrap_err();
+        assert!(error
+            .to_string()
+            .contains("duplicate MCP tool name after normalization: server_name-tool"));
+    }
+
+    #[test]
+    fn dynamic_mcp_normalizes_cursor_object_union_without_mutating_wire_schema() {
+        let original_schema = serde_json::json!({
+            "$schema": "https://json-schema.org/draft/2020-12/schema",
+            "anyOf": [
+                {
+                    "type": "object",
+                    "properties": {
+                        "rootPath": { "type": "string", "minLength": 1 }
+                    },
+                    "required": ["rootPath"],
+                    "additionalProperties": false
+                },
+                {
+                    "type": "object",
+                    "properties": {
+                        "rootPaths": {
+                            "type": "array",
+                            "items": { "type": "string", "minLength": 1 },
+                            "minItems": 1
+                        }
+                    },
+                    "required": ["rootPaths"],
+                    "additionalProperties": false
+                }
+            ]
+        });
+        let original_json = original_schema.to_string();
+        let mut tool = direct_mcp_tool("cursor-app-control-move_agent_to_cloned_root");
+        tool.input_schema_json = Some(original_json.clone());
+        let request = pb::AgentRunRequest {
+            mcp_tools: Some(pb::McpTools {
+                mcp_tools: vec![tool],
+            }),
+            ..Default::default()
+        };
+
+        let tools = dynamic_mcp(&request, &pb::RequestContext::default()).unwrap();
+        let (wire, definition) = tools
+            .get("cursor-app-control-move_agent_to_cloned_root")
+            .unwrap();
+
+        assert_eq!(definition.parameters["type"], "object");
+        assert_eq!(definition.parameters["anyOf"], original_schema["anyOf"]);
+        assert_eq!(
+            wire.input_schema_json.as_deref(),
+            Some(original_json.as_str())
+        );
+    }
+
+    #[test]
+    fn dynamic_mcp_preserves_valid_object_schema() {
+        let original_schema = serde_json::json!({
+            "type": "object",
+            "properties": {
+                "query": { "type": "string" }
+            },
+            "required": ["query"],
+            "additionalProperties": false
+        });
+        let mut tool = direct_mcp_tool("search");
+        tool.input_schema_json = Some(original_schema.to_string());
+        let request = pb::AgentRunRequest {
+            mcp_tools: Some(pb::McpTools {
+                mcp_tools: vec![tool],
+            }),
+            ..Default::default()
+        };
+
+        let tools = dynamic_mcp(&request, &pb::RequestContext::default()).unwrap();
+        let (_, definition) = tools.get("search").unwrap();
+
+        assert_eq!(definition.parameters, original_schema);
+    }
+
+    #[test]
+    fn dynamic_mcp_rejects_schemas_that_are_not_provably_objects() {
+        let invalid_schemas = [
+            serde_json::Value::Null,
+            serde_json::json!({ "type": "string" }),
+            serde_json::json!({ "properties": { "query": { "type": "string" } } }),
+            serde_json::json!({
+                "anyOf": [
+                    { "type": "object" },
+                    { "type": "string" }
+                ]
+            }),
+        ];
+
+        for schema in invalid_schemas {
+            let mut tool = direct_mcp_tool("unsafe_schema");
+            tool.input_schema_json = Some(schema.to_string());
+            let request = pb::AgentRunRequest {
+                mcp_tools: Some(pb::McpTools {
+                    mcp_tools: vec![tool],
+                }),
+                ..Default::default()
+            };
+
+            let error = dynamic_mcp(&request, &pb::RequestContext::default()).unwrap_err();
+            assert!(
+                error
+                    .to_string()
+                    .contains("MCP tool unsafe_schema input schema must describe an object"),
+                "unexpected error for {schema}: {error}"
+            );
+        }
+    }
+
     #[test]
     fn meta_mcp_routes_projects_descriptor_routing_without_runtime_discovery() {
         let context = pb::RequestContext {

server/src/cursor/request/model.rs

@@ -1,6 +1,9 @@
 use crate::{
     cursor::proto::agent::v1 as pb,
-    model::{ModelLatency, ModelSpec, ReasoningSpec, SubagentKind, SubagentModelOverride},
+    model::{
+        parse_token_count, ModelLatency, ModelSpec, ReasoningSpec, SubagentKind,
+        SubagentModelOverride,
+    },
     Error, Result,
 };
 
@@ -136,16 +139,6 @@ fn parse_bool(parameter: &pb::requested_model::ModelParameterValue) -> Result<bo
     }
 }
 
-fn parse_token_count(value: &str) -> Option<u64> {
-    let value = value.trim().to_ascii_lowercase();
-    let (number, multiplier) = match value.chars().last()? {
-        'k' => (&value[..value.len() - 1], 1_000),
-        'm' => (&value[..value.len() - 1], 1_000_000),
-        _ => (value.as_str(), 1),
-    };
-    number.parse::<u64>().ok()?.checked_mul(multiplier)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;

server/src/cursor/request/prepare.rs

@@ -192,6 +192,20 @@ pub(crate) async fn prepare(
         }
         _ => None,
     };
+    let request_context_message = match event_id.as_deref() {
+        Some(event_id) if !compacting && !background_completion => {
+            let message_id = format!("request-context:{event_id}");
+            match store.message(&conversation_id, &message_id).await? {
+                Some(message) => Some(message),
+                None => runtime::compile_request_context(
+                    event_id,
+                    &request_context,
+                    base_messages.as_deref().unwrap_or_default(),
+                )?,
+            }
+        }
+        _ => None,
+    };
     let initial_messages = if compacting {
         Vec::new()
     } else {
@@ -209,21 +223,27 @@ pub(crate) async fn prepare(
                 turn_user = Some(user);
                 vec![message]
             }
-            (Some(user), Some(event_id)) => match existing_runtime {
-                Some(message) => vec![message],
-                None => vec![
-                    runtime::compile(
-                        event_id,
-                        checkpoint_mode,
-                        &user,
-                        &request_context,
-                        &action_context,
-                        compiler,
-                        blob_sync,
-                    )
-                    .await?,
-                ],
-            },
+            (Some(user), Some(event_id)) => {
+                let runtime = match existing_runtime {
+                    Some(message) => message,
+                    None => {
+                        runtime::compile(
+                            event_id,
+                            checkpoint_mode,
+                            &user,
+                            &request_context,
+                            &action_context,
+                            compiler,
+                            blob_sync,
+                        )
+                        .await?
+                    }
+                };
+                request_context_message
+                    .into_iter()
+                    .chain(std::iter::once(runtime))
+                    .collect()
+            }
             (None, None) => Vec::new(),
             _ => {
                 return Err(Error::Protocol(
@@ -254,20 +274,7 @@ pub(crate) async fn prepare(
         };
         RunAction::Resume { pending_tool_round }
     };
-    let kind = match (request.subagent_type_name.as_deref(), parent) {
-        (None, _) => RunKind::Root,
-        (Some(name), Some((parent_run_id, parent_tool_call_id))) => RunKind::Subagent {
-            parent_run_id,
-            parent_tool_call_id,
-            kind: model::subagent_kind(name),
-            background: false,
-        },
-        (Some(_), None) => {
-            return Err(Error::Protocol(
-                "subagent Run is missing its parent Run and tool call".into(),
-            ));
-        }
-    };
+    let kind = run_kind(request.subagent_type_name.as_deref(), parent)?;
     let exec = exec_context(
         request,
         &request_context,
@@ -302,6 +309,21 @@ pub(crate) async fn prepare(
     ))
 }
 
+fn run_kind(subagent_type_name: Option<&str>, parent: Option<(RunId, String)>) -> Result<RunKind> {
+    match (subagent_type_name, parent) {
+        (None | Some("side-chat"), _) => Ok(RunKind::Root),
+        (Some(name), Some((parent_run_id, parent_tool_call_id))) => Ok(RunKind::Subagent {
+            parent_run_id,
+            parent_tool_call_id,
+            kind: model::subagent_kind(name),
+            background: false,
+        }),
+        (Some(_), None) => Err(Error::Protocol(
+            "subagent Run is missing its parent Run and tool call".into(),
+        )),
+    }
+}
+
 fn validate_prompt_root(messages: &[CanonicalMessage]) -> Result<()> {
     let prompts = messages
         .iter()
@@ -559,6 +581,23 @@ mod tests {
         assert!(mode_from_proto(99).is_err());
     }
 
+    #[test]
+    fn side_chat_without_task_parent_is_an_independent_root_run() {
+        assert!(matches!(
+            run_kind(Some("side-chat"), None).unwrap(),
+            RunKind::Root
+        ));
+    }
+
+    #[test]
+    fn task_subagent_without_parent_is_still_rejected() {
+        assert!(matches!(
+            run_kind(Some("explore"), None),
+            Err(Error::Protocol(message))
+                if message == "subagent Run is missing its parent Run and tool call"
+        ));
+    }
+
     #[test]
     fn current_user_message_consumes_the_mode_instead_of_history_mode() {
         let request = pb::AgentRunRequest {

server/src/cursor/request/runtime.rs

@@ -87,10 +87,6 @@ pub async fn compile(
             .map(|env| env.time_zone.as_str()),
     )?;
     let mut values = BTreeMap::from([
-        (
-            "REQUEST_CONTEXT",
-            section(context::compile_context(request_context, &time.today)),
-        ),
         ("OPEN_FILES", section(open_files(user))),
         (
             "SELECTED_CONTEXT",
@@ -122,6 +118,41 @@ pub async fn compile(
     .await
 }
 
+pub(super) fn compile_request_context(
+    event_id: &str,
+    request_context: &pb::RequestContext,
+    history: &[CanonicalMessage],
+) -> Result<Option<CanonicalMessage>> {
+    let time = Time::now(
+        request_context
+            .env
+            .as_ref()
+            .map(|env| env.time_zone.as_str()),
+    )?;
+    let text = context::compile_context(request_context, &time.today);
+    if text.is_empty() {
+        return Ok(None);
+    }
+    let message = CanonicalMessage::text(
+        format!("request-context:{event_id}"),
+        Role::User,
+        Origin::Prompt,
+        text,
+    );
+    Ok(should_project_request_context(history, &message).then_some(message))
+}
+
+fn should_project_request_context(
+    history: &[CanonicalMessage],
+    current: &CanonicalMessage,
+) -> bool {
+    history
+        .iter()
+        .rev()
+        .find(|message| message.message_id.starts_with("request-context:"))
+        .is_none_or(|previous| previous.content != current.content)
+}
+
 pub async fn compile_background(
     event_id: String,
     user: &pb::UserMessage,
@@ -252,3 +283,55 @@ impl Time {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn request_context_is_only_projected_when_its_content_changes() {
+        let first = CanonicalMessage::text(
+            "request-context:first",
+            Role::User,
+            Origin::Prompt,
+            "<rules>same</rules>",
+        );
+        let duplicate = CanonicalMessage::text(
+            "request-context:second",
+            Role::User,
+            Origin::Prompt,
+            "<rules>same</rules>",
+        );
+        let changed = CanonicalMessage::text(
+            "request-context:third",
+            Role::User,
+            Origin::Prompt,
+            "<rules>changed</rules>",
+        );
+        let runtime = CanonicalMessage::text(
+            "runtime:turn",
+            Role::User,
+            Origin::Runtime,
+            "<user_query>next</user_query>",
+        );
+
+        assert!(should_project_request_context(&[], &first));
+        assert!(!should_project_request_context(
+            &[first.clone(), runtime.clone()],
+            &duplicate
+        ));
+        assert!(should_project_request_context(
+            &[first.clone(), runtime.clone()],
+            &changed
+        ));
+        assert!(should_project_request_context(
+            &[first.clone(), changed, runtime],
+            &CanonicalMessage::text(
+                "request-context:fourth",
+                Role::User,
+                Origin::Prompt,
+                "<rules>same</rules>",
+            )
+        ));
+    }
+}

server/src/cursor/run_sse.rs

@@ -3,29 +3,24 @@ use axum::{
     http::{header, HeaderValue, Response, StatusCode},
 };
 use bytes::Bytes;
+use std::convert::Infallible;
 use tokio::sync::mpsc;
 use tokio_stream::StreamExt;
+use tokio_util::sync::CancellationToken;
 
 use crate::{
-    cursor::{observability::CursorTraceRecorder, CursorSessionRegistry},
+    cursor::{connect::END_STREAM_FLAG, observability::CursorTraceRecorder, CursorSessionRegistry},
     Result,
 };
 
 pub async fn stream(registry: &CursorSessionRegistry, request_id: &str) -> Result<Response<Body>> {
     let handle = registry.get_or_create(request_id).await?;
-    let mut receiver = handle.subscribe();
+    let receiver = handle.subscribe();
     let trace = handle.trace().cloned();
     if let Some(trace) = &trace {
         trace.response_started(StatusCode::OK.as_u16()).await;
     }
-    let body_stream = async_stream::stream! {
-        let mut trace = TraceStreamSink::new(trace, "byok_server");
-        while let Some(chunk) = receiver.recv().await {
-            trace.chunk(&chunk);
-            yield Ok::<Bytes, std::convert::Infallible>(chunk);
-        }
-        trace.finish(None);
-    };
+    let body_stream = local_body_stream(receiver, handle.cancellation(), trace);
     let mut response = Response::new(Body::from_stream(body_stream));
     *response.status_mut() = StatusCode::OK;
     response.headers_mut().insert(
@@ -41,6 +36,63 @@ pub async fn stream(registry: &CursorSessionRegistry, request_id: &str) -> Resul
     Ok(response)
 }
 
+fn local_body_stream(
+    mut receiver: mpsc::UnboundedReceiver<Bytes>,
+    cancellation: CancellationToken,
+    trace: Option<CursorTraceRecorder>,
+) -> impl tokio_stream::Stream<Item = std::result::Result<Bytes, Infallible>> {
+    async_stream::stream! {
+        let mut guard = LocalRunGuard::new(cancellation);
+        let mut trace = TraceStreamSink::new(trace, "byok_server");
+        while let Some(chunk) = receiver.recv().await {
+            let terminal = is_end_stream_frame(&chunk);
+            trace.chunk(&chunk);
+            if terminal {
+                guard.complete();
+                trace.finish(None);
+            }
+            yield Ok::<Bytes, Infallible>(chunk);
+            if terminal {
+                return;
+            }
+        }
+        guard.complete();
+        trace.finish(None);
+    }
+}
+
+fn is_end_stream_frame(frame: &Bytes) -> bool {
+    frame
+        .first()
+        .is_some_and(|flags| flags & END_STREAM_FLAG != 0)
+}
+
+struct LocalRunGuard {
+    cancellation: CancellationToken,
+    completed: bool,
+}
+
+impl LocalRunGuard {
+    fn new(cancellation: CancellationToken) -> Self {
+        Self {
+            cancellation,
+            completed: false,
+        }
+    }
+
+    fn complete(&mut self) {
+        self.completed = true;
+    }
+}
+
+impl Drop for LocalRunGuard {
+    fn drop(&mut self) {
+        if !self.completed {
+            self.cancellation.cancel();
+        }
+    }
+}
+
 pub async fn upstream(
     registry: CursorSessionRegistry,
     request_id: String,
@@ -127,7 +179,11 @@ impl TraceStreamSink {
 
 impl Drop for TraceStreamSink {
     fn drop(&mut self) {
-        self.finish(None);
+        if self.sender.is_some() {
+            self.finish(Some(
+                "response stream dropped before completion".to_string(),
+            ));
+        }
     }
 }
 
@@ -143,3 +199,38 @@ impl Drop for UpstreamRunGuard {
             .finish_upstream(self.request_id.clone(), self.generation);
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::cursor::{connect, proto::agent::v1 as pb};
+
+    #[tokio::test]
+    async fn local_stream_cancels_when_the_client_disconnects() {
+        let (sender, receiver) = mpsc::unbounded_channel();
+        let cancellation = CancellationToken::new();
+        sender
+            .send(connect::encode_message(&pb::AgentServerMessage::default()).unwrap())
+            .unwrap();
+        let mut stream = Box::pin(local_body_stream(receiver, cancellation.clone(), None));
+
+        stream.next().await.unwrap().unwrap();
+
+        drop(sender);
+        drop(stream);
+        assert!(cancellation.is_cancelled());
+    }
+
+    #[tokio::test]
+    async fn terminal_frame_does_not_cancel_a_completed_local_run() {
+        let (sender, receiver) = mpsc::unbounded_channel();
+        let cancellation = CancellationToken::new();
+        sender.send(connect::encode_end_stream()).unwrap();
+        let mut stream = Box::pin(local_body_stream(receiver, cancellation.clone(), None));
+
+        let terminal = stream.next().await.unwrap().unwrap();
+        assert!(is_end_stream_frame(&terminal));
+        drop(stream);
+        assert!(!cancellation.is_cancelled());
+    }
+}

server/src/cursor/session.rs

@@ -129,11 +129,19 @@ impl CursorSession {
                     checkpoint_worker_open = false;
                 }
                 Input::Completion(completion) => {
-                    self.forward_completion(completion, &mut completions)
-                        .await?;
+                    if let Some(completion) = self
+                        .forward_completion(completion, &mut completions)
+                        .await?
+                    {
+                        ready.push_back(completion);
+                    }
                 }
                 Input::CompletionResult(Some(result)) => {
-                    self.forward_completion(result?, &mut completions).await?;
+                    if let Some(completion) =
+                        self.forward_completion(result?, &mut completions).await?
+                    {
+                        ready.push_back(completion);
+                    }
                 }
                 Input::CompletionResult(None) => {
                     return Err(Error::Protocol("tool result channel closed".into()));
@@ -578,7 +586,7 @@ impl CursorSession {
         &self,
         mut completion: ToolCompletion,
         completions: &mut HashMap<String, ToolCompletion>,
-    ) -> Result<()> {
+    ) -> Result<Option<ToolCompletion>> {
         if let Some(image) = completion.take_read_image() {
             let blob_id = self.store.put_blob(&image.data, &[]).await?;
             completion.persist_read_image(&blob_id, &image)?;
@@ -600,7 +608,14 @@ impl CursorSession {
             .commands
             .send(ClientCommand::ToolResult(result.clone()))
             .await
-            .map_err(|_| Error::RunNotFound(self.context.request_id.clone()))
+            .map_err(|_| Error::RunNotFound(self.context.request_id.clone()))?;
+        let Some(dispatched) = self.tools.continue_after(&result.call_id).await? else {
+            return Ok(None);
+        };
+        for message in dispatched.messages {
+            self.handle.emit(&message)?;
+        }
+        Ok(dispatched.completion)
     }
 
     async fn forward_injection(&mut self, action: pb::InjectContextAction) -> Result<()> {

server/src/cursor/tab.rs

@@ -0,0 +1,67 @@
+use axum::{
+    body::Body,
+    extract::{Extension, State},
+    http::{Request, Response},
+    routing::post,
+    Router,
+};
+
+use crate::{
+    cursor::{proxy, CursorSessionRegistry},
+    Result,
+};
+
+pub const TAB_PATHS: [&str; 17] = [
+    "/aiserver.v1.AiService/StreamCpp",
+    "/aiserver.v1.AiService/StreamNextCursorPrediction",
+    "/aiserver.v1.AiService/GetCppEditClassification",
+    "/aiserver.v1.AiService/RefreshTabContext",
+    "/aiserver.v1.AiService/CppConfig",
+    "/aiserver.v1.AiService/CppEditHistoryStatus",
+    "/aiserver.v1.AiService/CppAppend",
+    "/aiserver.v1.AiService/CppEditHistoryAppend",
+    "/aiserver.v1.AiService/ReportAiCodeChangeMetrics",
+    "/aiserver.v1.AiService/WriteGitCommitMessage",
+    "/aiserver.v1.AiService/WriteGitBranchName",
+    "/aiserver.v1.CppService/AvailableModels",
+    "/aiserver.v1.CppService/RecordCppFate",
+    "/aiserver.v1.FileSyncService/FSSyncFile",
+    "/aiserver.v1.FileSyncService/FSIsEnabledForUser",
+    "/aiserver.v1.FileSyncService/FSConfig",
+    "/aiserver.v1.FileSyncService/FSUploadFile",
+];
+
+pub fn is_tab_path(path: &str) -> bool {
+    TAB_PATHS.contains(&path)
+}
+
+pub fn router() -> Router<CursorSessionRegistry> {
+    TAB_PATHS.into_iter().fold(Router::new(), |router, path| {
+        router.route(path, post(forward))
+    })
+}
+
+async fn forward(
+    State(registry): State<CursorSessionRegistry>,
+    Extension(upstream): Extension<proxy::CursorProxy>,
+    request: Request<Body>,
+) -> Result<Response<Body>> {
+    let settings = registry.store().tab_settings().await?;
+    match settings.service_url() {
+        Some(service_url) => proxy::forward_to_service(&upstream, request, service_url).await,
+        None => proxy::forward(Extension(upstream), request).await,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn matches_only_legacy_tab_routes() {
+        assert_eq!(TAB_PATHS.len(), 17);
+        assert!(is_tab_path("/aiserver.v1.AiService/StreamCpp"));
+        assert!(is_tab_path("/aiserver.v1.FileSyncService/FSUploadFile"));
+        assert!(!is_tab_path("/aiserver.v1.AiService/AvailableModels"));
+    }
+}

server/src/cursor/tools/codec/request.rs

@@ -49,6 +49,7 @@ pub fn request(id: u32, call: &ToolCall, context: &ExecContext) -> Result<pb::Ag
                 "request_smart_mode_approval",
                 "smart_mode_block_reason",
             )?,
+            requested_sandbox_policy: shell_sandbox_policy(call),
             close_stdin: true,
             conversation_id: Some(context.conversation_id.clone()),
             admin_command_denylist: context.admin_command_denylist.clone(),
@@ -365,6 +366,29 @@ pub fn abort(id: u32) -> pb::AgentServerMessage {
     }
 }
 
+fn shell_sandbox_policy(call: &ToolCall) -> Option<pb::SandboxPolicy> {
+    let permissions = call.arguments.get("required_permissions")?.as_array()?;
+    let perms: Vec<&str> = permissions
+        .iter()
+        .filter_map(Value::as_str)
+        .collect();
+    if perms.contains(&"all") {
+        Some(pb::SandboxPolicy {
+            r#type: pb::sandbox_policy::Type::InsecureNone as i32,
+            network_access: Some(true),
+            ..Default::default()
+        })
+    } else if perms.contains(&"full_network") {
+        Some(pb::SandboxPolicy {
+            r#type: pb::sandbox_policy::Type::WorkspaceReadwrite as i32,
+            network_access: Some(true),
+            ..Default::default()
+        })
+    } else {
+        None
+    }
+}
+
 fn shell_timeout(call: &ToolCall) -> Result<i32> {
     let value = call
         .arguments

server/src/cursor/tools/edit.rs

@@ -20,6 +20,13 @@ pub(crate) fn path(call: &ToolCall) -> Result<String> {
     string(call, field)
 }
 
+pub(crate) fn execution_path(call: &ToolCall) -> Result<Option<String>> {
+    match normalized(&call.name).as_str() {
+        "write" | "strreplace" | "editnotebook" => path(call).map(Some),
+        _ => Ok(None),
+    }
+}
+
 pub(crate) fn after_read(
     call: &ToolCall,
     result: &pb::ReadResult,

server/src/cursor/tools/mod.rs

@@ -1,11 +1,19 @@
-use std::collections::{BTreeMap, HashSet};
+use std::{
+    collections::{BTreeMap, HashSet},
+    sync::Arc,
+};
+
+use tokio::sync::Mutex;
 
 pub mod codec;
 mod dispatch;
 pub(crate) mod edit;
 pub(crate) mod result;
 pub mod runtime;
+mod schedule;
 pub(crate) mod stream;
+#[cfg(test)]
+mod tests;
 
 use crate::{
     model::{CanonicalMessage, MessageContent, Role, ToolCall},
@@ -14,6 +22,7 @@ use crate::{
 };
 
 use self::result::{ToolCompletion, ToolResultSender};
+use self::schedule::{DeferredEdit, EditSchedule};
 use super::{interaction, proto::agent::v1 as pb};
 use runtime::{CursorToolRuntime, ExecContext};
 
@@ -23,6 +32,7 @@ pub struct ToolDispatcher {
     results: ToolResultSender,
     search: WebSearch,
     fetch: WebFetch,
+    edit_schedule: Arc<Mutex<EditSchedule>>,
 }
 
 pub struct DispatchedTool {
@@ -54,6 +64,7 @@ impl ToolDispatcher {
             results,
             search: WebSearch::built_in(),
             fetch: WebFetch::built_in(),
+            edit_schedule: Arc::new(Mutex::new(EditSchedule::default())),
         }
     }
 
@@ -74,20 +85,62 @@ impl ToolDispatcher {
             if state.completed.contains(&call.call_id) {
                 continue;
             }
+            let message_index = first_tool_index + position;
+            let publish_started = !state.started.contains(&call.call_id);
+            let edit_path = if dynamic_mcp.contains_key(&call.name) {
+                None
+            } else {
+                edit::execution_path(call)?
+            };
+            if let Some(path) = edit_path {
+                let next = self.edit_schedule.lock().await.start_or_defer(
+                    path,
+                    DeferredEdit {
+                        call: call.clone(),
+                        message_index,
+                        publish_started,
+                        context: context.clone(),
+                    },
+                );
+                let Some(next) = next else {
+                    continue;
+                };
+                dispatched.push(
+                    self.start(
+                        &next.call,
+                        next.message_index,
+                        next.publish_started,
+                        dynamic_mcp,
+                        &next.context,
+                    )
+                    .await?,
+                );
+                continue;
+            }
             dispatched.push(
-                self.start(
-                    call,
-                    first_tool_index + position,
-                    !state.started.contains(&call.call_id),
-                    dynamic_mcp,
-                    context,
-                )
-                .await?,
+                self.start(call, message_index, publish_started, dynamic_mcp, context)
+                    .await?,
             );
         }
         Ok(dispatched)
     }
 
+    pub(crate) async fn continue_after(&self, call_id: &str) -> Result<Option<DispatchedTool>> {
+        let next = self.edit_schedule.lock().await.complete(call_id)?;
+        let Some(next) = next else {
+            return Ok(None);
+        };
+        self.start(
+            &next.call,
+            next.message_index,
+            next.publish_started,
+            &BTreeMap::new(),
+            &next.context,
+        )
+        .await
+        .map(Some)
+    }
+
     async fn start(
         &self,
         call: &ToolCall,

server/src/cursor/tools/result/exec/mod.rs

@@ -7,7 +7,7 @@ use crate::{
     Error, Result,
 };
 
-use super::{mcp_state, ReadImage, ToolCompletion};
+use super::{gate, mcp_state, ReadImage, ToolCompletion};
 use crate::cursor::tools::{
     edit,
     runtime::{ExecStage, PendingExec},
@@ -18,6 +18,15 @@ pub(crate) fn from_exec(
     wire_result: &pb::exec_client_message::Message,
 ) -> Result<ToolCompletion> {
     use pb::{exec_client_message::Message, tool_call::Tool};
+    let mut gated_shell = matches!(
+        wire_result,
+        Message::ShellResult(_) | Message::MiniSweAgentBashResult(_)
+    )
+    .then(|| wire_result.clone());
+    if let Some(message) = gated_shell.as_mut() {
+        gate::exec_message(message);
+    }
+    let wire_result = gated_shell.as_ref().unwrap_or(wire_result);
     if let Message::McpStateExecResult(result) = wire_result {
         return mcp_state::complete(pending, result);
     }

server/src/cursor/tools/result/exec/output.rs

@@ -119,17 +119,98 @@ fn delete(value: &pb::DeleteResult) -> Result<(String, bool)> {
 fn grep(value: &pb::GrepResult) -> Result<(String, bool)> {
     use pb::grep_result::Result as R;
     match value.result.as_ref().ok_or_else(|| missing("grep"))? {
-        R::Success(value) => Ok((
-            format!(
-                "grep success pattern={} mode={}",
-                value.pattern, value.output_mode
-            ),
-            false,
-        )),
+        R::Success(value) => Ok((grep_success(value), false)),
         R::Error(value) => Ok((value.error.clone(), true)),
     }
 }
 
+fn grep_success(value: &pb::GrepSuccess) -> String {
+    let mut lines = Vec::new();
+    if let Some(result) = &value.active_editor_result {
+        grep_union(result, &mut lines);
+    }
+    let mut workspaces = value.workspace_results.iter().collect::<Vec<_>>();
+    workspaces.sort_unstable_by_key(|(name, _)| *name);
+    for (_, result) in workspaces {
+        grep_union(result, &mut lines);
+    }
+    if lines.is_empty() {
+        format!(
+            "No matches found for pattern `{}` in {}",
+            value.pattern, value.path
+        )
+    } else {
+        lines.join("\n")
+    }
+}
+
+fn grep_union(value: &pb::GrepUnionResult, lines: &mut Vec<String>) {
+    use pb::grep_union_result::Result as R;
+    match value.result.as_ref() {
+        Some(R::Files(value)) => {
+            lines.extend(value.files.iter().cloned());
+            grep_truncation(
+                value.client_truncated,
+                value.ripgrep_truncated,
+                value.total_files,
+                "files",
+                lines,
+            );
+        }
+        Some(R::Count(value)) => {
+            lines.extend(
+                value
+                    .counts
+                    .iter()
+                    .map(|count| format!("{}:{}", count.file, count.count)),
+            );
+            grep_truncation(
+                value.client_truncated,
+                value.ripgrep_truncated,
+                value.total_matches,
+                "matches",
+                lines,
+            );
+        }
+        Some(R::Content(value)) => {
+            for file in &value.matches {
+                lines.extend(file.matches.iter().map(|matched| {
+                    let separator = if matched.is_context_line { '-' } else { ':' };
+                    let truncated = if matched.content_truncated {
+                        " [line truncated]"
+                    } else {
+                        ""
+                    };
+                    format!(
+                        "{}{separator}{}{separator}{}{truncated}",
+                        file.file, matched.line_number, matched.content
+                    )
+                }));
+            }
+            grep_truncation(
+                value.client_truncated,
+                value.ripgrep_truncated,
+                value.total_matched_lines,
+                "matched lines",
+                lines,
+            );
+        }
+        None => {}
+    }
+}
+
+fn grep_truncation(
+    client_truncated: bool,
+    ripgrep_truncated: bool,
+    total: i32,
+    unit: &str,
+    lines: &mut Vec<String>,
+) {
+    if client_truncated || ripgrep_truncated {
+        lines.push(format!("[Results truncated; {total} total {unit}]"));
+    }
+}
+
 fn diagnostics(value: &pb::DiagnosticsResult) -> Result<(String, bool)> {
     use pb::diagnostics_result::Result as R;
     match value
@@ -137,13 +218,7 @@ fn diagnostics(value: &pb::DiagnosticsResult) -> Result<(String, bool)> {
         .as_ref()
         .ok_or_else(|| missing("diagnostics"))?
     {
-        R::Success(value) => Ok((
-            format!(
-                "diagnostics path={} count={}",
-                value.path, value.total_diagnostics
-            ),
-            false,
-        )),
+        R::Success(value) => Ok((diagnostics_success(value), false)),
         R::Error(value) => Ok((value.error.clone(), true)),
         R::Rejected(value) => Ok((value.reason.clone(), true)),
         R::FileNotFound(value) => Ok((format!("file not found: {}", value.path), true)),
@@ -151,6 +226,76 @@ fn diagnostics(value: &pb::DiagnosticsResult) -> Result<(String, bool)> {
     }
 }
 
+fn diagnostics_success(value: &pb::DiagnosticsSuccess) -> String {
+    if value.diagnostics.is_empty() {
+        return format!("No diagnostics found in {}", value.path);
+    }
+    let mut lines = value
+        .diagnostics
+        .iter()
+        .map(|diagnostic| {
+            let location = diagnostic_location(&value.path, diagnostic.range.as_ref());
+            let mut labels = vec![diagnostic_severity(diagnostic.severity)];
+            if !diagnostic.source.is_empty() {
+                labels.push(diagnostic.source.as_str());
+            }
+            if !diagnostic.code.is_empty() {
+                labels.push(diagnostic.code.as_str());
+            }
+            if diagnostic.is_stale {
+                labels.push("stale");
+            }
+            format!(
+                "{}: [{}] {}",
+                location,
+                labels.join(" "),
+                diagnostic.message
+            )
+        })
+        .collect::<Vec<_>>();
+    if value.total_diagnostics != value.diagnostics.len() as i32 {
+        lines.push(format!(
+            "[Reported {} diagnostics; received {} details]",
+            value.total_diagnostics,
+            value.diagnostics.len()
+        ));
+    }
+    lines.join("\n")
+}
+
+fn diagnostic_location(path: &str, range: Option<&pb::Range>) -> String {
+    let Some(range) = range else {
+        return path.into();
+    };
+    let Some(start) = &range.start else {
+        return path.into();
+    };
+    let mut location = format!(
+        "{}:{}:{}",
+        path,
+        start.line.saturating_add(1),
+        start.column.saturating_add(1)
+    );
+    if let Some(end) = &range.end {
+        location.push_str(&format!(
+            "-{}:{}",
+            end.line.saturating_add(1),
+            end.column.saturating_add(1)
+        ));
+    }
+    location
+}
+
+fn diagnostic_severity(value: i32) -> &'static str {
+    match pb::DiagnosticSeverity::try_from(value) {
+        Ok(pb::DiagnosticSeverity::Error) => "error",
+        Ok(pb::DiagnosticSeverity::Warning) => "warning",
+        Ok(pb::DiagnosticSeverity::Information) => "information",
+        Ok(pb::DiagnosticSeverity::Hint) => "hint",
+        Ok(pb::DiagnosticSeverity::Unspecified) | Err(_) => "diagnostic",
+    }
+}
+
 fn mcp(value: &pb::McpResult) -> Result<(String, bool)> {
     use pb::mcp_result::Result as R;
     match value.result.as_ref().ok_or_else(|| missing("mcp"))? {
@@ -264,6 +409,106 @@ fn creates_subagent(call: &ToolCall) -> bool {
     )
 }
 
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use super::*;
+
+    #[test]
+    fn grep_output_contains_file_and_match_details() {
+        let value = pb::GrepResult {
+            result: Some(pb::grep_result::Result::Success(pb::GrepSuccess {
+                pattern: "Cursor".into(),
+                path: "/workspace".into(),
+                output_mode: "content".into(),
+                workspace_results: HashMap::from([
+                    (
+                        "workspace-b".into(),
+                        pb::GrepUnionResult {
+                            result: Some(pb::grep_union_result::Result::Files(
+                                pb::GrepFilesResult {
+                                    files: vec!["/workspace/Cargo.toml".into()],
+                                    total_files: 1,
+                                    ..Default::default()
+                                },
+                            )),
+                        },
+                    ),
+                    (
+                        "workspace-a".into(),
+                        pb::GrepUnionResult {
+                            result: Some(pb::grep_union_result::Result::Content(
+                                pb::GrepContentResult {
+                                    matches: vec![pb::GrepFileMatch {
+                                        file: "/workspace/README.md".into(),
+                                        matches: vec![pb::GrepContentMatch {
+                                            line_number: 7,
+                                            content: "Cursor BYOK".into(),
+                                            ..Default::default()
+                                        }],
+                                    }],
+                                    total_lines: 1,
+                                    total_matched_lines: 1,
+                                    ..Default::default()
+                                },
+                            )),
+                        },
+                    ),
+                ]),
+                active_editor_result: None,
+            })),
+        };
+
+        let (content, is_error) = grep(&value).unwrap();
+
+        assert!(!is_error);
+        assert!(content.contains("/workspace/README.md:7:Cursor BYOK"));
+        assert!(content.contains("/workspace/Cargo.toml"));
+        assert!(
+            content.find("/workspace/README.md").unwrap()
+                < content.find("/workspace/Cargo.toml").unwrap(),
+            "workspace map output must be deterministic"
+        );
+    }
+
+    #[test]
+    fn diagnostics_output_contains_each_diagnostic_detail() {
+        let value = pb::DiagnosticsResult {
+            result: Some(pb::diagnostics_result::Result::Success(
+                pb::DiagnosticsSuccess {
+                    path: "/workspace/src/main.rs".into(),
+                    diagnostics: vec![pb::Diagnostic {
+                        severity: pb::DiagnosticSeverity::Error as i32,
+                        range: Some(pb::Range {
+                            start: Some(pb::Position { line: 4, column: 8 }),
+                            end: Some(pb::Position {
+                                line: 4,
+                                column: 12,
+                            }),
+                        }),
+                        message: "cannot find value `name`".into(),
+                        source: "rustc".into(),
+                        code: "E0425".into(),
+                        is_stale: false,
+                    }],
+                    total_diagnostics: 1,
+                },
+            )),
+        };
+
+        let (content, is_error) = diagnostics(&value).unwrap();
+
+        assert!(!is_error);
+        assert!(content.contains("/workspace/src/main.rs:5:9"));
+        assert!(content.contains("-5:13"));
+        assert!(content.contains("error"));
+        assert!(content.contains("rustc"));
+        assert!(content.contains("E0425"));
+        assert!(content.contains("cannot find value `name`"));
+    }
+}
+
 fn missing(name: &str) -> Error {
     Error::Protocol(format!("{name} returned no result"))
 }

server/src/cursor/tools/result/gate.rs

@@ -0,0 +1,169 @@
+use crate::cursor::proto::agent::v1 as pb;
+
+const KIB: usize = 1024;
+const SHELL_STREAM_LIMIT: usize = 16 * KIB;
+const SHELL_CONTENT_LIMIT: usize = 32 * KIB;
+
+pub(super) fn model_content(tool: &pb::tool_call::Tool, content: &mut String) {
+    if matches!(tool, pb::tool_call::Tool::ShellToolCall(_)) {
+        *content = truncate_edges("Shell", content, SHELL_CONTENT_LIMIT);
+    }
+}
+
+pub(super) fn exec_message(message: &mut pb::exec_client_message::Message) {
+    use pb::exec_client_message::Message;
+    match message {
+        Message::ShellResult(result) | Message::MiniSweAgentBashResult(result) => {
+            gate_shell_result(result)
+        }
+        _ => {}
+    }
+}
+
+fn gate_shell_result(result: &mut pb::ShellResult) {
+    use pb::shell_result::Result;
+    match result.result.as_mut() {
+        Some(Result::Success(success)) => {
+            success.stdout = truncate_edges("Shell stdout", &success.stdout, SHELL_STREAM_LIMIT);
+            success.stderr = truncate_edges("Shell stderr", &success.stderr, SHELL_STREAM_LIMIT);
+            if let Some(interleaved) = success.interleaved_output.as_mut() {
+                *interleaved =
+                    truncate_edges("Shell interleaved output", interleaved, SHELL_CONTENT_LIMIT);
+            }
+        }
+        Some(Result::Failure(failure)) => {
+            failure.stdout = truncate_edges("Shell stdout", &failure.stdout, SHELL_STREAM_LIMIT);
+            failure.stderr = truncate_edges("Shell stderr", &failure.stderr, SHELL_STREAM_LIMIT);
+            if let Some(interleaved) = failure.interleaved_output.as_mut() {
+                *interleaved =
+                    truncate_edges("Shell interleaved output", interleaved, SHELL_CONTENT_LIMIT);
+            }
+        }
+        _ => {}
+    }
+}
+
+fn truncate_edges(tool_name: &str, content: &str, limit: usize) -> String {
+    if content.len() <= limit {
+        return content.to_string();
+    }
+    let original = content.len();
+    let mut shown = limit;
+    loop {
+        let notice = format!(
+            "\n\n[truncated: {tool_name} result exceeded {limit} bytes; omitted middle; showing {shown} of {original} bytes]\n\n"
+        );
+        let available = limit.saturating_sub(notice.len());
+        let head = utf8_prefix(content, available / 2);
+        let tail = utf8_suffix(content, available.saturating_sub(head.len()));
+        let next_shown = head.len().saturating_add(tail.len());
+        if next_shown == shown {
+            return format!("{head}{notice}{tail}");
+        }
+        shown = next_shown;
+    }
+}
+
+fn utf8_prefix(value: &str, limit: usize) -> &str {
+    let mut end = limit.min(value.len());
+    while end > 0 && !value.is_char_boundary(end) {
+        end -= 1;
+    }
+    &value[..end]
+}
+
+fn utf8_suffix(value: &str, limit: usize) -> &str {
+    let mut start = value.len().saturating_sub(limit);
+    while start < value.len() && !value.is_char_boundary(start) {
+        start += 1;
+    }
+    &value[start..]
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn shell_tool() -> pb::tool_call::Tool {
+        pb::tool_call::Tool::ShellToolCall(pb::ShellToolCall::default())
+    }
+
+    #[test]
+    fn shell_output_keeps_both_ends_within_its_budget() {
+        let mut content = format!("HEAD{}TAIL", " ".repeat(1024 * KIB));
+
+        model_content(&shell_tool(), &mut content);
+
+        assert!(content.len() <= SHELL_CONTENT_LIMIT);
+        assert!(content.starts_with("HEAD"));
+        assert!(content.ends_with("TAIL"));
+        assert!(content.contains("omitted middle"));
+    }
+
+    #[test]
+    fn non_shell_output_is_unchanged() {
+        let mut content = "x".repeat(64 * KIB);
+        let original = content.clone();
+
+        model_content(
+            &pb::tool_call::Tool::ReadToolCall(pb::ReadToolCall::default()),
+            &mut content,
+        );
+
+        assert_eq!(content, original);
+    }
+
+    #[test]
+    fn shell_streams_are_limited_before_rendering() {
+        let mut message = pb::exec_client_message::Message::ShellResult(pb::ShellResult {
+            result: Some(pb::shell_result::Result::Success(pb::ShellSuccess {
+                stdout: format!("HEAD{}TAIL", "x".repeat(64 * KIB)),
+                stderr: format!("ERROR_HEAD{}ERROR_TAIL", "y".repeat(64 * KIB)),
+                interleaved_output: Some(format!("START{}END", "z".repeat(64 * KIB))),
+                ..Default::default()
+            })),
+            ..Default::default()
+        });
+
+        exec_message(&mut message);
+
+        let pb::exec_client_message::Message::ShellResult(result) = message else {
+            panic!("expected Shell result");
+        };
+        let Some(pb::shell_result::Result::Success(success)) = result.result else {
+            panic!("expected Shell success");
+        };
+        assert!(success.stdout.len() <= SHELL_STREAM_LIMIT);
+        assert!(success.stdout.starts_with("HEAD"));
+        assert!(success.stdout.ends_with("TAIL"));
+        assert!(success.stderr.len() <= SHELL_STREAM_LIMIT);
+        assert!(success.stderr.starts_with("ERROR_HEAD"));
+        assert!(success.stderr.ends_with("ERROR_TAIL"));
+        assert!(success.interleaved_output.unwrap().len() <= SHELL_CONTENT_LIMIT);
+    }
+
+    #[test]
+    fn failed_shell_streams_are_limited() {
+        let mut message = pb::exec_client_message::Message::ShellResult(pb::ShellResult {
+            result: Some(pb::shell_result::Result::Failure(pb::ShellFailure {
+                stdout: "x".repeat(64 * KIB),
+                stderr: "y".repeat(64 * KIB),
+                interleaved_output: Some("z".repeat(64 * KIB)),
+                ..Default::default()
+            })),
+            ..Default::default()
+        });
+
+        exec_message(&mut message);
+
+        let pb::exec_client_message::Message::ShellResult(result) = message else {
+            panic!("expected Shell result");
+        };
+        let Some(pb::shell_result::Result::Failure(failure)) = result.result else {
+            panic!("expected Shell failure");
+        };
+        assert!(failure.stdout.len() <= SHELL_STREAM_LIMIT);
+        assert!(failure.stderr.len() <= SHELL_STREAM_LIMIT);
+        assert!(failure.interleaved_output.unwrap().len() <= SHELL_CONTENT_LIMIT);
+    }
+}

server/src/cursor/tools/result/local.rs

@@ -169,6 +169,13 @@ fn text(value: &Value, name: &str) -> String {
         .into()
 }
 
+fn normalized(name: &str) -> String {
+    name.chars()
+        .filter(|character| character.is_ascii_alphanumeric())
+        .flat_map(char::to_lowercase)
+        .collect()
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -203,10 +210,3 @@ mod tests {
         ));
     }
 }
-
-fn normalized(name: &str) -> String {
-    name.chars()
-        .filter(|character| character.is_ascii_alphanumeric())
-        .flat_map(char::to_lowercase)
-        .collect()
-}

server/src/cursor/tools/result/mod.rs

@@ -1,5 +1,6 @@
 mod await_shell;
 mod exec;
+mod gate;
 mod interaction;
 mod local;
 mod mcp;
@@ -87,9 +88,10 @@ impl ToolCompletion {
     pub(crate) fn new(
         call: &ToolCall,
         started_at_ms: u64,
-        result: ToolResult,
+        mut result: ToolResult,
         tool: pb::tool_call::Tool,
     ) -> Self {
+        gate::model_content(&tool, &mut result.content);
         Self {
             result,
             tool_call: pb::ToolCall {

server/src/cursor/tools/runtime.rs

@@ -346,6 +346,13 @@ impl CursorToolRuntime {
     }
 }
 
+pub(crate) fn now_ms() -> u64 {
+    std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .unwrap_or_default()
+        .as_millis() as u64
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -420,10 +427,3 @@ mod tests {
             .is_none());
     }
 }
-
-pub(crate) fn now_ms() -> u64 {
-    std::time::SystemTime::now()
-        .duration_since(std::time::UNIX_EPOCH)
-        .unwrap_or_default()
-        .as_millis() as u64
-}

server/src/cursor/tools/schedule.rs

@@ -0,0 +1,68 @@
+use std::collections::{HashMap, VecDeque};
+
+use crate::{model::ToolCall, Error, Result};
+
+use super::runtime::ExecContext;
+
+#[derive(Default)]
+pub(super) struct EditSchedule {
+    paths: HashMap<String, EditPathQueue>,
+    active_paths: HashMap<String, String>,
+}
+
+struct EditPathQueue {
+    active_call_id: String,
+    waiting: VecDeque<DeferredEdit>,
+}
+
+pub(super) struct DeferredEdit {
+    pub call: ToolCall,
+    pub message_index: usize,
+    pub publish_started: bool,
+    pub context: ExecContext,
+}
+
+impl EditSchedule {
+    pub fn start_or_defer(&mut self, path: String, edit: DeferredEdit) -> Option<DeferredEdit> {
+        if let Some(queue) = self.paths.get_mut(&path) {
+            queue.waiting.push_back(edit);
+            return None;
+        }
+        self.active_paths
+            .insert(edit.call.call_id.clone(), path.clone());
+        self.paths.insert(
+            path,
+            EditPathQueue {
+                active_call_id: edit.call.call_id.clone(),
+                waiting: VecDeque::new(),
+            },
+        );
+        Some(edit)
+    }
+
+    pub fn complete(&mut self, call_id: &str) -> Result<Option<DeferredEdit>> {
+        let Some(path) = self.active_paths.remove(call_id) else {
+            return Ok(None);
+        };
+        let queue = self.paths.get_mut(&path).ok_or_else(|| {
+            Error::Protocol(format!("active edit path disappeared for call {call_id}"))
+        })?;
+        if queue.active_call_id != call_id {
+            return Err(Error::Protocol(format!(
+                "edit path is active for {}, not {call_id}",
+                queue.active_call_id
+            )));
+        }
+        match queue.waiting.pop_front() {
+            Some(next) => {
+                queue.active_call_id = next.call.call_id.clone();
+                self.active_paths.insert(next.call.call_id.clone(), path);
+                Ok(Some(next))
+            }
+            None => {
+                self.paths.remove(&path);
+                Ok(None)
+            }
+        }
+    }
+}

server/src/cursor/tools/tests.rs

@@ -0,0 +1,139 @@
+use super::*;
+use serde_json::json;
+
+fn edit_call(index: usize, call_id: &str, path: &str, old: &str, new: &str) -> ToolCall {
+    ToolCall {
+        index,
+        call_id: call_id.into(),
+        model_call_id: "model:0".into(),
+        name: "StrReplace".into(),
+        arguments_text: String::new(),
+        arguments: json!({
+            "path": path,
+            "old_string": old,
+            "new_string": new,
+        }),
+    }
+}
+
+#[tokio::test]
+async fn same_path_edits_start_one_at_a_time() {
+    let runtime = CursorToolRuntime::default();
+    let dispatcher = ToolDispatcher::new(runtime.clone());
+    let calls = [
+        edit_call(0, "first", "/tmp/a.txt", "left", "LEFT"),
+        edit_call(1, "second", "/tmp/a.txt", "right", "RIGHT"),
+        edit_call(2, "other", "/tmp/b.txt", "other", "OTHER"),
+    ];
+
+    let dispatched = dispatcher
+        .start_batch(
+            &calls,
+            ToolBatchState {
+                completed: &HashSet::new(),
+                started: &HashSet::new(),
+                response_text: "",
+                response_thinking: "",
+            },
+            &[],
+            &BTreeMap::new(),
+            &ExecContext::default(),
+        )
+        .await
+        .unwrap();
+
+    assert_eq!(dispatched.len(), 2);
+    assert_eq!(exec(&dispatched[0]).exec_id, "first");
+    assert_eq!(exec(&dispatched[1]).exec_id, "other");
+
+    let mut file = "left right\n".to_string();
+    let first_write = advance_read(&runtime, exec(&dispatched[0]).id, &file).await;
+    file = write_text(&first_write);
+    assert_eq!(file, "LEFT right\n");
+    complete_write(&runtime, &first_write).await;
+
+    let second = dispatcher
+        .continue_after("first")
+        .await
+        .unwrap()
+        .expect("second same-path edit should start after the first completes");
+    assert_eq!(exec(&second).exec_id, "second");
+    let second_write = advance_read(&runtime, exec(&second).id, &file).await;
+    file = write_text(&second_write);
+    assert_eq!(file, "LEFT RIGHT\n");
+    complete_write(&runtime, &second_write).await;
+    assert!(dispatcher.continue_after("second").await.unwrap().is_none());
+}
+
+fn exec(dispatched: &DispatchedTool) -> &pb::ExecServerMessage {
+    dispatched
+        .messages
+        .iter()
+        .find_map(|message| match message.message.as_ref() {
+            Some(pb::agent_server_message::Message::ExecServerMessage(exec)) => Some(exec),
+            _ => None,
+        })
+        .expect("dispatched edit should contain an Exec request")
+}
+
+async fn advance_read(
+    runtime: &CursorToolRuntime,
+    id: u32,
+    content: &str,
+) -> pb::ExecServerMessage {
+    let event = codec::client_event(
+        &pb::ExecClientMessage {
+            id,
+            message: Some(pb::exec_client_message::Message::ReadResult(
+                pb::ReadResult {
+                    result: Some(pb::read_result::Result::Success(pb::ReadSuccess {
+                        output: Some(pb::read_success::Output::Content(content.into())),
+                        ..Default::default()
+                    })),
+                },
+            )),
+            ..Default::default()
+        },
+        runtime,
+    )
+    .await
+    .unwrap();
+    let codec::ClientExecEvent::Message(message) = event else {
+        panic!("edit read should advance to a write")
+    };
+    let Some(pb::agent_server_message::Message::ExecServerMessage(exec)) = message.message else {
+        panic!("edit read should emit an Exec write request")
+    };
+    exec
+}
+
+fn write_text(exec: &pb::ExecServerMessage) -> String {
+    let Some(pb::exec_server_message::Message::WriteArgs(args)) = exec.message.as_ref() else {
+        panic!("expected WriteArgs")
+    };
+    args.file_text.clone()
+}
+
+async fn complete_write(runtime: &CursorToolRuntime, exec: &pb::ExecServerMessage) {
+    let Some(pb::exec_server_message::Message::WriteArgs(args)) = exec.message.as_ref() else {
+        panic!("expected WriteArgs")
+    };
+    let event = codec::client_event(
+        &pb::ExecClientMessage {
+            id: exec.id,
+            message: Some(pb::exec_client_message::Message::WriteResult(
+                pb::WriteResult {
+                    result: Some(pb::write_result::Result::Success(pb::WriteSuccess {
+                        path: args.path.clone(),
+                        ..Default::default()
+                    })),
+                },
+            )),
+            ..Default::default()
+        },
+        runtime,
+    )
+    .await
+    .unwrap();
+    assert!(matches!(event, codec::ClientExecEvent::Completed(_)));
+}

server/src/harness/mod.rs

@@ -9,7 +9,10 @@ use parking_lot::RwLock;
 use serde::{Deserialize, Serialize};
 use tokio::sync::Mutex;
 
-use crate::{store::Store, Error, Result};
+use crate::{
+    store::{Store, TabMode, TabSettings},
+    Error, Result,
+};
 
 use self::{ca::CaManager, proxy::ProxyRuntime};
 
@@ -65,6 +68,7 @@ struct Inner {
     ca: CaManager,
     ca_initialization: Mutex<()>,
     backend_addr: RwLock<Option<SocketAddr>>,
+    tab_mode: Arc<RwLock<TabMode>>,
     proxy: Mutex<ProxyRuntime>,
 }
 
@@ -76,6 +80,7 @@ impl CursorHarness {
                 ca: CaManager::managed()?,
                 ca_initialization: Mutex::new(()),
                 backend_addr: RwLock::new(None),
+                tab_mode: Arc::new(RwLock::new(TabMode::default())),
                 proxy: Mutex::new(ProxyRuntime::default()),
             }),
         })
@@ -138,6 +143,12 @@ impl CursorHarness {
         self.status().await
     }
 
+    pub async fn set_tab_settings(&self, settings: TabSettings) -> Result<TabSettings> {
+        let saved = self.inner.store.set_tab_settings(settings).await?;
+        *self.inner.tab_mode.write() = saved.mode;
+        Ok(saved)
+    }
+
     async fn enable(&self) -> Result<()> {
         if !matches!(self.inner.ca.state()?, CaState::Ready) {
             return Err(Error::Config(
@@ -158,7 +169,15 @@ impl CursorHarness {
         }
         let ca = self.inner.ca.load()?;
         let requested_port = self.inner.store.port_settings().await?.proxy_port;
-        let (url, actual_port) = proxy.start(backend_addr, ca, requested_port).await?;
+        *self.inner.tab_mode.write() = self.inner.store.tab_settings().await?.mode;
+        let (url, actual_port) = proxy
+            .start(
+                backend_addr,
+                ca,
+                requested_port,
+                self.inner.tab_mode.clone(),
+            )
+            .await?;
         if let Err(error) = self.inner.store.set_proxy_port(actual_port).await {
             proxy.stop().await;
             return Err(error);

server/src/harness/proxy.rs

@@ -1,4 +1,4 @@
-use std::net::SocketAddr;
+use std::{net::SocketAddr, sync::Arc};
 
 use hudsucker::{
     certificate_authority::RcgenAuthority,
@@ -8,7 +8,13 @@ use hudsucker::{
 };
 use tokio::{net::TcpListener, sync::oneshot, task::JoinHandle};
 
-use crate::{cursor::proxy::UPSTREAM_URL_HEADER, Error, Result};
+use parking_lot::RwLock;
+
+use crate::{
+    cursor::{proxy::UPSTREAM_URL_HEADER, tab::is_tab_path},
+    store::TabMode,
+    Error, Result,
+};
 
 use super::ca::LoadedCa;
 
@@ -33,6 +39,7 @@ impl ProxyRuntime {
         backend: SocketAddr,
         ca: LoadedCa,
         requested_port: u16,
+        tab_mode: Arc<RwLock<TabMode>>,
     ) -> Result<(String, u16)> {
         if let Some(url) = self.url() {
             return Ok((url, self.port.unwrap_or_default()));
@@ -45,7 +52,7 @@ impl ProxyRuntime {
             .with_listener(listener)
             .with_ca(authority)
             .with_rustls_connector(aws_lc_rs::default_provider())
-            .with_http_handler(CursorRelay { backend })
+            .with_http_handler(CursorRelay { backend, tab_mode })
             .with_graceful_shutdown(async move {
                 let _ = done.await;
             })
@@ -89,6 +96,7 @@ async fn bind_proxy_listener(requested_port: u16) -> Result<TcpListener> {
 #[derive(Clone)]
 struct CursorRelay {
     backend: SocketAddr,
+    tab_mode: Arc<RwLock<TabMode>>,
 }
 
 impl HttpHandler for CursorRelay {
@@ -98,7 +106,8 @@ impl HttpHandler for CursorRelay {
         mut request: Request<Body>,
     ) -> RequestOrResponse {
         let original = request.uri().clone();
-        if is_cursor_host(original.host().unwrap_or_default()) && is_local_path(original.path()) {
+        let locally_routed = should_route_locally(original.path(), *self.tab_mode.read());
+        if is_cursor_host(original.host().unwrap_or_default()) && locally_routed {
             if let Ok(value) = original.to_string().parse() {
                 request.headers_mut().insert(UPSTREAM_URL_HEADER, value);
             }
@@ -157,6 +166,10 @@ fn is_local_path(path: &str) -> bool {
     )
 }
 
+fn should_route_locally(path: &str, tab_mode: TabMode) -> bool {
+    is_local_path(path) || (is_tab_path(path) && tab_mode != TabMode::Direct)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -179,5 +192,17 @@ mod tests {
             "/aiserver.v1.AnalyticsService/BootstrapStatsig"
         ));
         assert!(!is_local_path("/unrelated"));
+        assert!(should_route_locally(
+            "/aiserver.v1.AiService/StreamCpp",
+            TabMode::Public
+        ));
+        assert!(should_route_locally(
+            "/aiserver.v1.AiService/StreamCpp",
+            TabMode::Custom
+        ));
+        assert!(!should_route_locally(
+            "/aiserver.v1.AiService/StreamCpp",
+            TabMode::Direct
+        ));
     }
 }

server/src/model/mod.rs

@@ -9,6 +9,7 @@ mod projection;
 mod provider;
 mod run;
 mod runtime_tag;
+mod token_count;
 mod tool;
 mod usage;
 
@@ -23,5 +24,6 @@ pub use projection::*;
 pub use provider::*;
 pub use run::*;
 pub use runtime_tag::*;
+pub(crate) use token_count::*;
 pub use tool::*;
 pub use usage::*;

server/src/model/provider.rs

@@ -51,6 +51,7 @@ pub struct ProviderEndpoint {
     pub name: String,
     pub provider_type: ProviderType,
     pub base_url: String,
+    pub api_key: Option<String>,
     pub has_api_key: bool,
     pub custom_headers: serde_json::Value,
     pub extra_params: serde_json::Value,
@@ -61,7 +62,6 @@ pub struct ProviderEndpoint {
 #[derive(Clone, Debug)]
 pub struct ProviderEndpointSecret {
     pub endpoint: ProviderEndpoint,
-    pub api_key: String,
     pub custom_headers: serde_json::Value,
 }
 
@@ -138,7 +138,12 @@ pub fn normalize_base_url(value: &str) -> Result<String> {
     Ok(url.as_str().trim_end_matches('/').to_string())
 }
 
-pub fn model_hash(base_url: &str, provider_type: ProviderType, model_id: &str) -> Result<String> {
+pub fn model_hash(
+    base_url: &str,
+    api_key: &str,
+    provider_type: ProviderType,
+    model_id: &str,
+) -> Result<String> {
     let base_url = normalize_base_url(base_url)?;
     let model_id = model_id.trim();
     if model_id.is_empty() {
@@ -147,6 +152,8 @@ pub fn model_hash(base_url: &str, provider_type: ProviderType, model_id: &str) -
     let mut digest = Sha256::new();
     digest.update(base_url.as_bytes());
     digest.update([0]);
+    digest.update(api_key.as_bytes());
+    digest.update([0]);
     digest.update(provider_type.as_str().as_bytes());
     digest.update([0]);
     digest.update(model_id.as_bytes());
@@ -212,24 +219,41 @@ mod tests {
     use super::*;
 
     #[test]
-    fn hash_uses_normalized_url_type_and_model_only() {
+    fn hash_uses_normalized_url_key_type_and_model() {
         let first = model_hash(
             "HTTPS://Example.COM/v1/",
+            "secret",
             ProviderType::OpenAiChat,
             "model-a",
         )
         .unwrap();
         let second = model_hash(
             "https://example.com/v1",
+            "secret",
             ProviderType::OpenAiChat,
             "model-a",
         )
         .unwrap();
         assert_eq!(first, second);
-        assert_eq!(first, "f246010a");
         assert_ne!(
             first,
-            model_hash("https://example.com/v1", ProviderType::Anthropic, "model-a").unwrap()
+            model_hash(
+                "https://example.com/v1",
+                "different-secret",
+                ProviderType::OpenAiChat,
+                "model-a",
+            )
+            .unwrap()
+        );
+        assert_ne!(
+            first,
+            model_hash(
+                "https://example.com/v1",
+                "secret",
+                ProviderType::Anthropic,
+                "model-a",
+            )
+            .unwrap()
         );
     }
 

server/src/model/token_count.rs

@@ -0,0 +1,39 @@
+pub(crate) fn parse_token_count(value: &str) -> Option<u64> {
+    let value = value.trim().to_ascii_lowercase();
+    let (number, multiplier) = match value.chars().last()? {
+        'k' => (&value[..value.len() - 1], 1_000),
+        'm' => (&value[..value.len() - 1], 1_000_000),
+        _ => (value.as_str(), 1),
+    };
+    number.parse::<u64>().ok()?.checked_mul(multiplier)
+}
+
+pub(crate) fn format_token_count(tokens: u64) -> String {
+    if tokens >= 1_000_000 && tokens.is_multiple_of(1_000_000) {
+        format!("{}M", tokens / 1_000_000)
+    } else if tokens >= 1_000 && tokens.is_multiple_of(1_000) {
+        format!("{}K", tokens / 1_000)
+    } else {
+        tokens.to_string()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn token_counts_parse_plain_and_abbreviated_values() {
+        assert_eq!(parse_token_count("272000"), Some(272_000));
+        assert_eq!(parse_token_count("272K"), Some(272_000));
+        assert_eq!(parse_token_count("1m"), Some(1_000_000));
+        assert_eq!(parse_token_count("invalid"), None);
+    }
+
+    #[test]
+    fn token_counts_format_exact_thousands_and_millions() {
+        assert_eq!(format_token_count(272_000), "272K");
+        assert_eq!(format_token_count(1_000_000), "1M");
+        assert_eq!(format_token_count(272_001), "272001");
+    }
+}

server/src/provider/openai_responses.rs

@@ -27,6 +27,12 @@ struct ResponseToolState {
     ended: bool,
 }
 
+enum ResponseToolArguments<'a> {
+    None,
+    Delta(&'a str),
+    Snapshot(&'a str),
+}
+
 pub struct OpenAiResponsesProvider {
     client: reqwest::Client,
     config: ProviderConfig,
@@ -113,7 +119,6 @@ impl Provider for OpenAiResponsesProvider {
             let mut reasoning_items = Vec::new();
             let mut saw_tool = false;
             let mut saw_completed_item = false;
-            let mut saw_done_marker = false;
             let mut terminal = false;
             loop {
                 let event = tokio::select! {
@@ -122,7 +127,7 @@ impl Provider for OpenAiResponsesProvider {
                 };
                 let Some(event) = event else { break };
                 let event = event.map_err(|error| Error::Provider(format!("OpenAI Responses SSE: {error}")))?;
-                if event.data == "[DONE]" { saw_done_marker = true; break; }
+                if event.data == "[DONE]" { break; }
                 let value: Value = serde_json::from_str(&event.data)?;
                 let kind = value.get("type").and_then(Value::as_str).unwrap_or(&event.event);
                 match kind {
@@ -140,11 +145,11 @@ impl Provider for OpenAiResponsesProvider {
                         }
                         if text_open { text_open = false; yield ModelEvent::TextEnd; }
                     }
-                    "response.reasoning_summary_text.delta" => {
+                    "response.reasoning_summary_text.delta" | "response.reasoning_text.delta" => {
                         if !thinking_open { thinking_open = true; yield ModelEvent::ThinkingStart; }
                         if let Some(delta) = value.get("delta").and_then(Value::as_str) { yield ModelEvent::ThinkingDelta(delta.into()); }
                     }
-                    "response.reasoning_summary_text.done" => {
+                    "response.reasoning_summary_text.done" | "response.reasoning_text.done" => {
                         if thinking_open { thinking_open = false; yield ModelEvent::ThinkingEnd; }
                     }
                     "response.output_item.added" => {
@@ -152,28 +157,32 @@ impl Provider for OpenAiResponsesProvider {
                         if item.get("type").and_then(Value::as_str) == Some("function_call") {
                             let index = required_u64(&value, "output_index")? as usize;
                             saw_tool = true;
-                            for event in update_response_tool(index, item, None, false, &mut tools)? { yield event; }
+                            for event in update_response_tool(index, item, ResponseToolArguments::None, false, &mut tools)? { yield event; }
                         }
                     }
                     "response.output_item.done" => {
                         let item = value.get("item").unwrap_or(&Value::Null);
-                        saw_completed_item = true;
                         match item.get("type").and_then(Value::as_str) {
                             Some("reasoning") => {
                                 if thinking_open { thinking_open = false; yield ModelEvent::ThinkingEnd; }
                                 reasoning_items.push(item.clone());
                             }
                             Some("message") => {
+                                saw_completed_item = true;
                                 if let Some(final_text) = response_item_text(item) {
                                     for event in reconcile_response_text(&mut text_open, &mut text, &final_text) { yield event; }
                                 }
                                 if text_open { text_open = false; yield ModelEvent::TextEnd; }
                             }
                             Some("function_call") => {
+                                saw_completed_item = true;
                                 let index = required_u64(&value, "output_index")? as usize;
                                 saw_tool = true;
-                                let final_arguments = item.get("arguments").and_then(Value::as_str);
-                                for event in update_response_tool(index, item, final_arguments, true, &mut tools)? { yield event; }
+                                let arguments = item
+                                    .get("arguments")
+                                    .and_then(Value::as_str)
+                                    .map_or(ResponseToolArguments::None, ResponseToolArguments::Snapshot);
+                                for event in update_response_tool(index, item, arguments, true, &mut tools)? { yield event; }
                             }
                             _ => {}
                         }
@@ -182,39 +191,32 @@ impl Provider for OpenAiResponsesProvider {
                         let index = required_u64(&value, "output_index")? as usize;
                         if let Some(delta) = value.get("delta").and_then(Value::as_str) {
                             saw_tool = true;
-                            for event in update_response_tool(index, &Value::Null, Some(delta), false, &mut tools)? { yield event; }
+                            for event in update_response_tool(index, &Value::Null, ResponseToolArguments::Delta(delta), false, &mut tools)? { yield event; }
                         }
                     }
                     "response.function_call_arguments.done" => {
                         let index = required_u64(&value, "output_index")? as usize;
-                        let final_arguments = value.get("arguments").and_then(Value::as_str);
-                        for event in update_response_tool(index, &Value::Null, final_arguments, true, &mut tools)? { yield event; }
+                        match value.get("arguments").and_then(Value::as_str) {
+                            Some("") => {
+                                for event in update_response_tool(
+                                    index,
+                                    &Value::Null,
+                                    ResponseToolArguments::None,
+                                    false,
+                                    &mut tools,
+                                )? { yield event; }
+                            }
+                            arguments => {
+                                let arguments = arguments.map_or(
+                                    ResponseToolArguments::None,
+                                    ResponseToolArguments::Snapshot,
+                                );
+                                for event in update_response_tool(index, &Value::Null, arguments, true, &mut tools)? { yield event; }
+                            }
+                        }
                     }
                     "response.completed" => {
                         if let Some(usage) = value.pointer("/response/usage") { yield ModelEvent::Usage(responses_usage(usage)); }
-                        if let Some(output) = value.pointer("/response/output").and_then(Value::as_array) {
-                            for (index, item) in output.iter().enumerate() {
-                                match item.get("type").and_then(Value::as_str) {
-                                    Some("reasoning") => {
-                                        if thinking_open { thinking_open = false; yield ModelEvent::ThinkingEnd; }
-                                        if !reasoning_items.iter().any(|existing| existing.get("id") == item.get("id")) {
-                                            reasoning_items.push(item.clone());
-                                        }
-                                    }
-                                    Some("message") => {
-                                        if let Some(final_text) = response_item_text(item) {
-                                            for event in reconcile_response_text(&mut text_open, &mut text, &final_text) { yield event; }
-                                        }
-                                    }
-                                    Some("function_call") => {
-                                        saw_tool = true;
-                                        let final_arguments = item.get("arguments").and_then(Value::as_str);
-                                        for event in update_response_tool(index, item, final_arguments, true, &mut tools)? { yield event; }
-                                    }
-                                    _ => {}
-                                }
-                            }
-                        }
                         if thinking_open { thinking_open = false; yield ModelEvent::ThinkingEnd; }
                         if text_open { text_open = false; yield ModelEvent::TextEnd; }
                         for (index, tool) in tools.iter_mut().filter(|(_, tool)| tool.started && !tool.ended) {
@@ -249,11 +251,11 @@ impl Provider for OpenAiResponsesProvider {
                     _ => {}
                 }
             }
-            if !terminal && saw_done_marker && saw_completed_item {
+            if !terminal && saw_completed_item {
                 if thinking_open { yield ModelEvent::ThinkingEnd; }
                 if text_open { yield ModelEvent::TextEnd; }
                 if tools.values().any(|tool| !tool.ended) {
-                    Err(Error::Provider("OpenAI Responses [DONE] arrived with an incomplete tool call".into()))?;
+                    Err(Error::Provider("OpenAI Responses stream ended with an incomplete tool call".into()))?;
                 }
                 terminal = true;
                 if !reasoning_items.is_empty() {
@@ -303,7 +305,7 @@ fn reconcile_response_text(
 fn update_response_tool(
     index: usize,
     item: &Value,
-    arguments: Option<&str>,
+    arguments: ResponseToolArguments<'_>,
     done: bool,
     tools: &mut std::collections::BTreeMap<usize, ResponseToolState>,
 ) -> Result<Vec<ModelEvent>> {
@@ -314,11 +316,17 @@ fn update_response_tool(
     if let Some(name) = item.get("name").and_then(Value::as_str) {
         tool.name.get_or_insert_with(|| name.into());
     }
-    if let Some(arguments) = arguments {
-        if arguments.starts_with(&tool.arguments) {
-            tool.arguments.push_str(&arguments[tool.arguments.len()..]);
-        } else if !tool.arguments.ends_with(arguments) {
-            tool.arguments.push_str(arguments);
+    match arguments {
+        ResponseToolArguments::None => {}
+        ResponseToolArguments::Delta(delta) => tool.arguments.push_str(delta),
+        ResponseToolArguments::Snapshot(snapshot) if snapshot == tool.arguments => {}
+        ResponseToolArguments::Snapshot(snapshot) if snapshot.starts_with(&tool.arguments) => {
+            tool.arguments.push_str(&snapshot[tool.arguments.len()..]);
+        }
+        ResponseToolArguments::Snapshot(_) => {
+            return Err(Error::Provider(
+                "OpenAI Responses final tool arguments do not match streamed arguments".into(),
+            ));
         }
     }
 
@@ -453,6 +461,7 @@ fn responses_content(parts: &[ContentPart], text_type: &str) -> Result<Vec<Value
             ContentPart::Text { text } => Some(Ok(json!({"type":text_type, "text":text}))),
             ContentPart::Image { mime_type, data } => Some(Ok(json!({
                 "type":"input_image",
+                "detail":"auto",
                 "image_url":format!("data:{mime_type};base64,{}", STANDARD.encode(data)),
             }))),
         })
@@ -508,8 +517,11 @@ fn responses_usage(value: &Value) -> Usage {
 
 #[cfg(test)]
 mod tests {
-    use super::responses_input;
+    use std::collections::BTreeMap;
+
+    use super::{responses_input, update_response_tool, ResponseToolArguments, ResponseToolState};
     use crate::model::{ContentPart, ProjectedContent, ProjectedMessage, Role, ToolResultContent};
+    use crate::provider::ModelEvent;
 
     #[test]
     fn read_image_stays_in_its_function_call_output() {
@@ -539,5 +551,73 @@ mod tests {
         assert_eq!(input[0]["call_id"], "call");
         assert_eq!(input[0]["output"][0]["type"], "input_text");
         assert_eq!(input[0]["output"][1]["type"], "input_image");
+        assert_eq!(input[0]["output"][1]["detail"], "auto");
+    }
+
+    #[test]
+    fn tool_argument_deltas_are_ordered_bytes_and_final_snapshots_are_idempotent() {
+        let item = serde_json::json!({"call_id": "call-1", "name": "Shell"});
+        let mut tools = BTreeMap::<usize, ResponseToolState>::new();
+        let mut events = update_response_tool(
+            0,
+            &item,
+            ResponseToolArguments::Delta(r#"{"block_until_ms":300"#),
+            false,
+            &mut tools,
+        )
+        .unwrap();
+        events.extend(
+            update_response_tool(
+                0,
+                &item,
+                ResponseToolArguments::Delta("00"),
+                false,
+                &mut tools,
+            )
+            .unwrap(),
+        );
+        events.extend(
+            update_response_tool(
+                0,
+                &item,
+                ResponseToolArguments::Delta("}"),
+                false,
+                &mut tools,
+            )
+            .unwrap(),
+        );
+        events.extend(
+            update_response_tool(
+                0,
+                &item,
+                ResponseToolArguments::Snapshot(r#"{"block_until_ms":30000}"#),
+                true,
+                &mut tools,
+            )
+            .unwrap(),
+        );
+
+        let arguments = events
+            .iter()
+            .filter_map(|event| match event {
+                ModelEvent::ToolCallArgumentsDelta { delta, .. } => Some(delta.as_str()),
+                _ => None,
+            })
+            .collect::<String>();
+        assert_eq!(arguments, r#"{"block_until_ms":30000}"#);
+        assert_eq!(
+            serde_json::from_str::<serde_json::Value>(&arguments).unwrap()["block_until_ms"],
+            30000
+        );
+
+        assert!(update_response_tool(
+            0,
+            &item,
+            ResponseToolArguments::Snapshot(r#"{"block_until_ms":30000}"#),
+            true,
+            &mut tools,
+        )
+        .unwrap()
+        .is_empty());
     }
 }

server/src/provider/router.rs

@@ -82,7 +82,7 @@ impl Provider for ProviderRouter {
                     ProviderType::Anthropic => ProviderKind::Anthropic,
                 },
                 request_url,
-                api_key: endpoint.api_key,
+                api_key: endpoint.endpoint.api_key.clone().unwrap_or_default(),
                 custom_headers: custom_headers(&endpoint.custom_headers)?,
                 max_output_tokens: model.max_output_tokens,
                 request_timeout,

server/src/run/engine.rs

@@ -460,11 +460,8 @@ impl RunEngine {
             .iter()
             .map(|message| message.message_id.as_str())
             .collect::<HashSet<_>>();
-        let compactable = messages
-            .iter()
-            .filter(|message| !current_ids.contains(message.message_id.as_str()))
-            .cloned()
-            .collect::<Vec<_>>();
+        let (compactable, retained_request_context) =
+            auto_compaction_partition(messages, &current_ids);
         if compactable.is_empty() {
             return Ok((revision, None));
         }
@@ -533,7 +530,8 @@ impl RunEngine {
             },
             runtime_event_id: Some(event_id),
         };
-        let mut replacement = vec![summary_message];
+        let mut replacement = retained_request_context.into_iter().collect::<Vec<_>>();
+        replacement.push(summary_message);
         replacement.extend(prepared.initial_messages.iter().cloned());
         let revision = self
             .store
@@ -565,6 +563,29 @@ impl RunEngine {
     }
 }
 
+fn auto_compaction_partition(
+    messages: &[CanonicalMessage],
+    current_ids: &HashSet<&str>,
+) -> (Vec<CanonicalMessage>, Option<CanonicalMessage>) {
+    let latest_request_context = messages
+        .iter()
+        .rposition(|message| message.message_id.starts_with("request-context:"));
+    let compactable = messages
+        .iter()
+        .enumerate()
+        .filter(|(index, message)| {
+            Some(*index) != latest_request_context
+                && !current_ids.contains(message.message_id.as_str())
+        })
+        .map(|(_, message)| message.clone())
+        .collect();
+    let retained = latest_request_context
+        .and_then(|index| messages.get(index))
+        .filter(|message| !current_ids.contains(message.message_id.as_str()))
+        .cloned();
+    (compactable, retained)
+}
+
 fn should_auto_compact(prepared: &PreparedRun, messages: &[CanonicalMessage]) -> bool {
     if prepared.action != RunAction::Start {
         return false;
@@ -724,7 +745,7 @@ fn failure_message(failure: &RunFailure) -> String {
 
 #[cfg(test)]
 mod tests {
-    use super::{estimate_context_tokens, hydrate_tool_images};
+    use super::{auto_compaction_partition, estimate_context_tokens, hydrate_tool_images};
     use crate::{
         model::{
             CanonicalMessage, ContentPart, Origin, ProjectedContent, ProjectedMessage, PromptSpec,
@@ -732,6 +753,7 @@ mod tests {
         },
         store::Store,
     };
+    use std::collections::HashSet;
 
     #[test]
     fn context_estimate_grows_with_prompt_history() {
@@ -756,6 +778,42 @@ mod tests {
         assert!(estimate_context_tokens(&prompt, &long) > estimate_context_tokens(&prompt, &short));
     }
 
+    #[test]
+    fn auto_compaction_preserves_only_the_latest_request_context() {
+        let first_context = CanonicalMessage::text(
+            "request-context:first",
+            Role::User,
+            Origin::Prompt,
+            "old rules",
+        );
+        let old_runtime =
+            CanonicalMessage::text("runtime:first", Role::User, Origin::Runtime, "old query");
+        let latest_context = CanonicalMessage::text(
+            "request-context:second",
+            Role::User,
+            Origin::Prompt,
+            "new rules",
+        );
+        let current_runtime = CanonicalMessage::text(
+            "runtime:current",
+            Role::User,
+            Origin::Runtime,
+            "current query",
+        );
+        let messages = vec![
+            first_context.clone(),
+            old_runtime.clone(),
+            latest_context.clone(),
+            current_runtime,
+        ];
+        let current_ids = HashSet::from(["runtime:current"]);
+
+        let (compactable, retained) = auto_compaction_partition(&messages, &current_ids);
+
+        assert_eq!(compactable, vec![first_context, old_runtime]);
+        assert_eq!(retained, Some(latest_context));
+    }
+
     #[tokio::test]
     async fn read_image_is_loaded_only_for_the_provider_projection() {
         let directory = tempfile::tempdir().unwrap();

server/src/store/providers.rs

@@ -36,7 +36,12 @@ impl Store {
         let mut hashes = Vec::with_capacity(models.len());
         let mut unique_hashes = HashSet::with_capacity(models.len());
         for model in models {
-            let hash = model_hash(&base_url, model.endpoint_type, &model.model_id)?;
+            let hash = model_hash(
+                &base_url,
+                provider.api_key.as_deref().unwrap_or_default(),
+                model.endpoint_type,
+                &model.model_id,
+            )?;
             if !unique_hashes.insert(hash.clone()) {
                 return Err(Error::Config(format!(
                     "8-character model hash collision: {hash}"
@@ -122,11 +127,16 @@ impl Store {
             .provider(provider_id)
             .await?
             .ok_or_else(|| Error::RunNotFound(format!("provider {provider_id}")))?;
-        let api_key = input.api_key.as_deref().unwrap_or(&current.api_key);
+        let api_key = input
+            .api_key
+            .as_deref()
+            .or(current.endpoint.api_key.as_deref())
+            .unwrap_or_default();
         let custom_headers = merge_custom_headers(&current.custom_headers, &input.custom_headers)?;
         let base_url = normalize_base_url(&input.base_url)?;
-        let base_url_changed = base_url != current.endpoint.base_url;
-        let models = if base_url_changed {
+        let identity_changed = base_url != current.endpoint.base_url
+            || api_key != current.endpoint.api_key.as_deref().unwrap_or_default();
+        let models = if identity_changed {
             sqlx::query("SELECT * FROM provider_models WHERE provider_id = ?")
                 .bind(provider_id)
                 .fetch_all(&self.pool)
@@ -140,7 +150,7 @@ impl Store {
         let mut next_hashes = Vec::with_capacity(models.len());
         let mut unique_hashes = HashSet::with_capacity(models.len());
         for model in &models {
-            let hash = model_hash(&base_url, model.endpoint_type, &model.model_id)?;
+            let hash = model_hash(&base_url, api_key, model.endpoint_type, &model.model_id)?;
             if !unique_hashes.insert(hash.clone()) {
                 return Err(Error::Config(format!(
                     "8-character model hash collision: {hash}"
@@ -271,6 +281,7 @@ impl Store {
         for input in inputs {
             let hash = model_hash(
                 &provider.endpoint.base_url,
+                provider.endpoint.api_key.as_deref().unwrap_or_default(),
                 input.endpoint_type,
                 &input.model_id,
             )?;
@@ -327,6 +338,7 @@ impl Store {
             .expect("model provider must exist");
         let next_hash = model_hash(
             &provider.endpoint.base_url,
+            provider.endpoint.api_key.as_deref().unwrap_or_default(),
             input.endpoint_type,
             &input.model_id,
         )?;
@@ -478,6 +490,7 @@ fn validate_model_batch(inputs: &[ProviderModelInput]) -> Result<()> {
 
 fn endpoint_from_row(row: sqlx::sqlite::SqliteRow) -> Result<ProviderEndpoint> {
     let api_key: String = row.try_get("api_key")?;
+    let has_api_key = !api_key.is_empty();
     let headers: serde_json::Value = serde_json::from_str(row.try_get("custom_headers_json")?)?;
     let extra_params: serde_json::Value = serde_json::from_str(row.try_get("extra_params_json")?)?;
     Ok(ProviderEndpoint {
@@ -485,7 +498,8 @@ fn endpoint_from_row(row: sqlx::sqlite::SqliteRow) -> Result<ProviderEndpoint> {
         name: row.try_get("name")?,
         provider_type: ProviderType::from_str(row.try_get("provider_type")?)?,
         base_url: row.try_get("base_url")?,
-        has_api_key: !api_key.is_empty(),
+        api_key: has_api_key.then_some(api_key),
+        has_api_key,
         custom_headers: redact_custom_headers(&headers),
         extra_params,
         created_at_ms: row.try_get("created_at_ms")?,
@@ -494,12 +508,10 @@ fn endpoint_from_row(row: sqlx::sqlite::SqliteRow) -> Result<ProviderEndpoint> {
 }
 
 fn secret_from_row(row: sqlx::sqlite::SqliteRow) -> Result<ProviderEndpointSecret> {
-    let api_key: String = row.try_get("api_key")?;
     let custom_headers: serde_json::Value =
         serde_json::from_str(row.try_get("custom_headers_json")?)?;
     Ok(ProviderEndpointSecret {
         endpoint: endpoint_from_row(row)?,
-        api_key,
         custom_headers,
     })
 }
@@ -680,6 +692,33 @@ mod tests {
         assert_eq!(store.provider_models(false).await.unwrap().len(), 2);
     }
 
+    #[tokio::test]
+    async fn allows_same_endpoint_and_model_with_different_api_keys() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("credential-models.db").display()
+        ))
+        .await
+        .unwrap();
+        let first_provider = provider();
+        let mut second_provider = provider();
+        second_provider.name = "Second".into();
+        second_provider.api_key = Some("different-secret".into());
+
+        let (_, first_model) = store
+            .create_provider_with_model(&first_provider, &model("model-a"))
+            .await
+            .unwrap();
+        let (_, second_model) = store
+            .create_provider_with_model(&second_provider, &model("model-a"))
+            .await
+            .unwrap();
+
+        assert_ne!(first_model.model_hash, second_model.model_hash);
+        assert_eq!(store.provider_models(false).await.unwrap().len(), 2);
+    }
+
     #[tokio::test]
     async fn adds_multiple_models_to_existing_provider_atomically() {
         let directory = tempfile::tempdir().unwrap();
@@ -750,6 +789,7 @@ mod tests {
             models[0].model_hash,
             model_hash(
                 &updated_provider.base_url,
+                input.api_key.as_deref().unwrap(),
                 models[0].endpoint_type,
                 &models[0].model_id,
             )
@@ -764,6 +804,117 @@ mod tests {
         assert_eq!(detached, None);
     }
 
+    #[tokio::test]
+    async fn updating_provider_api_key_rehashes_its_models() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("provider-key-update.db").display()
+        ))
+        .await
+        .unwrap();
+        let (created_provider, original) = store
+            .create_provider_with_model(&provider(), &model("model-a"))
+            .await
+            .unwrap();
+        insert_call(&store, &created_provider, &original).await;
+
+        let mut input = provider();
+        input.api_key = Some("different-secret".into());
+        store
+            .update_provider(created_provider.provider_id, &input)
+            .await
+            .unwrap();
+
+        assert!(store
+            .provider_model(&original.model_hash)
+            .await
+            .unwrap()
+            .is_none());
+        let models = store.provider_models(false).await.unwrap();
+        assert_eq!(models.len(), 1);
+        assert_eq!(
+            models[0].model_hash,
+            model_hash(
+                &created_provider.base_url,
+                "different-secret",
+                models[0].endpoint_type,
+                &models[0].model_id,
+            )
+            .unwrap()
+        );
+        let detached: Option<String> =
+            sqlx::query_scalar("SELECT model_hash FROM llm_calls WHERE call_id = ?")
+                .bind("call-1")
+                .fetch_one(store.pool())
+                .await
+                .unwrap();
+        assert_eq!(detached, None);
+    }
+
+    #[tokio::test]
+    async fn updating_provider_without_changing_api_key_preserves_model_hashes() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("provider-key-keep.db").display()
+        ))
+        .await
+        .unwrap();
+        let (created_provider, original) = store
+            .create_provider_with_model(&provider(), &model("model-a"))
+            .await
+            .unwrap();
+
+        // Editor keeps the configured key: sending it back must not rehash models.
+        store
+            .update_provider(created_provider.provider_id, &provider())
+            .await
+            .unwrap();
+        assert!(store
+            .provider_model(&original.model_hash)
+            .await
+            .unwrap()
+            .is_some());
+
+        // Editor cleared the field: keep the current key, still no rehash.
+        let mut without_key = provider();
+        without_key.api_key = None;
+        store
+            .update_provider(created_provider.provider_id, &without_key)
+            .await
+            .unwrap();
+        assert!(store
+            .provider_model(&original.model_hash)
+            .await
+            .unwrap()
+            .is_some());
+        assert_eq!(store.provider_models(false).await.unwrap().len(), 1);
+    }
+
+    #[tokio::test]
+    async fn providers_expose_the_configured_api_key_for_editing() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("provider-key-echo.db").display()
+        ))
+        .await
+        .unwrap();
+        let created = store.create_provider(&provider()).await.unwrap();
+        assert_eq!(created.api_key.as_deref(), Some("secret"));
+        let listed = store.providers().await.unwrap();
+        assert_eq!(listed.len(), 1);
+        assert_eq!(listed[0].api_key.as_deref(), Some("secret"));
+        assert!(listed[0].has_api_key);
+
+        let without_key = ProviderEndpointInput { api_key: None, ..provider() };
+        let empty = store.create_provider(&without_key).await.unwrap();
+        assert_eq!(empty.api_key, None);
+        assert!(!empty.has_api_key);
+        assert_eq!(store.providers().await.unwrap().len(), 2);
+    }
+
     async fn insert_call(store: &Store, provider: &ProviderEndpoint, model: &ProviderModel) {
         sqlx::query(
             "INSERT INTO llm_calls(call_id, run_id, conversation_id, provider_call_index, model_hash, provider_type, provider_url, request_type, request_url, model_id, display_name, status, created_at_ms, message_count, tool_count, detailed) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",