J2 operator journey: cockpit-started agent reaches GP activity

Context

The v13.2 ROADMAP gate names the product proof directly: the operator starts an agent from the cockpit UI instead of a terminal, and the resulting agent work becomes visible in the cockpit. Epic #14781 makes that a named integration journey (J2). Ticket #14782 makes the same path the F1 release-gate falsifier: the operator-run artifact is evidence for "measured, not claimed."

This ticket is the J2 leaf. It is deliberately a testing/integration contract, not a Fleet Manager implementation lane: Vega owns the Fleet Manager surface work, and the existing Fleet Manager tickets remain the substrate this journey consumes.

The Problem

The roadmap can still fail even when every component leaf is locally green. If the cockpit button, control-plane actuator, agent wake, Golden Path pickup, and cockpit activity readback are never exercised as one journey, the release can claim "download and steer" while the product is still a set of islands.

J2 is the repeatable seam proof. It must fail loudly on a crossed boundary defect instead of letting the team weaken the test or replace real seams with mocks.

The Architectural Reality

• Parent epic #14781 defines J2 as: cockpit -> start an agent from UI -> agent wakes -> picks a GP item -> PR/activity appears in cockpit activity adapter.
• Measurement sibling #14782 defines F1 as the operator-run artifact with zero terminal commands between cockpit-open and agent-first-action.
• ADR-0026 governs the lifecycle-write / read-observe exposure seam. The daemon-core restart actuator leaf #14760 is closed; this journey consumes the control-plane shape and must not put lifecycle-write authority onto a client bridge.
• Fleet activity sources already exist as closed adapter leaves #14572 and #14573. The remaining ActivityStream live-binding work stays on #14606; this ticket reads that surface, it does not implement Fleet Manager internals.
• Deterministic tour/replay support is tracked by #14640. Until that substrate is ready, the tour segment is a dependency gate, not local one-off glue.

Structure-map gate: ran npm run --silent ai:structure-map -- --files --loc; relevant owning folders include apps/agentos, ai/daemons/orchestrator/control-plane, ai/services/fleet, ai/services/graph, and test/playwright.

The Fix

Add the J2 journey proof as one coherent integration leaf:

1. A whitebox e2e opens the real cockpit UI and starts an agent through the UI/control-plane path.
2. The journey observes the agent wake and pick a Golden Path item through existing Agent OS state, not a test-only shortcut.
3. The cockpit activity surface shows the resulting activity through the bounded activity adapters/stream.
4. The same scenario yields an operator-run artifact for #14782 F1 and a deterministic tour segment once #14640 is ready.

Contract Ledger Matrix

Decision Record impact

aligned-with ADR-0026. The journey consumes the existing lifecycle-write / read-observe exposure seam and must not amend it or bypass it.

Acceptance Criteria

• A custom Playwright whitebox e2e opens the real cockpit UI and starts an agent from the UI, with no terminal commands between cockpit-open and first agent action.
• The journey crosses real seams: cockpit UI -> control-plane actuator -> agent wake/orchestrator -> Golden Path item selection -> cockpit activity adapter/readback.
• No crossed boundary is mocked in the journey proof. Doubles remain allowed only in lower-level unit specs; a J2 failure on a seam is a seam bug.
• The cockpit readback assertion uses the existing bounded activity adapter/stream path; missing/stale activity renders as a visible degraded state, never fake quietness.
• The run emits or captures the #14782 F1 operator-run artifact with provenance suitable for release notes.
• The tour segment is either implemented through #14640's deterministic runner or remains explicitly blocked by #14640; no local one-off replay glue.
• Test evidence uses Neo's custom Playwright configs, not default npx playwright test.

Out of Scope

• Fleet grid, health bar, ActivityStream component implementation, and other Fleet Manager UX work.
• Dock/QT layout work.
• Launch-playbook publication or public announcement sequencing.
• Control-plane authority implementation beyond consuming the already-shaped seam.
• Golden Path v2 feature work outside the minimum journey observation required here.

Avoided Traps

• Do not turn the proof into a terminal-driven script; that would falsify the release gate.
• Do not implement Fleet Manager internals here; consume existing tickets and block honestly when they are not ready.
• Do not weaken the e2e if a crossed seam fails. File or route the seam bug.
• Do not create tour-specific one-off code while #14640 owns deterministic tour execution.

Related

Parent: #14781. Measurement sibling: #14782. Consumes/depends on #14760, #14572, #14573, #14606, and #14640. Roadmap anchor: v13.2 "operator starts an agent from the cockpit UI instead of a terminal."

Live duplicate sweeps before drafting: GitHub searches for "operator journey", "start an agent from the UI", "control-plane actuator", "PoC falsifier", "agent wakes", and "activity adapter" found #14781/#14782 and related substrate tickets, but no equivalent J2 journey leaf. KB ticket sweep found no explicit #14781/J2 ticket. Memory Core sweep found historical FM/control-plane context, not a duplicate J2 leaf.

Origin Session ID: 6ab85930-3c14-4b18-b3b3-97989d1e75c6 Retrieval Hint: "J2 operator journey cockpit starts agent from UI control-plane agent wakes Golden Path activity adapter PoC falsifier"