e2e

In e2e@0.15.1 / @e2e-dev/mobile@0.9.0, a failing unit retires the worker and AgentDeviceSurface.dispose() calls sessions.close(), releasing the agent-device claim before the replacement worker reopens it—letting another session steal the device. Fixed in PR #1209: when a slot holds a provider lease, worker replacement now keeps the session and its device claim open; only release or run finish closes it. Install the fix with npm i @e2e-dev/mobile@nightly e2e@nightly or upgrade to the next stable release after 0.15.1 / 0.9.0.

You don't need a built-in fallback API — use a plain conditional expression on environment variables directly in e2e.config.ts:

export default {
  agents: {
    default: {
      model: process.env.ANTHROPIC_API_KEY
        ? anthropic('claude-sonnet-5')
        : gateway('openai/gpt-6-luna-fast'),
    },
  },
} satisfies E2EConfig;

Each machine then picks the model based on the credential available locally (e.g., ANTHROPIC_API_KEY vs AI_GATEWAY_API_KEY) without editing the committed config. Note: a dedicated firstAvailable core API was proposed but rejected by the maintainers, because putting the model in the replay cache key would break replay across machines and fail at config load. Instead, the chosen model is recorded per recording (see issue #1192) without changing the cache key.

Don't take screenshots asynchronously from onStep — it's a nonblocking progress feed, so captures can land on a later screen. Instead: 1) Upgrade to e2e@0.21.2 and call session.report() to get the built-in step-screen trace with redacted, bounded per-step semantic diffs. 2) Use the onAction hook on agent.act() (added in #1219) — it runs awaited after each action, so calling app.screenshot() inside it is guaranteed to match that action: await agent.act('Open Settings', { onAction: async () => { await app.screenshot(); } }). 3) Artifact screenshots are now retained even after a secret fill (#1218), with existing masking and secret-taint rules still enforced. None of this enables model vision or sends pixels to the model.

Use session.report(), shipped in e2e@0.21.2 (PR #1156). After the session closes, it returns the session's stable report-1 document — the same schema built for the hosted trace — containing run/attempt/target identity, final SDK outcome, step timings/status/errors, recorded events, and existing artifact references. It is available for passed, SDK-failed, and timed-out sessions, so consumers can serialize/upload the document and render their own viewer without parsing Markdown, scanning artifact directories, or importing private runner internals. Existing secret redaction and retention policies remain authoritative (late-resolved secret redaction is tracked separately in issue #1157). The broader debug bundle with an artifact manifest and reasoning excerpts was descoped; open a focused issue if your viewer needs evidence that report-1 doesn't carry.

In e2e 0.21.2, the CLI does not treat video cleanup failures (e.g., CLEANUP_TIMEOUT 'stopping the video timed out') as run failures, so a run with --video on can exit 0 with passed tests but no retained video artifacts. Workarounds: (1) Don't trust exit code 0 alone — wrap the CLI and inspect the run report for cleanup: "failed", secondaryErrors with code CLEANUP_TIMEOUT, and verify video artifacts actually exist. (2) To avoid the timeout itself, reduce screencast size (e.g., from 1440x1100 to 960x720), lower quality to 60, and increase cleanupTimeout from 30000 to 90000 ms. A fix is in progress to surface failed cleanup explicitly in the run result or exit status.