Input from Claude Opus 4.7 (1M context) (Claude Code):

❋ # PR Review Summary

Status: Approved

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Approve
• Rationale: Surgical fix that closes the empirically-anchored Claude-side composer-focus gap (Cmd+3 selects the Code tab without always moving focus into the prompt). The opt-in focusSeedKey primitive is the right shape — keeps Antigravity on its already-idempotent Cmd+Shift+I route untouched while giving Claude a per-harness-defaulted Space seed before the destructive Cmd+A / Cmd+X clear sequence. Pairs cleanly with #10644 to make Claude wake delivery as robust as the Antigravity path. Bridge-daemon spec parallel-isolation fix is a substrate-correctness win on its own.

Peer-Review Opening: Strong execution on a tight composer-focus boundary. The opt-in-per-harness shape (focusSeedKey defaulting to 'space' only when appName === 'Claude') is the right architectural decision — operator can override or null-out per subscription, and future harnesses adopt the primitive only after their matrix-validated routes prove insufficient. The new ordering test (activate < tab < space < clear) directly verifies the load-bearing invariant that focus-seeding must precede the destructive prompt-clear. Ship it.

---

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #10660
• Related Graph Nodes: #10658 (composer-focus epic; this PR is the empirical-substrate-fix lane within it), #10647 (Wake Incident Safety Tree), #10649 (cross-harness prompt-landing matrix — this PR's shape is what the matrix would have surfaced as the Claude-side gap), #10644 / PR #10652 (companion Antigravity shortcut fix)

---

🔬 Depth Floor

Challenge (per §7.1):

The focusSeedKey keystroke branch is hardcoded:

if (focusSeedKey === 'space' || focusSeedKey === ' ') {
    osascriptArgs.push('-e', '      key code 49');
} else {
    osascriptArgs.push('-e', `      keystroke "${focusSeedKey}"`);
}

Two observations:

1. Untrusted-input surface: the else branch interpolates focusSeedKey directly into the osascript string. If a future operator (or a compromised subscription metadata writer) sets focusSeedKey to a value containing " or \\, the resulting osascript would either fail to parse OR execute injected AppleScript. Today the input is from operator-controlled WAKE_SUBSCRIPTION metadata, so the threat surface is narrow — but worth a one-line escape (e.g., reject anything that's not a single printable character at the validation layer in WakeSubscriptionService.subscribe()). Non-blocking; the empirical risk is low because the only writer is the operator.

2. Two-key allowlist via string-equality is fragile: 'space' and ' ' both map to key code 49. Any other value goes through keystroke. There's no test for the else branch, only for the Space branch. Future operator passing focusSeedKey: 'enter' would emit keystroke "enter" rather than key code 36. Probably out-of-scope for this PR's stated scope (Claude Space seed), but worth flagging as a follow-up if other focus seeds end up needed.

Per feedback_pr_review_iteration_calibration calibration: neither is a Cycle 2 blocker. Flagging as non-blocking observations.

Rhetorical-Drift Audit (per §7.4):

• PR description: "Claude-only Space focus seed" + "before the destructive Cmd+A / Cmd+X prompt-clear sequence" — mechanically accurate
• Anchor & Echo on the new bridge-daemon comment: "Claude Desktop's Cmd+3 selects the Code tab but does not always move focus into the prompt. Seed focus with Space before the destructive Cmd+A/Cmd+X clear path. Keep this opt-in per harness; Antigravity already owns an idempotent Cmd+Shift+I route." — accurate; cites the empirical anchor without embedding session-specific PIDs / timestamps; framework-portable
• No [RETROSPECTIVE] tag inflation
• Linked anchors: #10660 (close-target), #10658 (epic context), #10647 (parent), #10649 (matrix sibling) all resolve and frame correctly

Findings: Pass.

---

🧠 Graph Ingestion Notes

• [KB_GAP]: N/A
• [TOOLING_GAP]: N/A — but the bridge-daemon spec isolation fix (moving crypto.randomUUID() from describe-scope to beforeEach) closes a latent parallel-spec race that empirically reduced the spec to "single-worker-only" in default Playwright config. Worth flagging as a meta-discipline win — same parallel-spec-isolation pattern as the cleanup memory I saved earlier today.
• [RETROSPECTIVE]: This PR is the natural completion of the Claude-side wake delivery loop alongside #10644 (Antigravity Cmd+Shift+I) and #10645 (cache hydration). Together with the wake safety gate (#10648) and the protocol document (#10650), the substrate has full per-harness focus-and-deliver coverage. The opt-in focusSeedKey primitive is also future-proofing: when Codex Desktop comes online with its own composer-focus quirks, it can adopt a per-harness focus seed without re-architecting the bridge.

---

🛂 Provenance Audit

N/A — bug fix below the §7.3 threshold.

---

🎯 Close-Target Audit

• Close-targets identified: Resolves #10660
• #10660 verified not epic-labeled
• Syntax-exact: Resolves #10660 on its own line
• Related: #10658, Related: #10647, Related: #10649 are non-magic references, correctly used

Findings: Pass.

---

📡 MCP-Tool-Description Budget Audit

PR touches ai/mcp/server/memory-core/openapi.yaml — audit applies.

• Single-line addition: focusSeedKey: {type: string, nullable: true} — minimal, no description bloat
• No internal cross-refs in the schema description (none added)
• No architectural narrative — pure type-shape addition consistent with sibling fields (tabShortcut, tmuxSession)
• No external standard URL needed
• Well within 1024-char hard cap (single-line schema addition)

Findings: Pass.

---

🔌 Wire-Format Compatibility Audit

• harnessTargetMetadata now optionally includes focusSeedKey: string|null — backwards-compat: existing subscriptions without focusSeedKey continue to work (undefined → defaulted Claude-only)
• Existing Antigravity subscription path explicitly verified unchanged (expect(args.join(' ')).not.toContain('key code 49') in the existing Antigravity test)
• WakeSubscriptionService.subscribe() JSDoc updated to list focusSeedKey in the metadata enumeration
• No downstream consumer needs handler updates

Findings: Pass.

---

🔗 Cross-Skill Integration Audit

• No skill files / AGENTS.md / AGENTS_STARTUP.md changes — this is a substrate primitive, not a workflow convention
• WakeSubscriptionService schema change is documented at the schema layer + JSDoc layer + spec layer (round-trip metadata test)
• No new MCP tool surface
• Bridge-daemon comment block teaches the convention (Claude-only default, opt-in for future harnesses) — future agent reading the bridge-daemon code understands when to extend this primitive

Findings: All checks pass — no integration gaps.

---

🧪 Test-Execution Audit

• Branch checked out locally (review/10661-cycle1 tracking origin/codex/10660-claude-space-focus-seed)
• Targeted specs executed: npx playwright test -c test/playwright/playwright.config.unit.mjs test/playwright/unit/ai/scripts/bridge-daemon.spec.mjs test/playwright/unit/ai/mcp/server/memory-core/services/WakeSubscriptionService.spec.mjs test/playwright/unit/ai/mcp/server/memory-core/McpServerToolLimits.spec.mjs — 44 passed (8.7s)
• New ordering test for Claude focus seed verifies the load-bearing invariant: activateIndex < tabIndex < spaceIndex < clearIndex. The order is precisely what the Anchor & Echo block claims; if a future refactor moved the seed AFTER the clear, the test would fail loudly.
• Existing Antigravity test gained negative assertion (not.toContain('key code 49')) — confirms Space seed is Claude-only, doesn't leak into the Antigravity path

Findings: Tests pass; new behavior empirically verified end-to-end including the cross-harness negative case.

---

📋 Required Actions

No required actions — eligible for human merge.

Suggested follow-ups (non-blocking, do NOT delay merge):

1. One-line input escape at the validation layer. WakeSubscriptionService.subscribe() could reject focusSeedKey values that aren't single printable characters (or 'space', ' '). The osascript-injection threat surface is narrow today (only operators write subscription metadata), but a one-line allowlist at the schema-validation layer is cheap insurance. Worth a small follow-up PR or sub-issue under #10658.

2. Coverage for the keystroke "${focusSeedKey}" else branch. The Space branch is tested directly; the else branch is uncovered. If future harnesses adopt non-Space seeds (e.g., Codex Desktop), the test needs to verify the else path emits keystroke correctly. Could land as part of #10649 matrix execution evidence or as a small follow-up.

Both are out-of-scope for this PR's stated #10660 scope (Claude-only Space seed). The PR closes its named target cleanly.

---

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 95 — 5 points deducted for the untested keystroke "${focusSeedKey}" else branch + lack of a one-line input-escape at the validation layer. Substrate placement, opt-in-per-harness shape, and the focus-before-destructive-clear ordering are all correct.
• [CONTENT_COMPLETENESS]: 95 — 5 points deducted because the Anchor & Echo doesn't explicitly explain the key code 49 vs keystroke branching rationale. The opt-in-per-harness rationale, the Cmd+3-doesn't-focus claim, and the antigravity-already-idempotent contrast are all clear.
• [EXECUTION_QUALITY]: 95 — 5 points deducted because the else branch is empirically untested. The Space path is verified end-to-end including order assertion, and the Antigravity negative case is verified — solid coverage on the named target.
• [PRODUCTIVITY]: 100 — I actively considered: (a) does the PR close #10660's stated scope, (b) does it leave the bridge-daemon parallel-spec race latent, (c) does it stop short of the related #10658 epic scope. None apply — closes target cleanly, fixes adjacent spec discipline, leaves epic-level scope for follow-ups appropriately.
• [IMPACT]: 80 — Major: closes the empirically-anchored Claude composer-focus gap that #10658 audit identified. Pairs with #10644/#10645/#10650 to complete the per-harness wake-delivery substrate.
• [COMPLEXITY]: 35 — Low-to-moderate: 6 files (3 source + 3 spec). Schema/service edits are minimal one-liners; bridge-daemon edit is ~10 lines + one Anchor & Echo block; spec isolation fix is a small but architecturally-meaningful refactor.
• [EFFORT_PROFILE]: Quick Win — High operational ROI on a focused per-harness primitive that future harnesses can adopt without re-architecting the bridge.

---

Approving. Tobi to execute the squash-merge per pull-request §6 Human-Only invariant. With #10661 merged, the per-harness wake delivery substrate completes the focus-and-deliver loop across Claude / Antigravity / Codex per the matrix the #10650 protocol gates on.

— Opus