Merge branch 'main' into test-e2e-grpc-not-set

Author: ekam-walia

.github/workflows/ci.yaml

@@ -358,33 +358,45 @@ jobs:
         uses: actions/setup-node@v7
         with:
           node-version-file: ui/.nvmrc
-          cache: "npm"
-          cache-dependency-path: ui/package-lock.json
 
-      # Honor the pinned npm from ui/package.json "packageManager" so npm ci
-      # resolves the lock file with the same npm version it was generated with.
+      # Before any cache step that shells out to yarn: the pinned version in
+      # ui/package.json "packageManager" is Yarn 4, and without corepack the shim
+      # on the runner is a different one that cannot read this lock file.
       - name: Enable Corepack
         run: corepack enable
 
+      - name: Cache Yarn downloads
+        uses: actions/cache@v4
+        with:
+          path: ui/.yarn/cache
+          key: yarn-${{ runner.os }}-${{ hashFiles('ui/yarn.lock') }}
+          restore-keys: yarn-${{ runner.os }}-
+
       - name: Install dependencies
         working-directory: ./ui
-        run: npm ci
+        run: yarn install --immutable
+
+      - name: Typecheck
+        working-directory: ./ui
+        run: yarn typecheck
 
       - name: Run lint
         working-directory: ./ui
-        run: npm run lint
+        run: yarn lint
 
-      - name: Run unit tests (Jest)
+      - name: Run unit tests
         working-directory: ./ui
-        run: npm run test
+        run: yarn test
 
-      - name: Install Playwright browser (Chromium)
+      # Both engines the suite declares. Installing only one leaves that project
+      # failing to launch, which reads as a broken app rather than a missing browser.
+      - name: Install Playwright browsers
         working-directory: ./ui
-        run: npx playwright install --with-deps chromium
+        run: yarn playwright install --with-deps chromium firefox
 
-      - name: Run Storybook tests (Vitest + Playwright)
+      - name: Run browser tests
         working-directory: ./ui
-        run: npm run test:vitest
+        run: yarn test:pw
 
   # This job builds the Docker images for the controller, UI, ADKs, and CLI on arm64.
   build:

.github/workflows/ui-chromatic.yaml

@@ -108,6 +108,85 @@ Common commands:
 - Do not commit or push unless asked.
 - Keep PRs focused. Explain non-obvious invariants and operational tradeoffs, not line-by-line implementation details.
 
+## The web interface (`ui/`)
+
+A Vite single-page app. It is a static bundle served by nginx: there is no server
+process, so there are no server components, no server-side data fetching and no
+file-system routing.
+
+**Stack:** Vite + React 19, TypeScript, antd 6 for components, Emotion for styling
+(the `css` prop, via `jsxImportSource`), SWR for reads, Yarn 4. React Router owns
+routing; there is no file-system routing and no server rendering.
+
+### Commands
+
+Run these from `ui/`:
+
+| Task | Command |
+|------|---------|
+| Dev server | `yarn dev` |
+| Unit tests | `yarn test` |
+| End-to-end, no cluster needed | `yarn test:pw` (Chromium and Firefox) |
+| End-to-end against a real cluster | `yarn test:pw:live` |
+| Type check | `yarn typecheck` |
+| Lint | `yarn lint` |
+
+Only lint **errors** gate a change; a handful of warnings are pre-existing.
+
+`scripts/setup-cluster/setup-cluster.sh` builds a Kind cluster with kagent on it in one
+command, for work that needs a real backend.
+
+### Settings reach the app at runtime, not at build time
+
+Configuration is read from `window.environmentVariables`, which the container
+rewrites from its own environment on every start. So one image serves every
+deployment, and a setting is an operator's decision rather than something frozen
+into a build. Locally the same values come from `ui/.env` (git-ignored;
+`ui/.env.example` documents each one).
+
+Two consequences worth knowing before touching that code:
+
+- The script that supplies them is **synchronous** in `index.html`. Several modules
+  read settings at import time, so anything awaited would be read before it arrived.
+- `import.meta.env` is for build-time flags only. A value that an operator should be
+  able to change belongs in `window.environmentVariables`.
+
+### Fixtures are opt-in
+
+`ENABLE_MOCK_UI=true` serves the whole API from an in-browser mock (MSW) with no
+cluster at all, and `?mock=ok|empty|error|slow` picks which scenario the fixtures
+play. **It is off unless asked for**, in a dev server exactly as in a built image: a
+page that quietly serves fixtures when the backend is down looks healthy while
+showing data that was never real.
+
+When mock mode is on it overrides every backend setting, and anything reporting who
+is signed in correctly reports nobody — there is no backend to have signed in to.
+
+### Extension points
+
+One `VendorExtensionConfig` contributes navigation entries and overrides, routes and
+route handles, slots, form fields, table columns, API overrides, providers, theme
+tokens, shell regions, branding, provider icons and agent links. Components read
+every colour, radius and font from those tokens, so overriding them restyles
+components an extension never touches. When adding a feature, check whether it
+belongs behind an extension point rather than as a branch inside a shared component.
+
+The full guide is [ui/docs/vendor-extensions.md](ui/docs/vendor-extensions.md).
+
+### Conventions specific to this codebase
+
+- **Say when data is not real.** A page showing fixtures says so on the page. Never
+  suppress an error because a mock flag is set — a broken backend must not render as
+  healthy mock data.
+- **Normalise at the client boundary.** Go marshals a nil slice as JSON `null`, so
+  any collection the controller has nothing for arrives as null. Fix it once where the
+  response is parsed, not at each use.
+- **Fixtures must match the controller, not each other.** A fixture, a type and a
+  test can agree perfectly and all three be wrong; that has happened here more than
+  once and each time only a real cluster objected. Check the CRD.
+- **Prefer a smaller honest test suite** over a green one that proves nothing.
+  Coverage debt belongs in `playwright/DEFERRED.md`, not in skipped specs.
+
 ## 9. References
 
 - [STYLE.md](STYLE.md)

.github/workflows/ui-playwright.yaml

@@ -326,6 +326,10 @@ MCP:
 
 - Discover ready AgentInstances.
 - Invoke them only through the public gateway A2A path.
+- Expose durable A2A turns through the MCP Tasks extension, including polling,
+  cancellation, and input-required continuation.
+- Keep synchronous `tools/call` fallback for clients without Tasks support.
+- Store no MCP-owned task or session state.
 - Never expose Actor or private MCP endpoints.
 
 Content:

CLAUDE.md

@@ -94,7 +94,6 @@ type Client interface {
 	GetSessionShareByToken(ctx context.Context, token string) (*SessionShare, error)
 	ListSessionSharesBySession(ctx context.Context, sessionID string) ([]SessionShare, error)
 	DeleteSessionShare(ctx context.Context, token, sessionID, userID string) error
-	RecordShareAccess(ctx context.Context, userID string, shareID int64) error
 
 	// Agent memory (vector search) methods
 	StoreAgentMemory(ctx context.Context, memory *Memory) error
@@ -124,12 +123,19 @@ type Client interface {
 	CreateAgentInstance(context.Context, *apiv1alpha1.AgentInstance, string) (*apiv1alpha1.AgentInstance, bool, error)
 	ForkAgentInstance(context.Context, string, string, string, string, string) (*apiv1alpha1.AgentInstance, bool, error)
 	GetAgentInstance(context.Context, string, string, string) (*apiv1alpha1.AgentInstance, error)
-	ListAgentInstances(context.Context, string, string, bool, map[string]string, string, int) ([]*apiv1alpha1.AgentInstance, error)
+	ListAgentInstances(context.Context, AgentInstanceQuery) ([]*apiv1alpha1.AgentInstance, error)
+	// UpdateAgentInstanceName sets the instance's display name, scoped to its owner.
+	// Takes namespace, id, owner and the new name.
+	UpdateAgentInstanceName(context.Context, string, string, string, string) (*apiv1alpha1.AgentInstance, error)
 	MarkAgentInstanceReady(context.Context, string, string) (*apiv1alpha1.AgentInstance, error)
 	TransitionAgentInstance(context.Context, *apiv1alpha1.AgentInstance, apiv1alpha1.AgentInstanceState, apiv1alpha1.AgentInstanceOperation) (*apiv1alpha1.AgentInstance, error)
 	DeleteAgentInstance(context.Context, string) error
 	CreateAgentInstanceShare(context.Context, AgentInstanceShare) (*AgentInstanceShare, error)
 	ListAgentInstanceShares(context.Context, string, string, string, string, int) ([]AgentInstanceShare, error)
+	// GetAgentInstanceShareByTokenHash resolves a share token to its share and the
+	// owner of the instance it grants access to. Takes the digest, because only the
+	// digest is stored.
+	GetAgentInstanceShareByTokenHash(context.Context, []byte) (*AgentInstanceShare, error)
 	DeleteAgentInstanceShare(context.Context, string, string, string) error
 	// CreateAgentInstanceTask reserves the instance's single active-task slot.
 	CreateAgentInstanceTask(context.Context, string, []byte, *a2a.Task) (*a2a.Task, bool, error)

docs/plans/api-v2-execution-plan.md

@@ -270,14 +270,38 @@ type RuntimeRevision struct {
 	GoldenSnapshot         string
 }
 
+// AgentInstanceQuery narrows a page of AgentInstances. Zero values mean "do not
+// filter on this", so an empty query lists the caller's own instances in the
+// namespace.
+type AgentInstanceQuery struct {
+	Namespace   string
+	UserID      string
+	AllUsers    bool
+	MatchLabels map[string]string
+	// AgentTemplate and Harness name the agent whose conversations are wanted.
+	// They are matched against the (AgentTemplate, Harness) pair the instance's
+	// prepared revision was built from, not against its labels, so they select
+	// instances stored before either field existed.
+	AgentTemplate string
+	Harness       string
+	AfterID       string
+	Limit         int
+}
+
 type AgentInstanceShare struct {
 	ID         string
 	Namespace  string
 	InstanceID string
-	Creator    string
 	Permission string
 	TokenHash  []byte
 	CreatedAt  time.Time
+	// OwnerUserID is the user the shared AgentInstance belongs to.
+	//
+	// Populated only by the token lookup, which joins it in — that is what the
+	// share grants. A visitor is authenticated as themselves and the token widens
+	// what their account may reach to what the *owner* can see, so the instance
+	// read has to run as the owner or it finds nothing.
+	OwnerUserID string
 }
 
 // AgentInstanceTaskSnapshot identifies the immutable Substrate snapshot at a