Add cross-harness wake prompt-landing matrix

Context

Child of #10647. The 2026-05-03 regression showed that the team repeatedly accepted layer-local success as full-loop wake success. The clearest example is #10644: the bridge believed it delivered to Antigravity because osascript exited 0, but the payload landed in editor/file content instead of the agent prompt surface.

The wake substrate needs a validation matrix that says exactly what must be true before anyone can call wake delivery working again.

Duplicate Sweep Notes

Creation sweep performed as part of #10647:

• Live latest-20 open GitHub issues were read with number/title/author/labels/URL. Adjacent entries included #10644 (Antigravity-specific prompt-surface bug), #10645 (Codex bootstrap), #10633/#10627 (heartbeat/recovery dependencies), #10601 (parent recovery epic), and #10517 (routing semantics). None define a cross-harness prompt-landing acceptance matrix.
• Local resource search found #10440 noting the old trap: a bootstrap check passed even though full-loop wake delivery had not been validated. That is historical adjacency, not an open matrix ticket.
• ask_knowledge_base(type: 'ticket') found no equivalent ticket for prompt-landing validation.

The Problem

"Wake delivered" is currently ambiguous. It can mean any of:

• the A2A message was persisted;
• list_messages can read it;
• a wake subscription exists;
• a coalesced/raw event was emitted;
• bridge-daemon attempted an adapter;
• osascript exited 0;
• the correct app became frontmost;
• the payload actually reached the agent prompt input;
• the agent accepted/submitted the prompt without corrupting files or spawning a wrong session.

Only the last two prove the user-visible effect. The substrate repeatedly regressed because earlier layers were treated as sufficient proof.

Historical Memory Context

Relevant Memory Core anchors:

• summary_0763a9bf-1052-4a2f-99f3-a8e0e14f1671 — wake delivery previously required multiple fixes across permissions, tab focus, atomic paste, raw/digest envelope, and clone state.
• summary_a8592d87-f132-42c5-af6f-0e066fd5f428 — a raw/coalescing change broke another delivery path, showing that wire-format and adapter compatibility must be tested as a loop.
• summary_bf59d6c4-e250-44a2-b4b2-5bffae40ab5f — wake substrate reviews uncovered silent metadata stripping and led to mailbox protocol discipline.

The Architectural Reality

The validation surface spans:

• Memory Core A2A storage/listing (add_message, list_messages).
• WakeSubscriptionService bootstrap/active subscription state.
• Coalescing/raw wake event shape.
• Bridge daemon or native harness transport.
• Harness-specific prompt surface behavior:
  • Claude Desktop / Claude app tab or prompt field.
  • Antigravity IDE agent composer, not editor files.
  • Codex Desktop prompt/thread surface.

No single unit test can prove all of this for all harnesses. The matrix should combine deterministic unit/integration tests with explicit manual or semi-automated live validation evidence.

The Fix

Create a cross-harness wake validation matrix and wire it into the #10647 reactivation gate.

Minimum matrix columns:

1. message persisted;
2. unread/list state correct;
3. subscription bootstrap and metadata correct;
4. wake event emitted with expected envelope;
5. adapter selected correct app/session target;
6. prompt payload lands in the agent prompt/composer surface;
7. no editor/file content is modified;
8. no fresh session is spawned unless explicit sunset/unsubscribe permits it;
9. recipient can act on the prompt or a clear blocked signal is emitted;
10. evidence artifact captured in PR/comment/log.

Minimum rows:

• Claude Desktop / Claude app wake path.
• Antigravity IDE / Gemini wake path.
• Codex Desktop wake path.

Acceptance Criteria

• A repo doc or test-runbook defines the matrix columns above and names the evidence required for each cell.
• The matrix explicitly states that A2A storage success and bridge adapter success are not sufficient proof of prompt delivery.
• The matrix covers Claude Desktop, Antigravity IDE, and Codex Desktop as distinct harness rows.
• Antigravity row includes a negative assertion: the payload must not land in editor/file content (#10644 regression class).
• Fresh-session rows include a negative assertion: no fresh session may spawn unless explicit sunset/unsubscribe state authorizes it.
• At least one deterministic test or mock validates bridge adapter intent for each supported app name/strategy where feasible.
• Live/manual evidence requirements are documented for UI-only assertions that cannot be proven in headless unit tests.
• #10647 heartbeat reactivation gate references this matrix as required evidence before reset.

Out of Scope

• Fixing any individual harness row failure. Failures should create or link specific tickets such as #10644/#10645.
• Building a full native UI automation layer for every harness in this ticket.
• Replacing #10517 HarnessPresence/wakePolicy routing semantics.
• Reactivating heartbeat.

Avoided Traps

• Trap: bridge log says delivered, therefore done. Rejected. Prompt landing is the observable effect.
• Trap: require impossible full automation before documenting the gate. Rejected. Manual evidence is acceptable where harness UI automation is brittle, but the matrix must make the gap explicit.
• Trap: test only the currently broken harness. Rejected. Cross-harness regression is the pattern.
• Trap: collapse storage/listing into wake delivery. Rejected. A2A mailbox can work while wake delivery is broken.

Related

• Parent epic: #10647.
• Grandparent epic: #10601.
• Antigravity prompt-surface regression: #10644.
• Codex bootstrap regression: #10645.
• Routing model: #10517.
• Historical full-loop validation miss: #10440.

Origin Session ID: 89b259c3-27ec-4afb-baaf-fd39b55bffe1

Retrieval Hint: wake prompt landing matrix bridge delivered not enough Antigravity file write Codex Claude harness validation.