Bind AgentOS Neural Link E2Es to the current page session

Summary

AgentOS Neural Link E2Es must bind to the App Worker session owned by the Playwright page under test, not to an older same-name agentos session left in the shared bridge/runtime.

This is the follow-up from the post-#14847/#14855 E2E review run. The dashboard source shape was fixed separately; the remaining failure is the E2E session-binding/test-contract layer.

Problem

A focused ordering run with Cockpit.spec.mjs before the Neural Link AgentOS specs exposed stale-session behavior:

./node_modules/.bin/playwright test -c test/playwright/playwright.config.e2e.mjs \
  test/playwright/e2e/agentos/Cockpit.spec.mjs \
  test/playwright/e2e/agentos/FleetCockpitLifecycleNL.spec.mjs \
  test/playwright/e2e/neural-link/WindowOps.spec.mjs \
  --workers=1

Observed symptoms from the review evidence:

• FleetCockpitLifecycleNL.spec.mjs can bind via raw waitForSession('agentos'), which is app-name lookup rather than current-page identity lookup.
• The selected session can be an old AgentOS session whose stores/windows do not reflect the page just opened by the test.
• WindowOps.spec.mjs still failed after the AgentOS dashboard source-shape PR because the test contract did not prove that the selected window topology belonged to the bound AgentOS page/session.

Architectural Reality

Scope

Fix the AgentOS Neural Link E2E session binding and test assertions so these specs are stable when another AgentOS page/session already exists in the bridge.

Expected shape:

• Refactor AgentOS Neural Link E2Es away from raw app-name waitForSession('agentos') binding.
• Use the page-bound neuralLink.connectToApp('AgentOS') path or an equivalent exact current-page App Worker id helper.
• Fix WindowOps.spec.mjs so source-window discovery validates the RuntimeService topology for the bound app worker.
• Harden the fixture/service only if local evidence shows the current page worker id path is still insufficient.

Acceptance Criteria

• FleetCockpitLifecycleNL.spec.mjs does not use raw waitForSession('agentos') app-name binding.
• WindowOps.spec.mjs verifies source windows using the actual topology fields for the bound AgentOS app worker.
• This focused ordering passes:

./node_modules/.bin/playwright test -c test/playwright/playwright.config.e2e.mjs \
  test/playwright/e2e/agentos/Cockpit.spec.mjs \
  test/playwright/e2e/agentos/FleetCockpitLifecycleNL.spec.mjs \
  test/playwright/e2e/neural-link/WindowOps.spec.mjs \
  --workers=1

• The fix does not depend on manually killing unrelated long-lived bridge sessions as normal test setup.
• If local port isolation is required for deterministic validation, the PR documents the exact supported command/config and why.

Out of Scope

• LivePreview Learn-route popout failure (#14856).
• Grid E2E failures (#14852, #14853, #14854).
• Reworking the AgentOS dashboard source shape already handled by #14847 / #14855.
• Broad bridge protocol redesign unless a minimal fixture hardening is proven necessary.

Avoided Traps

• Do not make test order the fix; the ordering is the regression probe.
• Do not accept app-name fallback as proof of current-page binding.
• Do not delete or skip the failing specs.
• Do not convert this to black-box DOM-only coverage; this is specifically a Neural Link whitebox E2E contract.

Duplicate Sweep

Checked immediately before creation on 2026-07-05T17:19:02Z:

• Latest 20 open issues via gh issue list --state open --limit 20.
• Targeted search for waitForSession('agentos'), oldest same-name session, and FleetCockpitLifecycleNL: no open duplicate.
• A2A mailbox latest lifecycle messages: no peer lane claims this session-binding follow-up; Grace owns #14857, unrelated dock producer work.
• Knowledge Base query surfaced archived #12897 as prior art for the same failure mode, not an active open duplicate.

Related: #14855, #14847, #12897.