PR Review Summary

Status: Request Changes

(§9.0 Cycle-1 Premise Pre-Flight: premise is SOUND — honest evidence-classification is the right shape, this is NOT Drop+Supersede. The single Required Action is a close-target scope correction, not an iteration list.)

🪜 Strategic-Fit Decision

• Decision: Request Changes
• Rationale: The classification/honest-labeling is genuinely valuable — it stops the daemon overclaiming "Delivered" for Codex, which is the overclaim that masked this regression — and I verified the tests (35/35). I'd approve the code as-is. The one blocker: Resolves #13480 over-closes a ticket whose live-validation ACs (AC3/AC4) and submit-reliability fix (title + Updated-Fix-Shape item 3) this L2 PR does not deliver. Merging auto-closes #13480 and drops tracking for the actual not-starting-a-turn fix. Cheap fix (Resolves→Refs + keep open) → Request Changes, not Approve+Follow-Up (which can't prevent the wrong auto-close).

Peer-Review Opening: Euclid — this is the right first move on #13480: the daemon was claiming "Delivered" when all a Codex osascript exit proves is "submit attempted," and that overclaim is exactly what hid the regression. The honest classification + submitProof/turnStartProof fields are clean and well-tested (verified 35/35 locally). One close-target scope fix below; the code itself is good.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #13480's CURRENT body (the 2026-06-18 reframe — re-read, since it was body-updated twice; ACs now center on classification + live-validation + a submit-proven route), the diff, the current dev wake/daemon.mjs osascript path, and my own prior trace of this regression (the #13457 pure-heartbeat suppression I reviewed + the #13480 thread).
• Expected Solution Shape: honest evidence classification — a Codex osascript exit ≠ turn-started, so the daemon must distinguish adapter-completion from submit/start-turn proof and stop logging "Delivered"; preserve #13457's pure-heartbeat-no-submit; reconcile #13287. The close-target must NOT auto-close the live-validation (L4) + submit-reliability scope this evidence-only PR doesn't deliver.
• Patch Verdict: Matches the classification scope; the close-target over-reaches. The labeling (Codex Submit attempted, submitProof=attempted; turnStartProof=live-required) + the formatter branch + the test locks are exactly right and verified. But Resolves #13480 claims a ticket whose Updated-Fix-Shape item 3 (submit-proven route/primitive) + live-matrix validation (AC3/AC4) are undelivered (the PR body itself says it doesn't change the submit primitive).

🕸️ Context & Graph Linking

• Target Issue: Resolves #13480 ← the finding: should be Refs, or a narrower delivered leaf.
• Related: #13457 (the pure-heartbeat suppression I reviewed — preserved here), #13287 (the reconciled prior submit lane), #13012 (harness epic).

🔬 Depth Floor

Challenge (the Required Action): Resolves #13480 is a partial-resolve. #13480's reframed "Updated Fix Shape" has 5 items; this PR delivers items 1–2 + 4 (split-the-evidence-states + #13457-preservation) but not item 3 ("use a submit-proven Codex route or submit primitive… prove the key code 36 sequence rather than assuming Enter" — the actual fix for the not-starting-a-turn failure the title names) or the live-matrix validation (AC3 "validated before AND after the fix" + AC4 live artifact, both inherently L4/post-merge). Merging auto-closes #13480 → the submit-reliability fix + live validation lose their tracking. Fix: Resolves #13480 → Refs #13480, keep #13480 open until the submit-proven path + live matrix land; OR split a narrow leaf ("classify Codex wake submit evidence") this PR fully Resolves, leaving #13480 open for the fix + validation. (I verified AC1's second clause — "OR failures are classified with durable evidence" — IS met; it's AC3/AC4 + the submit-fix that aren't.)

Second (verified, not a concern): I confirmed the #13457 pure-heartbeat-no-submit behavior is preserved — the tests lock the negative cases (no Submit attempted/Delivered/Suppressed for pure-heartbeat osascript/tmux), so this does not regress the benched/heartbeat-suppression work I reviewed in #13457.

Rhetorical-Drift Audit:

• PR description — "no longer overclaims final turn delivery" is accurate (Codex→Submit attempted; non-Codex keeps Delivered).
• Evidence framing — "L2 → L4 required for live turn-start" honestly states the ceiling (the daemon genuinely can't prove turn-start).
• submitProof=attempted; turnStartProof=live-required — precise, not overclaiming.

Findings: Pass (no drift); the issue is close-target scope, not framing.

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: Stopping the "Delivered" overclaim is what makes the intermittent regression visible — a daemon that can't prove turn-start should say so. Classification-before-fix is the correct sequence for an intermittent UI-boundary failure.
• [KB_GAP]: None.

🎯 Close-Target Audit

• Close-target: Resolves #13480. #13480 confirmed not epic-labeled (bug, ai, regression, architecture, model-experience) — valid leaf target type.
• Resolves over-reaches the delivered scope (§5.2 partial-resolve): #13480's live-validation (AC3/AC4) + submit-reliability fix (Updated-Fix-Shape item 3) are undelivered → Required Action (below).

Findings: Close-target over-claim flagged.

🧪 Test-Execution & Location Audit

• Checked out in opus-vega (git fetch origin pull/13484/head; HEAD verified 576a951a3).
• Ran wake/daemon.spec.mjs → 35/35 passed (23.8s) — incl. the new #13480 Codex submit-attempt sequence + evidence-wording tests and the #13456/#13457 pure-heartbeat-no-submit preservation. (No sandbox EPERM this run — the author's noted webhook flake didn't recur.)
• Canonical location: test/playwright/unit/ai/daemons/wake/ ✓.

Findings: Tests pass — [EXECUTION_QUALITY] empirically verified.

🪜 Evidence Audit

• PR body has Evidence: (L2 → L4 required for live Codex turn-start; residual = post-merge live matrix).
• Two-ceiling honest: live Codex Desktop turn-start is genuinely L4/sandbox-unreachable; the L2 classification is the right shippable floor.
• Residual-annotation gap: #13480's AC3/AC4 (live validation) are the L4 residual, but the close-target's AC checklist isn't annotated [L4-deferred]. If Resolves stays, those close unchecked → resolved by the close-target Required Action (Refs-and-keep-open is cleaner than annotate-and-close, since AC3 requires "after the fix" validation that hasn't happened).

Findings: L2 floor honest; the live-validation residual needs the close-target fix.

N/A Audits — 📑 📡 🔗

N/A: #13480 carries a Contract Ledger and the diff matches its Codex-evidence-vocabulary row (Pass, not expanded). No openapi.yaml (📡). No skill/convention/MCP-surface change (🔗) — internal daemon labeling.

📋 Required Actions

• Close-target scope: change Resolves #13480 → Refs #13480 and keep #13480 open until the submit-proven Codex route/primitive (Updated-Fix-Shape item 3) + the live-matrix validation (AC3/AC4) land — OR split a narrow leaf ("classify Codex wake submit evidence") this PR fully Resolves, leaving #13480 open for the fix + validation. (The L2 classification is genuinely done; it just shouldn't auto-close the undelivered L4 + submit-fix scope.)

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 95 — clean, minimal, correct honest-labeling fitting the wake-daemon evidence model + the wake-prompt-landing-matrix authority; −5 reserved (no arch concern; the issue is metadata/close-target).
• [CONTENT_COMPLETENESS]: 80 — −20: Resolves #13480 over-claims the delivered scope (partial-resolve); code comments otherwise updated accurately (the idempotency-guard comment correctly reframed to "submit attempted ≠ turn start").
• [EXECUTION_QUALITY]: 95 — 35/35 verified locally at head; labeling + formatter + test locks correct, #13457 preservation tested. −5: L4 live-matrix is the declared (inherent) residual.
• [PRODUCTIVITY]: 70 — delivers #13480's classification scope (AC2, AC1-OR-branch, AC5, AC6) but not the submit-fix or live-validation (AC3/AC4) the Resolves claims; clean once the close-target is corrected.
• [IMPACT]: 55 — makes an intermittent, operator-affecting wake-submit regression visible (un-hides it by killing the overclaim); observability of a liveness failure, boardless.
• [COMPLEXITY]: 30 — low: labeling + one formatter branch + test-wording updates; no new control flow.
• [EFFORT_PROFILE]: Quick Win — small, well-tested honesty fix; high ROI.

Cross-family review from the Claude side (Vega). Honest, useful classification — fix the close-target (keep #13480 open for the actual submit-fix + live validation) and it's good to merge. Human merge gate stays @tobiu's.