Goal

Clean up the E2E test debt found during the v2026.5.11 release gate so future release checks distinguish real regressions from stale test assumptions.

Scope

In scope:

• Update prompt-mention coverage so it searches files that exist in the active test workspace instead of assuming the repo checkout is the active project.
• Fix session-composer-dock tests that use invalid question seed data or hit nested-scroll/streaming boundaries instead of the intended user contract.
• Fix the release-notes zh setup test so LANGUAGE_KEY is available inside the browser addInitScript context.
• Update or retire stale session header-menu tests now that sidebar rename/delete coverage owns the current user path.
• Update the home model-chip width assertion if the intrinsic chip width is the intended design.

Out of scope:

• Do not change product behavior unless a test cleanup proves a real regression.
• Do not fold this into the v2026.5.11 release chore.
• Do not broaden into a full E2E suite rewrite.

Relevant files or context

Likely files:

• packages/app/e2e/prompt/prompt-mention.spec.ts
• packages/app/e2e/session/session-composer-dock.spec.ts
• packages/app/e2e/release-notes/release-notes-toast.spec.ts
• packages/app/e2e/session/session.spec.ts
• packages/app/e2e/app/home.spec.ts

Release-gate findings:

• prompt-mention: the failing test searched for packages/app/package.json, but the browser was pointed at /Users/yuhan/PawWork; the smaller active-project @mention smoke in prompt-slash-open.spec.ts passed.
• session-composer-dock: one failure hit nested scroll / ongoing rendering behavior rather than a stable long-session user path; another used a one-option question seed rejected by the schema.
• release-notes: the zh test referenced LANGUAGE_KEY inside browser context without passing it into addInitScript.
• session.spec: header-menu rename/archive/delete paths are stale after the current sidebar IA; sidebar rename smoke passes.
• home.spec: the model chip now uses intrinsic sizing instead of the old fixed 176px assertion.
• home.spec: the root route smoke still expects session-new-home, but / currently renders the no-project empty state when no project is seeded. Triage whether this is fixture setup debt, intended IA drift, or a real regression before changing product behavior.
• Windows advisory non-E2E note: unit-windows-opencode-session timed out in cancel records MessageAbortedError on interrupted process during PR test: replace hash scroll source snapshot #532 validation; unit-windows-app passed, so this did not block [Task] Replace use-session-hash-scroll source snapshot with behavior test #528 but should be tracked with release-gate debt if it recurs.
• Windows advisory opencode workspace-adaptor flake: packages/opencode/test/plugin/workspace-adaptor.test.ts:458 (plugin.workspace > cold-start Workspace.get retries sync after owner bootstrap for remote adaptors) failed on Windows after PR feat(ui): render release notes toast with markdown and visible scroll #538 (608068369). Recent windows-advisory history showed 6 failures in 10 runs across unrelated commits, so treat it as timing-sensitive advisory debt rather than a feat(ui): render release notes toast with markdown and visible scroll #538 regression; stabilize retries/timeouts or skip on Windows in a scoped follow-up.

Verification

• Run each updated targeted E2E locally with --workers=1.
• Confirm at least one positive @mention file suggestion path uses files created inside the active test project.
• Confirm the composer-dock tests use valid question data and test a stable user-visible scroll/dock contract.
• Confirm release-notes zh test fails before the setup fix and passes after.
• Confirm no product copy or runtime behavior changes unless a real regression is found and documented.

Execution mode

Agent can implement directly after v2026.5.11 release is complete.