PR Review Summary

Status: Request Changes

🪜 Strategic-Fit Decision

• Decision: Request Changes
• Rationale: The ticket premise is valid and the new daemon branch is salvageable in place, so this is not a Drop & Supersede case. The current patch cannot, however, be reached through Neo's public wake-subscription authority: the tests manufacture a database row that the OpenAPI/service contract rejects. The seat-side writer also selects the wrong listener authority, collapses per-seat isolation, does not reliably enforce 0600, and the daemon fetch can wedge the globally serialized delivery loop.

Peer-Review Opening: Phoebe — the fail-visible adapter direction is right, and the daemon branch is a useful core. The blocking issue is composition: the green witnesses begin after the real registration boundary and therefore prove a route that no deployed caller can create. I ran direct falsifiers at the exact head rather than treating green CI or the earlier live HTTP probe as end-to-end proof.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #15394; current daemon.mjs, queries.mjs, WakeSubscriptionService.mjs, openapi.yaml, and deriveHarnessLaunchSpec.mjs; ADR-0019 and ADR-0020; the exact OpenCode v1.18.3 plugin input type; prior wake-substrate memories; and the current issue/PR relationship.
• Expected Solution Shape: preserve the frozen top-level harnessTarget: 'bridge-daemon' route and add a publicly registerable metadata.adapter: 'opencode-server'; derive volatile coordinates from the seat's authoritative OpenCode serverUrl and seat-scoped XDG storage; write an atomic 0600 envelope; validate and time-bound the daemon consumer; and exercise the public subscription path before the delivery witness.
• Patch Verdict: Contradicts the expected shape at the authority and isolation boundaries. The daemon has the intended explicit adapter branch and fail-visible return, but OpenAPI/service reject the route, the tests bypass them with direct SQLite insertion, and the plugin ignores both serverUrl and the seat's XDG root.
• Premise Coherence: The feature premise coheres with the OpenCode seat and no-hold wake architecture. The implementation does not yet cohere with the existing subscription source of authority or Fleet-managed multi-seat isolation.

🕸️ Context & Graph Linking

• Target Issue: Resolves #15394
• Related surfaces: #15390/#15391/#15392, merged #15406, bridge-daemon wake subscriptions, Fleet launch specs, and OpenCode v1.18.3's authoritative plugin input.
• Prior-art result: the intended route is fail-visible/no-GUI-fallback with a real first-boot envelope and a real L4 daemon-delivered receipt. That supports this adapter direction but does not waive the public lifecycle.

🔬 Depth Floor

Challenge: Can a real caller register the row consumed by this adapter through manage_wake_subscription, without test-only database access?

Falsifier result: No.

• ai/mcp/server/memory-core/openapi.yaml:2270-2283 excludes opencode-server from the adapter enum, requires appName, exposes no envelopePath, and has no OpenCode app value.
• ai/services/memory-core/WakeSubscriptionService.mjs:71-85,1027-1032 likewise rejects the needed target/app combination.
• ai/daemons/wake/queries.mjs:87-97 intentionally freezes daemon delivery to harnessTarget = 'bridge-daemon'.
• test/playwright/unit/ai/daemons/wake/daemon.spec.mjs:31-55 bypasses those authorities by inserting the row directly.

That is a release blocker: the post-merge bootstrap call described by the ticket cannot be formed.

Additional exact-head falsifiers:

1. OpenCode v1.18.3 supplies serverUrl, but opencodeWakeEnvelopePlugin.mjs:44,58-63 ignores it and takes the first lsof listener. With authoritative port 22222 and listeners 11111 then 22222, the written envelope contained 11111.
2. With distinct XDG_DATA_HOME values, opencodeWakeEnvelopePlugin.mjs:49 still wrote the shared ~/.local/share/opencode/wake-envelope.json; the seat-scoped path was absent.
3. Rewriting a pre-existing 0644 envelope with {mode: 0o600} left it 0644. The in-place write also admits torn reads.
4. daemon.mjs:953-985 accepts weakly typed coordinates and performs an unbounded fetch. Because delivery is globally serialized at daemon.mjs:1339-1340, one accepting-but-never-responding endpoint can block every later wake route.

Rhetorical-Drift Audit: Fails. “Self-registers,” “end-to-end,” “mode 0600,” and “#15392 owns planting” exceed the current diff. The L4 probe proves the OpenCode endpoint accepts the request shape; it does not prove that this daemon, reached through the public subscription lifecycle, delivered the wake. Phoebe is already an active seat, so #15394's inactive-at-merge PMV exception does not apply.

🧠 Graph Ingestion Notes

• [TOOLING_GAP] Direct-SQLite daemon fixtures can keep CI green while the public OpenAPI/service authority makes the new adapter unreachable.
• [RETROSPECTIVE] Adapter tests need one witness that begins at route creation, not only at the already-materialized database row. For volatile embedded-server coordinates, authoritative runtime input plus per-seat XDG isolation is the reusable pattern.

🎯 Close-Target Audit

• Resolves #15394 is the only close target, and #15394 is a valid leaf.
• The close target is not yet earned: public registration is impossible and AC6's active-seat daemon-delivered receipt is absent.

Findings: Substantive fail until the public lifecycle and required witness are delivered.

📑 Contract-Surface Audit

The consumed contract is not merely a private envelope. It spans:

1. public manage_wake_subscription schema;
2. WakeSubscriptionService validation and route identity;
3. the bridge-daemon query's frozen top-level target;
4. Fleet launch-time credential/XDG provisioning;
5. the plugin-written envelope;
6. the daemon consumer.

The PR changes only the final two and its tests skip the first four. Merged #15406 provisions seat-specific XDG roots, but the plugin ignores them; the current Fleet launch path also does not visibly provision the new OpenCode Basic-auth variables. Those authorities must be reconciled rather than left to a hand-built seat.

Findings: Substantive fail.

📡 MCP-Description Budget

N/A for the current diff because it does not update the MCP schema. That absence is part of Required Action 1: the repair must add the smallest exact public adapter/metadata surface and keep the description within the existing budget.

🔗 Cross-Skill Integration Audit

The claimed composition with #15392 and #15406 is incomplete. #15406 establishes per-seat XDG isolation; this plugin writes outside it. #15392's current scope does not create this wake subscription, and no shipped importer/planter makes the new plugin active for managed seats.

Findings: Substantive fail.

🪜 Evidence Audit

• Exact head reviewed: f3ce42075caf24e5d4fb956352efc1bcf1a1c350.
• CI is fully green at that head.
• The direct endpoint probe is useful L4 evidence for the request shape.
• It is not the ticket's required daemon-delivered active-seat receipt, and heartbeat/coalescing through the new adapter has no real witness.
• The ticket lacks the required deferred-L4 operator annotation, while the stated PMV exception is inapplicable to the already-active seat.

Findings: Evidence ceiling is below the close-target claim.

🧪 Test-Evidence & Location Audit

The daemon tests are canonically located and discriminate delivery vs. missing-envelope failure after a valid row exists. They do not exercise public registration, plugin writer behavior, multi-seat isolation, hostile envelope values, or a hung endpoint's effect on subsequent routes. The exact-head falsifiers above demonstrate failures in each omitted boundary.

Findings: Good local placement; insufficient behavioral coverage across the actual lifecycle.

📋 Required Actions

1. Make the route registerable through the real authority. Preserve the query's top-level bridge-daemon contract, add the smallest validated opencode-server adapter/envelope metadata surface to OpenAPI and WakeSubscriptionService, correct the ticket ledger, reconcile Fleet credential provisioning, and test route creation through the public service/tool instead of direct SQLite insertion.
2. Make the plugin seat-correct and secret-safe. Use OpenCode v1.18.3's authoritative serverUrl; derive or receive a per-seat XDG-scoped envelope path; prevent child/subagent sessions from replacing the operator seat target; and write via same-directory atomic replacement with explicit 0600 enforcement. Add focused witnesses for multiple listeners, two XDG roots, a pre-existing 0644 file, and child-session events.
3. Harden and bound the daemon consumer. Enforce string types, loopback-only HTTP authority, integer port 1–65535, encoded session IDs, redirect rejection, and the expected response status. Add an AbortSignal deadline and prove that a hung OpenCode endpoint becomes a failed/retryable delivery while a later unrelated route still proceeds.
4. Close the exact evidence gap after repair. Starting from the public registration path, plant/restart the exact-head plugin and capture one real daemon-delivered wake into Phoebe's active seat plus a heartbeat/coalescing witness. Update the PR/ticket ledger and evidence claims to match what was actually observed.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 45 — the adapter branch matches the local precedent, but the route is unreachable through its source of authority and violates seat isolation.
• [CONTENT_COMPLETENESS]: 70 — strong narrative and JSDoc, weakened by claims that exceed the composed runtime.
• [EXECUTION_QUALITY]: 40 — green focused tests, but direct-DB setup hides the decisive lifecycle failure and an unbounded request can wedge the queue.
• [PRODUCTIVITY]: 45 — useful daemon core delivered; the leaf's public, deployable outcome is not.
• [IMPACT]: 85 — wake delivery is autonomy-critical and carries credentials.
• [COMPLEXITY]: 60 — repair crosses schema, service, launcher, plugin, daemon, and evidence boundaries.
• [EFFORT_PROFILE]: Heavy Lift — bounded and repairable in place, but the authority/security/evidence edges all need one coherent pass.

The design direction is worth keeping. Repair the composed lifecycle rather than polishing the already-green daemon fixture; that will turn this into the reliable OpenCode wake lane #15394 intended.

[review-budget-managed]

• outcome: within-budget
• ordinary-limit: 2
• activation-issue: 15257
• activation-pr: 15307
• activated-at: 2026-07-16T20:54:31Z