Task 13.5: match a release bundle's debug id to the composed map

[krassx]

Summary

• A Hermes release build on the WOD_LX1 reports the same debug id the composed source map carries: 3c110b23-2416-5bc6-acac-e35ac7e69aa6 for index.android.bundle.
• debug-files upload is still not executed. The Task 13.5 checkbox stays open because the dashboard has not symbolicated a frame. No Bugsee app was created. The example credentials are the CI placeholder.

Test plan

• WOD_LX1 AMRJCP4718402860, release, E2E_RELEASE=1: runtime debug_ids equals the composed map's debug_id
• iPhone XS was offline and was not run
• iOS Simulator was not run: beta3 compiles logException out there, so a Release simulator build does not retain debug_ids
• Dashboard symbolication, once a token for an existing app is available

Made with Cursor

[cursor]

Stale comment

Deep review — Task 13.5 local half (Gradle 9 inject + composed debug id)

Reviewed 5cd1209 against surrounding 13.1/13.4 code (hermes-sourcemaps.js, ensureHermesHooks, debugIds.ts, bugsee-xcode.sh, existing exception e2e). The production change is the Gradle 9 Project.exec replacement so hermes-sourcemaps.js finish actually runs after compose. The new exc-map-id path correctly leaves the inject stub alone. The dashboard / debug-files upload half is still open, and the plan checkbox stays unchecked — that matches the evidence.

I did not run unit tests or a device e2e in this environment.

Findings

P2 — Expo plugin will not migrate an existing project.exec hook

• Location: packages/react-native/plugin/src/gradle.ts (ensureHermesHooks, used with the new ExecOperations line)
• Problem: ensureHermesHooks returns the file unchanged once PRESERVE_REFUSAL is present. A 13.4 prebuild already contains that marker and project.exec. Re-running the config plugin (including expo prebuild without --clean) never rewrites the exec call. ensureAppAppliesPlugin(next) === next in plugin.test.ts locks that write-once behavior in.
• Impact: The Gradle 9 inject fix is what this PR exists to ship for customers. The bare example was patched by hand; Expo apps that already have the 13.4 hook keep the removed API. createBundle*JsAndAssets then throws MissingMethodException on project.exec, so the composed map is never injected and a release JS stack cannot join the uploaded map.
• Scenario: Prebuild with 13.4 on RN 0.87 / Gradle 9.4 (this example's wrapper). Upgrade @bugsee/react-native to this commit. Run expo prebuild without --clean. android/app/build.gradle still has project.exec. assembleRelease hits the finish doLast (preserved.isFile() is true) and fails before hermes-sourcemaps.js finish.
• Fix: When the marker is already present, rewrite only the Bugsee finish-hook project.exec { (the 20-space line inside that doLast) to the ExecOperations call. Do not globally replace project.exec in the whole file — bundleTask is not in scope elsewhere. Add a test that starts from a 13.4 hook and asserts the exec line changed. Prefer capturing ExecOperations at configureEach time (objects.newInstance / injected interface) rather than bundleTask.services.get inside doLast; that is the public Gradle 9 replacement.

P2 — iOS device suite is enabled but the composed map path is Android-only

• Location: examples/bare/e2e/composed-debug-id.test.ts (composedMapPath)
• Problem: describeReleaseProof runs on iOS device, but the default map is android/app/build/generated/sourcemaps/react/release/index.android.bundle.map. iOS writes $CONFIGURATION_TEMP_DIR/main.jsbundle.composed.map (bugsee-xcode.sh). BUGSEE_COMPOSED_MAP exists, but nothing requires it on iOS.
• Impact: The iOS half of this proof cannot succeed as written. You either throw on a missing Android map, or compare the iPhone runtime UUID to a leftover Android UUIDv5 (different bundle content → different id) and fail for the wrong reason.
• Scenario: E2E_RELEASE=1 E2E_PLATFORM=ios E2E_IOS_TARGET=device on the XS, as the describe gate allows. beforeAll reads the Android path first, before any launch.
• Fix: Skip iOS unless BUGSEE_COMPOSED_MAP is set, or persist the iOS composed map to a stable checkout path during the Xcode hook and default to that when ON_IOS.

Overall risk: Medium

The example's ExecOperations lookup did what 13.5 needed on WOD_LX1 (debug_ids {"index.android.bundle":"3c110b23-2416-5bc6-acac-e35ac7e69aa6"} matching the composed map). The customer-facing Expo hook does not apply that same fix to trees that already have the 13.4 snippet. iOS is opted in without a map it can read.

Merge recommendation

Request changes. Do not merge until the plugin rewrites an existing project.exec finish hook (with a test). The iOS default path should not run against the Android map.

Most important issues to fix

1. Migrate the already-written Hermes finish hook off project.exec in ensureHermesHooks.
2. Stop the iOS device describe from using the Android composed map unless an override is provided.

What looks good

• Replacing Project.exec is the correct Gradle 9 fix; all finish arguments are already absolute, so working-directory differences vs Project.exec do not matter here.
• exc-map-id skips registerDebugIds, which would have replaced the inject stub and hidden the real UUID. Launch still goes through startRun (URI on Android, launch args on iOS), so a release binary can select the scenario even though e2e-scenario.json is baked in.
• expect(run.dev).toBe(false), airplane mode, nonce-scoped reason, and debug_id/debugId equality on the map are the right local-half checks.
• The plan note is honest that upload and dashboard symbolication were not done.

  

_{Sent by Cursor Automation: Bugsee code review}

[cursor]

Deep review — Task 13.5 local half (181a692)

Re-reviewed 181a692 against surrounding 13.1/13.4 code (hermes-sourcemaps.js finishAfterCompose, ensureHermesHooks, debugIds.ts / payload.ts, bugsee-xcode.sh, exceptions.test.ts payloadOf, startRun URI/launch-args). Both findings from 5cd1209 are fixed. The dashboard / debug-files upload half is still open by design; the plan checkbox stays unchecked.

This environment has no node_modules, so plugin unit tests were not executed here. At posting time GitHub CI on this HEAD was still running (lint, typecheck, unit, mutation pending).

Prior findings

P2 Expo project.exec migration — fully addressed. migrateFinishHookExec scopes the replace to the Bugsee bundleTask.doLast that contains "finish" after PRESERVE_REFUSAL, so a 13.4 hook is rewritten and a sibling project.exec is left alone. The new fixture starts from that stale line and asserts a second apply is a no-op. The lookup stays bundleTask.services.get(ExecOperations); that is the call the WOD_LX1 Gradle 9.4.1 release actually ran.

P2 iOS device vs Android map — fully addressed. describeReleaseProof now requires BUGSEE_COMPOSED_MAP on iOS device. Without it the suite is skipped and does not open index.android.bundle.map.

Findings

No new P0–P3. I looked for silent migration misses (brace matching vs the generated snippet, "finish" region, CRLF), ExecOperations cwd (all finish argv paths are already absolute), exc-map-id replacing the inject stub, awaitBundles / payloadOf vs the existing exception e2e, iOS afterAll ( clearBundles → clearIosBundles → terminateIosApp), and release steering (startRun still uses the Android URI / iOS launch args). None of those produced a credible failure on this HEAD.

Overall risk: Low

The production change is the Gradle 9 Project.exec replacement plus an upgrade rewrite for trees that already have the 13.4 hook. Android local half was device-proven on WOD_LX1 (debug_ids {"index.android.bundle":"3c110b23-2416-5bc6-acac-e35ac7e69aa6"} matching the composed map). iOS remains opt-in via BUGSEE_COMPOSED_MAP; that is documented, not a false failure.

Merge recommendation

Safe to merge once CI on 181a692 is green. Do not treat the 13.5 checkbox as done: upload and dashboard symbolication are still open.

Most important issues to fix

None on this HEAD.

What looks good

• Replacing Project.exec is the correct Gradle 9 fix; finish arguments are already absolute, so working-directory differences vs Project.exec do not matter here.
• The migration is deliberately narrow (^([ \t]*)project\.exec \{$ inside that doLast only).
• exc-map-id skips registerDebugIds, which would have replaced the inject stub and hidden the real UUID. Launch still goes through startRun (URI on Android, launch args on iOS), so a release binary can select the scenario even though e2e-scenario.json is baked in.
• expect(run.dev).toBe(false), airplane mode, nonce-scoped reason, and debug_id/debugId equality on the map are the right local-half checks.
• The plan note is honest that upload and dashboard symbolication were not done.

  

_{Sent by Cursor Automation: Bugsee code review}