This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@e2e-dev/[email protected]

Minor Changes

• #730 1554360 Thanks @szdziedzic! - easSimulators() authenticates with the eas-cli login when EXPO_TOKEN is not set: it reads the session eas login keeps in ~/.expo/state.json (under USERPROFILE on Windows) and sends it in the expo-session header, as eas-cli does, so a machine signed in to eas-cli needs no token. EXPO_TOKEN still wins when both are there. A session is stopped as the account that started it, even if the login changes during the run. Only the production login is read, since the sessions API is api.expo.dev. With neither a token nor a login, the lease fails with EXPO_TOKEN is not set and eas-cli is not logged in.

• #731 88cef2c Thanks @szdziedzic! - easSimulators() takes projectId from the app config when the option is absent: extra.eas.projectId, the id eas init writes, read from app.config.json or app.json beside e2e.config.ts, or from a dynamic app.config.ts (or .js, .mjs, .cjs, .mts, .cts) as the project's own expo config --type public evaluates it in the run's environment, without .env files, as eas-cli runs it. It reads the config once per run, before the first session starts; a config that links no project fails the lease naming the file. projectId is optional now and still wins when it is passed. Reading it needs the @e2e-dev/mobile that passes projectRoot to device providers.

@e2e-dev/[email protected]

Minor Changes

• #731 88cef2c Thanks @szdziedzic! - A DeviceProvider's acquire request carries projectRoot, the directory the config's relative paths resolve against, so a provider reads project files there instead of process.cwd().

• #740 bb904d4 Thanks @NathanWalker! - device.fold(pose) puts a foldable iOS simulator (iPhone Duo) in a hinge pose: 'closed' lights the outer display, 'half-open' (a 130° book) and 'open' the inner one. It runs agent-device's fold command, which sends the simulator's hinge event and reads the angle back from CoreDevice before it resolves, so the next observation reads the app on the other display. Like setOrientation, it counts as an action for the transition budget. A device without a hinge is UNSUPPORTED_CAPABILITY, and Android is refused before any device command. The pose type is exported as FoldPose.

[email protected]

Patch Changes

• #710 6b25a3a Thanks @okwasniewski! - A failure where no model answered (MODEL_PROVIDER_FAILED, MODEL_UNAVAILABLE) no longer evicts replay cache entries. It says nothing about the app, so a read-write run without model access keeps every recording it did not confirm, including a step that replayed whole before an agent.assert that could not reach its model and a replay that handed off to the model mid-step, and still saves the steps verified before the failure.

• #736 c9c3e14 Thanks @programmrz! - Agent steps on DeepSeek's V4 models no longer fail on their first turn with MODEL_PROVIDER_FAILED. In thinking mode, DeepSeek's default, the API refuses a forced tool choice with Thinking mode does not support this tool_choice; that refusal now reads as one, so the step retries with auto and the tool-calls-only rule, as it already did for Anthropic, and thinking can stay on.

• #735 28b09f3 Thanks @okwasniewski! - e2e init adds zod next to ai when it sets up a model gateway. zod is a peer of ai and every provider package; npm and pnpm install peers on their own, Yarn does not, so a Yarn project failed to load the generated config with CONFIG_LOAD_FAILED. A Yarn project scaffolded with a model gateway before this fix needs yarn add -D zod.

• #708 4ee772c Thanks @okwasniewski! - The package ships the documentation pages: coding agents can read them offline from node_modules/e2e/docs.

• #742 4f17480 Thanks @okwasniewski! - expect(value).toBeDefined() now passes on null, as it does in Playwright, Jest, and Vitest; only undefined fails it. It used to reject null too, so expect(null).not.toBeDefined() passed where Playwright fails. expect.poll(read).toBeDefined() follows: a read that returns null now passes on the first try instead of waiting. For the old non-nullish check, use toEqual(expect.anything()).

• #739 8d38206 Thanks @NathanWalker! - The agent loop handles providers that downgrade a forced tool choice instead of refusing it, as the AI SDK's Anthropic provider does for Claude models that reject forced tool use (Claude Sonnet 5.5 through @ai-sdk/anthropic). Such a request succeeds with auto, so the HTTP 400 path never ran: the loop kept asking for a forced choice, the SDK logged an AI SDK Warning ... toolChoice line on every turn, steps ran without the tool-calls-only instruction, and a turn the model answered in prose failed the step with MODEL_PROVIDER_FAILED (the SDK's ToolChoiceViolationError). Now a toolChoice call warning switches the step to auto and gives its remaining turns the instruction, a prose answer under a forced choice is retried with auto like a refusal, and later steps on that model start in auto.