PR #20347 maintainer takeover proof
Head: 83aa2ebdc1f55439a0090f2f94e8bc7ab25dd2da
Base: 1451e7cb23034d29774338608368a15d8845df36
Production SHA256: 344bdee2915ee22aa06d2d18cfd257307b651a470367ddbed8e542feba189332

Focused tests: original 11 added lifecycle cases rewritten as 4 explicit behavior cases in existing ClaudeAdapterV2.test.ts.
Exact-main negative control: 4 fail, 286 pass.
Final adapter suite: 290 pass. Adapter plus replay suite prior to behavior-neutral one-line cleanup: 434 pass.
Final scoped server typecheck exit 0; lint exit 0 (one existing unused layer warning); format exit 0.
Two independent whole-diff review rounds, Claude Opus 5.5 and Codex GPT-6-Sol: final round no findings. Round1 redundant condition simplified.

Native SDK 0.3.276 / Claude CLI 2.1.303 on headless Linux.
Production adapter, real SDK+CLI native tool execution, deterministic loopback Anthropic tool responses.
Background Agent: root terminal observed before child native tools released. Post-root Read accepted; post-root Write declined. Two approvals, no marker created, child wake detected.
Completed foreground Agent: native SendMessage resumed same task in second run; second root terminal observed before child tools released. Two post-root approvals, native Read accepted and Write declined; no marker created; child wake detected.

Integrated real web client on isolated worktree server with same deterministic model stub:
root run completed_at 2026-10-19T08:39:41.903Z
file-read request created 2026-10-19T08:39:51.484Z; client Approve; resolved accept
file-change request created 2026-10-19T08:40:01.639Z; client Decline; resolved decline
continuation run completed_at 2026-10-19T08:40:02.779Z
marker /tmp/claude-09/ui-bg-marker.txt absent
Screenshots show root completion text and child approval, not a fabricated interface.

Scope: latest settled run only, and only if it completed; current live query identity and running child ownership required. Earlier-run children after newer root settlement remain denied.
Not verified live: full-access callback fallback (CLI normally bypasses callback), no-agentID SDK variation, stop/restart rejection paths (focused tests cover), remote model behavior, macOS/Windows/Electron/iOS/Android/remote/tunnel.
Human authorization-lifecycle review remains required.