Author parallel reviews and resolve workflow attention #28

Parent

#3

What to build

Complete parallel review as one authoring/execution/attention feature: configure distinct frozen-head reviewers, see exactly why Wait for all is pending, and resolve gates, stale reviews, and native requests from an actionable run-grouped attention view.

Acceptance criteria

• The same editor supports independently labeled reviewer branches with stable IDs, actual skills, focus/instructions, provider/model, input/report/deadlines, and canonical review permission policy. Same-skill branches remain distinct after reorder/reopen. Show All N reviewers required and Wait for all, the committed PR reference/frozen input, and backend-prepared isolated reviewer worktrees. No new budgets/concurrency setting.
• Join predicates use published aggregate/branch fields and existing typed route controls. Label changes preserve references; branch removal exposes dangling predicates. Backend validation blocks invalid fork/join pairing, unsupported reporting policy, and invalid/unverifiable frozen input according to its contract.
• The run view uses the complete persisted required branch set, not a compact recent-attempt tail. Show each branch/generation, focus, deadline, thread, report/execution phase, frozen head, workspace evidence when available, and stop reason, with bounded complete history and textual/keyboard-accessible statuses.
• Keep report count and settled-branch count separate: 2/3 reports and 3/3 with a running execution both remain pending. Only persisted join results enable the next route. Early changes never imply sibling cancellation; failed/missing/timed-out/canceled/stale/unverifiable branches explain the aggregate. Do not mix generations.
• Stale-review rerun uses the existing backend Retry and authoritative recovery target to verify a new head and create a new review generation. Gate actions retain the displayed revision/frozen reference, idempotency, and stale-result handling from the shared run controls; no implicit review or auto-merge.
• Own the narrow attention read model and generic summary/count contribution: stable server-owned reason kinds, newest-first paging, total distinct unresolved runs, per-attempt pending native IDs, and exact run/attempt/gate page-state links. Ordinary activity is progress, not a user decision. Never parse reasons from prose or mistake a capped list length for the total.
• Show every native request, including multiple branches and multiple requests per thread, using existing host inspection/lifecycle data. Generic navigation opens existing native controls. A later request shows its kind/queue position and explains that earlier requests in that same native queue are answered first; native ordering/response protocol remains unchanged. Refresh when a request is no longer pending; each item clears only on its own underlying resolution.
• Viewing/marking read/snoozing changes presentation only, never reports, approvals, native requests, gates, or server-owned workflow attention. Preserve native attention priorities. Share one attention controller per connected environment and release it with the last consumer; no all-transcript count scans or perpetual progress animation.
• Keyboard/touch controls, semantic labels, textual counts/reasons/direction, bounded views, and disconnect/reconciliation/capability gating apply to authoring, evidence, attention, and actions. Authenticated remote/deep links retain the exact environment/project; reload/restart preserves attention and trace.
• Focused real-backend tests cover same-skill branches, more than five reviewers, 2/3 and reported-but-running states, early changes, failed/missing/timed-out branches, backend cancellation, changed/unverifiable heads, rerun generations, native request queue navigation/resolution, read/snooze isolation, large attention totals, and subscription cleanup.
• Perform the parent acceptance walkthrough locally and again over an authenticated remote connection across the integrated authoring/run/attention surfaces, including implementation, checks, code/security/UX reviewers, nested conditions/Otherwise, Repeat once, At limit human review, native threads, reconnect/restart, and an unresolved branch. Obtain browser authorization for one real-client pass; use deterministic external replay and real authorization/persistence/services. Add concise workflow user guidance and keep fixes localized; larger uncovered gaps become separate blockers rather than speculative scope.

Blocked by

• Author checks, typed decisions, and bounded rework #27 — Author checks, typed decisions, and bounded rework.

Ticket specification reviewed and agreed by Codex and Claude Opus 5.5.