The new normalization in `src/snapshot-tests/__tests__/json-normalize.test.ts` validates the rs/1 capture envelope (`protocol`, `capturedAtMs`, `expiresAtMs`, `elements`, `actions`) but three stored JSON fixtures still serialize the old runtime-snapshot shape (`rs: "1"`, `targets`, `scroll`, `udid`, no `protocol`/`capturedAtMs`/`expiresAtMs`/`elements`/`actions`): `src/snapshot-tests/__fixtures__/mcp/json/ui-automation/snapshot-ui--success.json`, `…/mcp/json/ui-automation/wait-for-ui--success.json`, and `…/cli/json/ui-automation/wait-for-ui--success.json`. Because `snapshot_ui.ts:100-101` passes `snapshot.payload` (built by `createRuntimeSnapshotRecord` in `src/mcp/tools/ui-automation/shared/runtime-snapshot.ts:556–591`, which emits the new fields) directly into `createCaptureSuccessResult`, the normalized output cannot match the stored fixtures. This violates the skill guardrails 'Fixture updates map to intentional behavior changes' and 'Missing fixtures are generated through the snapshot update flow' — these fixtures appear to have been left behind by the rs/1 refactor. (Note: the CLI `snapshot-ui--success.json` is a different `type: ui-hierarchy` envelope and is not affected.)

The e2e smoke test at `src/smoke-tests/__tests__/e2e-mcp-ui-automation.test.ts:226` calls `long_press` with `{ x: 150, y: 300, duration: 500 }`, but the schema no longer accepts `x` or `y` — it requires `elementRef`. The test will now fail with a validation error. Update it to call `long_press` with a valid `elementRef` (obtained after a `snapshot_ui` call).

Replace the inline `createBootedScreenshotMockExecutor` with the existing `createCommandMatchingMockExecutor` helper from `mock-executors.ts`, which already handles multi-command matching for exactly this kind of multi-step tool flow.

Replace the inline `createBootedScreenshotMockExecutor` with the existing `createCommandMatchingMockExecutor` helper from `mock-executors.ts`, which already handles multi-command matching for exactly this kind of multi-step tool flow.

The swipe `'success'` test case (previously asserting `isError: false` against `swipe--success`) is entirely replaced by an error-only test, removing all snapshot verification of the swipe success path; the pre-existing `swipe--success.json` fixture is now an unreferenced orphan.

`wait_for_uiLogic` sets `ctx.nextStepParams` (lines 364–367) but never sets `ctx.nextSteps`, and `wait_for_ui.yaml` declares no `nextSteps` templates; `postProcessSession` only merges `nextStepParams` into template-based steps, so both suggested tools (`snapshot_ui`, `wait_for_ui`) will never surface to the model — set `ctx.nextSteps` directly (as `snapshot_ui.ts` does via `createRuntimeSnapshotNextSteps`) instead.

When `executeAxeCommand` throws (e.g. `AxeError` or `SystemError`) after a successful UI action, the returned `uiError.code` is `'ACTION_FAILED'`, but `message` states "UI action succeeded." Agents inspecting the error code will incorrectly conclude the UI action itself failed and may retry it, potentially double-applying the action.

After `applyViewportVisibility` sets `publicElement.actions = []` for elements that don't intersect the viewport, `inferScrollableContainers` runs next and can push `swipeWithin` back onto those same off-screen `scroll-view`, `list`, and `other` containers — because it only checks `isVisible(frame)` (frame dimensions) and `actions.includes('swipeWithin')`, neither of which reflect the cleared viewport state. This surfaces invalid swipe-action hints to agents for elements they cannot actually interact with.

When the post-action `describe-ui` call fails for any reason other than a parse error, the returned `uiError.code` is `'ACTION_FAILED'`, but the message explicitly says "UI action succeeded." An agent receiving `code: 'ACTION_FAILED'` will likely retry the action (e.g., a tap or text entry), causing duplicate/unintended interactions.

The pattern `/statictext|text/` is a plain substring match, so any role/type whose lowercased concatenation contains "text" (e.g. "axcontextmenu" contains "con**text**") is incorrectly classified as `'text'` before the `/menu/` check is ever reached, suppressing correct menu-role action derivation for agents.

Replace the four-level nested ternary (which includes an IIFE for the `gone` branch) with an `if/else` chain or a helper function — the skill explicitly prohibits nested ternaries and requires preferring switch/if-else chains for multiple conditions.

Replace the four-level nested ternary (which includes an IIFE for the `gone` branch) with an `if/else` chain or a helper function — the skill explicitly prohibits nested ternaries and requires preferring switch/if-else chains for multiple conditions.

Replace the four-level nested ternary (which includes an IIFE for the `gone` branch) with an `if/else` chain or a helper function — the skill explicitly prohibits nested ternaries and requires preferring switch/if-else chains for multiple conditions.