[Fix] Flaky mocked e2e subtasks test when parent resume is served the child fixture

[zoomote]

Opened on behalf of Elliott de Launay. Follow up by mentioning @zoomote, in the web UI, or in Discord.

Related GitHub Issue

Closes: #1001

Description

The mocked e2e test Roo Code Subtasks > child completing on its first response returns to parent flaked in CI (run 30098606515, merge queue for the unrelated PR #996; a re-run of the same SHA passed): the parent completed as "Fast child completed" instead of "Fast parent resumed" because aimock served the child fixture for the parent's resume request.

Two compounding fixture weaknesses in apps/vscode-e2e/src/fixtures/subtasks.ts, both fixed here:

1. Child fixtures matched parent-resume turns. The fast-child and SUBTASK_CHILD fixtures matched a bare marker that the corresponding parent prompt embeds verbatim. When the parent's tool result is serialized as a role: "tool" message, the original parent prompt (marker included) becomes the last user message and the child fixture — registered earlier — wins. Both fixtures are now predicates that keep aimock's last-user-message scoping but exclude requests containing the parent marker, mirroring the [Flaky Test] e2e-mock: subtasks cross-profile fixture collision + task identity prompt timeout #561 fix for the cross-profile fixtures.
2. Parent-resume fixtures could miss. They guarded on the literal new_task tool-call id, which validateAndFixToolResultIds can rewrite when reopenParentFromDelegation rebuilds the parent's history. The regular, fast, and interrupt parent-resume fixtures now match on the injected subtask-result content (Subtask <id> completed.\n\nResult:\n<summary> / the child result text) instead of the fragile id.

What reviewers should pay attention to: the new lastUserMessageContains helper deliberately replicates aimock's userMessage last-message scoping inside a predicate, so multi-turn child flows (ask_followup → completion) behave exactly as before.

Test Procedure

• TEST_FILE=subtasks.test xvfb-run -a pnpm --filter @roo-code/vscode-e2e test:ci:mock — 9/9 subtask e2e tests pass, including the previously failing one.
• pnpm test (full unit suite) — 6851 passed.
• pnpm --filter @roo-code/vscode-e2e check-types and lint — clean.

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved GitHub Issue (see "Related GitHub Issue" above).
• Scope: My changes are focused on the linked issue (one major feature/fix per PR).
• Self-Review: I have performed a thorough self-review of my code.
• Testing: New and/or updated tests have been added to cover my changes (if applicable).
• Documentation Impact: I have considered if my changes require documentation updates (see "Documentation Updates" section below).
• Contribution Guidelines: I have read and agree to the Contributor Guidelines.

Screenshots / Videos

N/A — test-infrastructure-only change, no UI surface.

Documentation Updates

• No documentation updates are required.

Additional Notes

Test-fixture-only change; no product behavior is affected.

Get in Touch

edelauna

[zoomote]

2 issues outstanding. See task

• apps/vscode-e2e/src/fixtures/subtasks.ts:158-160 — The regular parent turn-1 fixture (userMessage: /SUBTASK_PARENT_MARKER/) is the only parent turn-1 fixture without sequenceIndex: 0. In the role: "tool" resume serialization that caused this flake class, the last user message is the original parent prompt, so this fixture matches the parent's resume turn and re-serves new_task before the guarded child fixture (line 178) and parent-resume fixture (line 199) are ever evaluated — aimock serves the first content match in registration order. The new guards at 178/199 are therefore unreachable in exactly the path that motivated them; add sequenceIndex: 0 here for parity with the fast/interrupt/api-hang/abandon/xprofile parent fixtures.
• apps/vscode-e2e/src/fixtures/subtasks.ts:145 — The fast parent-resume guard [SUBTASK_FAST_PARENT_MARKER, "Fast child completed"] can also match the parent's initial request, because SUBTASK_FAST_PARENT_PROMPT embeds the child prompt verbatim and the child prompt itself contains Fast child completed (same for the api-hang guard [SUBTASK_API_HANG_PARENT_MARKER, SUBTASK_API_HANG_CHILD_RESULT] at line 271, where this PR removed the tool-call-id condition). Correctness now depends on the earlier sequenceIndex: 0 fixture winning turn 1; on a retry of the parent's first call (turn-1 fixture consumed, no tool result yet) these fixtures would serve the parent-completion response prematurely. The SUBTASK_RESULT_INJECTION guard used for the regular/interrupt resumes cannot match turn 1 — consider using it here too.

_{Reviewed dcae4b7}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!