Pin Codex wake submit boundary in tests and matrix

Context

PR #13290 correctly pins the Codex wake boundary discovered while investigating #13287: prompt insertion / adapter command success is not the same as prompt submission and turn start.

During review, Ada and Grace both identified a close-target problem: #13290 used Resolves #13287, but #13287 is the live behavioral ticket for AC2, where actionable A2A message wakes may land in Codex without submitting. That live L4 behavior remains open. This ticket gives #13290 a truthful leaf close-target for the L2 work it fully delivers, while #13287 remains open as the behavioral investigation and live validation home.

The Problem

Agent PR body lint requires at least one Resolves #N line for agent-authored PRs. Without a narrower leaf, the only available close-target was #13287, which would prematurely close the operator-reported live-submit bug even though the PR explicitly declares the L4 residual.

The concrete substrate problem solved by #13290 is narrower and complete: the wake system lacked a precise, test-backed boundary between:

• a wake payload landing or being accepted by an adapter;
• the payload being submitted into Codex;
• a Codex turn actually starting without the operator pressing Enter.

The Architectural Reality

V-B-A source sweep before creation:

• #13287 is open and assigned to @neo-gpt; it remains the live behavioral tracking ticket for heartbeat-vs-A2A Codex submission.
• Active @neo-gpt wake subscription V-B-A via manage_wake_subscription({action:'list'}) on 2026-06-15 showed one active bridge-daemon route with metadata {appName:'Codex', tabShortcut:null, focusSeedKey:'r'} and no explicit adapter, so the live route defaults to macOS osascript, not an explicit codex-app-server route.
• ai/daemons/wake/daemon.mjs#deliverDigest() chooses the adapter from subscription metadata, not from event type; heartbeat and message events coalesced into the same subscription use the same adapter branch.
• ai/daemons/wake/daemon.mjs#buildWakeDigest() intentionally emits different payload shapes: pure heartbeat digests append the lifecycle directive, while message/task/permission digests omit it.
• ai/docs/wake-prompt-landing-matrix.md now needs an explicit submitted/turn-start row so adapter success cannot be reported as full wake success.

The Fix

Fully deliver the L2 boundary pin implemented by PR #13290:

• Add an explicit Prompt Submitted / Turn Starts criterion to ai/docs/wake-prompt-landing-matrix.md.
• Update the Codex app-server daemon JSDoc/log wording so send-message-v2 success is classified as dispatch/injection proof, not full submit/turn-start proof.
• Add wake-daemon unit coverage proving the Codex UI osascript route includes paste plus key code 36 for direct message wakes and mixed message+heartbeat digests.
• Keep #13287 open for live Codex L4 validation and root-cause work.

Contract Ledger

Decision Record impact

Aligned with ADR 0002 / wake-substrate standards. No ADR amendment: this ticket clarifies evidence taxonomy and adapter success semantics; it does not change the wake substrate architecture.

Acceptance Criteria

• The wake matrix contains an explicit prompt submitted / turn starts criterion distinct from prompt landing and adapter dispatch.
• Codex app-server send-message-v2 daemon wording does not overclaim full submission or turn-start proof.
• Codex UI bridge unit coverage proves direct A2A message wake digests include paste plus Enter intent.
• Codex UI bridge unit coverage proves mixed message+heartbeat digests include paste plus Enter intent and omit the heartbeat-only lane directive.
• PR #13290 uses this issue as its Resolves close-target and keeps #13287 open via Related / Refs semantics.

Out of Scope

• Proving live Codex Desktop prompt submission and turn start for actionable A2A message wakes. That remains #13287.
• Changing the active @neo-gpt wake subscription adapter.
• Reworking stale retry/read-state behavior from #13281 / PR #13285.
• Adding a new submit-capable Codex app-server command.

Avoided Traps

• Do not close #13287 from an L2 unit/doc boundary pin. #13287 is the live behavioral bug and still needs L4 evidence.
• Do not infer that codex debug app-server send-message-v2 submits a turn. Current evidence supports injection/dispatch only.
• Do not infer adapter divergence from heartbeat-vs-message symptoms unless live route metadata proves different adapters; current route metadata does not.

Related

• PR #13290: implements this ticket.
• #13287: live Codex A2A message wake submit / turn-start investigation.
• #13012: Agent Harness epic.
• #13067 / PR #13073: Codex app-server adapter route.
• #13281 / PR #13285: stale retry read-state behavior, adjacent but not duplicate.

Release classification: post-release Agent Harness reliability; not v13 release-blocking.

Duplicate sweep: live latest-open GitHub sweep at 2026-06-15T02:23:59Z checked the latest 20 open issues with number/title/author/labels/url. #13287 is related but broader and intentionally remains open; no equivalent L2 boundary-pin leaf was found. Live GitHub search for Codex wake boundary submit intent Enter matrix returned no open issues. A2A in-flight claim sweep at 2026-06-15T02:23:59Z checked the latest 30 messages across read states; no competing [lane-claim] / [lane-intent] for this split-ticket scope was present. KB ticket search for Codex wake submission boundary send-message-v2 key code 36 issue duplicate surfaced older Codex wake/app-server context (#10664, #10649, #10666, #10517, #10662) but no current equivalent leaf.

Origin Session ID: 019ec8a7-1f8e-75a3-b223-fe59cc444776

Retrieval Hint: "PR #13290 Codex wake submit boundary matrix send-message-v2 key code 36 #13287 close-target"