<!-- Add a changeset with `yarn changeset`, or skip the check by labeling the PR with "no changelog". -->
# Why
This is the next part of https://github.com/expo/universe/pull/32694, which allows the requested/required permission role for the build job's EXPO_TOKEN to have to be specified in build mutation.
# How
The first step is to allow specifying this in eas.json. The proposed structure is:
```
...
{
"build": {
"production": {
"permissions": { "account": { "role": "release_manager" } }
}
}
}
...
```
This will, upon updating eas-cli in a separate PR, set the requested role in the build creation mutation. If the creator doesn't possess the required role, it will throw.
# Test Plan
CI
<!-- Add a changeset with `yarn changeset`, or skip the check by labeling the PR with "no changelog". -->
# Why
Regenerate graphql to fix CI ahead of others PRs failing due to needing to sync.
# How
regenerate.
# Test Plan
`yarn tsc`, CI
# Why
Workers can set `NVM_NODEJS_ORG_MIRROR` through the Node.js cache settings. If `nvm install` fails with that mirror, runtime setup currently stops even when the default Node.js download source is available.
# How
Retry `nvm install` without the mirror if it failed with the mirror. This helps with situation when we fail to have refreshed Nodes list in our cache or our cache is down and stuff like that.
# Test Plan
CI should pass.
# Why
We recently changed the default Android system image from API 30 to API 35. Looks like API 35 needs more RAM to be working well.
<details>
<summary>Details</summary>
Android 35 Google APIs x86_64 can lose ADB during large APK transfers. Guest tracing shows `adbd` getting `ENOMEM` from `write()`, then closing VSOCK before the emulator closes host TCP. Six fresh live EAS devices with 4 GiB passed six Maestro tests and 60 additional APK downloads. Two of three corrected lower-memory control devices lost ADB.
</details>
# How
Use 4096 MB when the resolved system-image package starts with `system-images;android-35` or newer.
# Test Plan
- Existing emulator utility and startup suites: 30 tests passed. No tests added. Updated one existing API 35 AOSP RAM expectation from 2048 to 4096.
- Type checks for all 13 projects, repository lint, targeted formatting, and `git diff --check` passed.
- [Live validation of the 4 GiB setting](https://expo.dev/accounts/sjchmiela/projects/sdk-57-production/workflows/01a165c0-9d4d-777b-920f-3827ff2f89b4): three devices, three Maestro tests, 30 APK downloads passed.
- [Default API 30 comparison](https://expo.dev/accounts/sjchmiela/projects/sdk-57-production/workflows/01a16833-cf25-7eac-b472-e64dc6084c3d): 2/3/4 GiB, three fresh workers per setting, complete.
- [API 34 comparison](https://expo.dev/accounts/sjchmiela/projects/sdk-57-production/workflows/01a16833-cf57-70be-a4e8-72a20ca8d6b2): 2/4 GiB, three fresh workers per setting, complete.
- [API 35 comparison](https://expo.dev/accounts/sjchmiela/projects/sdk-57-production/workflows/01a16837-36ae-7bfe-9518-28fe72b9cd5c): 2/3/4 GiB, three fresh workers per setting, complete.
| Image | RAM request | Full passes (flow plus ten downloads) |
| --- | --- | --- |
| Default API 30 | 2 / 3 / 4 GiB | 3/3 / 3/3 / 3/3 |
| API 34 Google APIs x86_64 | 2 / 4 GiB | 3/3 / 3/3 |
| API 35 Google APIs x86_64 | 2 / 3 / 4 GiB | 1/3 / 0/3 / 3/3 |
API 35 at 2 and 3 GiB each had two Maestro failures with `device offline`. One 3 GiB device passed Maestro but lost ADB during extra download four. All 4 GiB devices passed all 30 extra downloads. Three devices per cell do not establish a long-term failure rate.
* [eas-cli] Share one GraphQL client between the project ID check and the command
The project ID check and LoggedInContextField each built their own urql
client, so the second AppByIdQuery for the same project missed the cache.
SessionManager now keeps one client per credential.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Update test mocks for the analytics and GraphQL client changes on main
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: Claude Opus 5.5 <[email protected]>
Hook parsing in native builds runs between READ_PACKAGE_JSON and
INSTALL_DEPENDENCIES and logged through the phase-less logger, so every
native build with hooks showed an "Unknown build phase" log group.
Ignore unknown and unrun hook keys silently in both the native and the
steps parser, let the eas/checkout step print the working-directory line
itself, and drop the after-hook failure line that the failing hook step
already logs.
* [eas-cli] make command descriptions lowercase without trailing periods
Most command descriptions in eas --help already start lowercase and end
without a period; align the remaining 11 capitalized and 12 period-terminated
ones with that style. Proper nouns, [EXPERIMENTAL] prefixes, and later
paragraphs of multi-line descriptions are unchanged.
* [eas-cli] Drop trailing periods from simulator command descriptions
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>
# Why
I want to set up billing and limits for sandboxes so we need to be able to say how long can a sandbox run for.
# How
Added simple `max_duration_seconds` like we have in agent runs and device run sessions.
# Test Plan
CI should pass.
Look up why a failed simulator session failed in both simulator:start and
simulator:get. For unknown errors, read the failed step's cause from the job
run log, using only the masked log message, with the download bounded by size
and time. Strip terminal control characters before printing.
* [eas-cli] Show why a simulator session failed to start
simulator:start only printed the session status when a session errored
before it was ready. The reason is already on the session's job run, so
fetch turtleJobRun.errors and print each message (and docs link) on its
own line.
* [eas-cli] Read simulator job run errors only on failure, and from logs for unknown errors
Fetch job run errors with a separate query when a session fails instead of
on every poll of the shared session query. When the job run only reports an
unknown error, print the failed step errors from its logs. Also show the
errors when a session fails after it is ready.
* [eas-cli] Show only structured job run errors in simulator:start
Drop the job run log fallback from this change, which printed unmasked
err.message lines, read logs for stopped sessions and had no timeout or size
limit. The bounded, masked log fallback lands in the simulator:get follow-up.
* [eas-cli] Bound and quiet the simulator job run errors lookup
Give the lookup a 10s deadline and the session's Ctrl+C signal, and cancel
its request by unsubscribing from urql, so it can't hold up stopping the
session. Don't retry it or print the transient-error message, since it's
best-effort. Cap each error message at 500 characters.
* [worker] Mask secrets in logged errors
The upload transform masked only each record's msg, so a serialized err
(message, stack and other fields) was uploaded with secrets intact. That
includes unknown build phase failures, which log { err } with an empty msg.
Mask the strings inside err on a copy, since bunyan shares the record with
other streams.
* [worker] Mask secrets in every field of uploaded log records
Mask all fields except bunyan's metadata, not only err: other fields such as
error, data and the job dump also carried secrets. Round-trip each value
through JSON first so a toJSON method can't bring back the original, mask keys
and numbers, and replace values nested too deeply instead of uploading them
unmasked.
* [build-tools] Check Android launch inputs before the emulator boots
Agent-device sessions checked Android launch inputs only when launching,
after the emulator booted and the build installed. Check NUL characters
and the command length at the start of the step, as on iOS, leaving room
for the launch component that is only known after install.
* Add changeset for early Android launch input checks
* [eas-build-job] Share the Android launch input checks
Move the Android shell quoting and the early launch input checks into
@expo/eas-build-job, so www can check launch inputs when a session is
created, for every session type.
* [eas-build-job] Check only the known part of the Android launch command early
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>
* [build-tools] Quote Android launch arguments for the device shell
adb shell joins its arguments with spaces and the device shell parses the
result again, so a launch argument such as "hello world" reached am start
as two words and the launch failed. Quote each word, reject commands that
adb cannot carry, and reject empty arguments that are not extra values.
* [build-tools] Pass empty Android launch arguments through and reject NUL
Guessing which am start option an empty argument belongs to rejected
valid input such as -e key "" and -d "", and accepted input that am
start then rejects. Pass empty arguments through, and explain a failed
launch when launch_args has an empty element. Reject NUL characters,
which spawn cannot pass, with a user error.
Poll serve-sim's event log and reset the existing idle timeout when a
new preview input event appears.
Co-authored-by: Claude Opus 5.5 <[email protected]>
* [eas-cli] Add a checkout prompt experiment to the Free plan build limit error
* [eas-cli] Reuse the account billing URL helper in the checkout experiment
* [eas-cli] Add simulator:action command
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Fold simulator actions into one table
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Give the simulator API lookup its own timeout
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Drop event-log from simulator:action
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Rename to simulator:control and add sidebar actions
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Drive simulator:control from the serve-sim action catalog
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Fix value order, upload cleanup and output handling in simulator:control
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Check output paths and local files before simulator:control runs
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Upload only when the catalog condition matches
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Send upload IDs and JSON lists from simulator:control
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* [eas-cli] Complete the simulator control session fixture
Co-authored-by: Codex <[email protected]>
* [eas-cli] Close simulator control after discovery failure
Co-authored-by: Codex <[email protected]>
* [eas-cli] Keep simulator action descriptions on one line
Co-authored-by: Codex <[email protected]>
* [eas-cli] List simulator actions by default
Co-authored-by: Codex <[email protected]>
* [eas-cli] Reject empty simulator control session IDs
Document the separator for values beginning with --.
Co-authored-by: Codex <[email protected]>
* [eas-cli] Clarify local screenshot output usage
Co-authored-by: Codex <[email protected]>
* [eas-cli] Compact simulator action discovery and add action help
---------
Co-authored-by: Claude Opus 5.5 <[email protected]>
Co-authored-by: Codex <[email protected]>
* [eas-cli] Add experimentation client
* [eas-cli] Document experimentation context order and override semantics
* [eas-cli] Take the account unit from the command instead of the app config
* [eas-cli] Skip experiments in CI
* [eas-cli] Make the GraphQL request timeout take effect
* [eas-cli] Document which experiment units are shared across surfaces
* [eas-cli] Test that non-timeout fetch errors pass through the timeout wrapper
* [eas-cli] Fetch the experimentation config lazily on first variant read
* [eas-cli] Cap the actor lookup in the experimentation context field
* [eas-cli] add --frame-ancestor to simulator:start and print the embed URL
Co-Authored-By: Claude Fable 5.1 <[email protected]>
* [eas-cli] cover the embed output fallbacks
Co-Authored-By: Claude Fable 5.1 <[email protected]>
* [eas-cli] format the embed output tests
Co-Authored-By: Claude Fable 5.1 <[email protected]>
---------
Co-authored-by: Claude Fable 5.1 <[email protected]>
# Why
Google Play submissions sometimes fail with generic errors like `Precondition check failed.` (HTTP 400, `FAILED_PRECONDITION`). We currently drop everything except `error.message` and `errors[].reason`, so neither the user nor we can tell what Google actually objected to.
This replaces #4650, which also put Google's details into user-facing messages and structured logs, and added a redacting sanitizer. That made the code a lot more complex. This PR keeps build logs unchanged and only makes the raw response available in Sentry. Once real payloads show up there, we can add targeted classifiers to `mapGooglePlayError` with clear user-facing messages.
# How
- `GooglePlayApiError` now keeps:
- the JSON error response body, as a string capped at 8 KB
- the request URL path, without the query (the upload session query contains a secret)
Non-JSON bodies are still dropped, as before.
- `GooglePlayApiError.getErrorMetadata()` finds the API error in the cause chain, through retry `AggregateError`s and commit-outcome-unknown wrappers. It returns `googlePlayRequestPath` and `googlePlayErrorResponse`.
- `mapGooglePlayError` and `GooglePlayCommitOutcomeUnknownError` attach that as `ExpoError.metadata`. The worker already sends it to Sentry as `extras.errorMetadata`, and it is not part of the external build error.
- The body is stored as a string because `@sentry/node` 7 normalizes extras to depth 3, so nested `details` would otherwise show up as `[Object]`.
User-facing messages are unchanged.
# Test Plan
- New tests in `GooglePlaySubmission.test.ts`:
- a `FAILED_PRECONDITION` commit with `details` ends up in metadata but not in the message
- metadata survives retries when the commit outcome is unknown
- the upload-session query and non-JSON bodies are not stored
- `yarn test src/steps/utils/android src/steps/functions/__tests__/submitToGooglePlay.test.ts` (153 passed), plus `yarn typecheck`, `yarn lint`, `yarn fmt:check`
🤖 Generated with [Claude Code](https://claude.com/claude-code)