feat(workflows): author parallel reviews and resolve workflow attention

[Igloczek]

Summary

Parallel reviewers are authored and inspected in the same workflow Graph. Run evidence separates accepted reports from settled executions, and server-owned attention groups native requests, human gates and recovery by run.

Code / Security / UX → Wait for all → recorded join
                                   → repeat / human gate / recovery
Native questions + approvals       → one attention run → existing thread controls

Refs #28 and parent #3. Layer 4/5: #30 → #31 → #34 → #35 → #36. Graph remains the editor/run representation; canonical inspector Routes, Definition and Steps support it without a separate Routes screen.

Current head edf703c3bb3949c807e29c119b4f13b39616a85a is based on approved #27 508f50cbc218dcf04db384216dbc69eb11a2cee2, on main 294aa94c427c2bd9690af172f07374bfc9877a9c. The initial nine-commit rebase was clean: 37 original owning files retain approved bytes, and Workflow.ts only inherits the approved persisted-action decoder fix. This head adds one exact gate-node expectation to the inherited legacy-cache regression, making the layer 39 paths. All product code is unchanged by that test adaptation. Fresh compatibility proof is complete. Independent R8 and primary integration review both approve this exact head across all 39 paths and all eleven criteria; no material findings remain. The independent reviewer reran the complete selected legacy-cache regression: one passed, seven filtered declarations excluded. Current-head CI passed, and GitHub confirms the draft is mergeable.

The previous head 8b6376be01ffe2d062f27cc8b12fbed3ff4defb4 received full independent R7 and primary approval for all eleven criteria, including the integrated local and authenticated remote walkthrough. Its complete physical evidence remains source-qualified reuse, with no new current-head pixel claim.

Evidence

• Before: the preceding authoring layer could not configure a complete parallel review or grouped attention flow. After: nine original images and the original rerun recording, with capture heads and limitations identified.
• Actual local and authenticated HTTPS workflows cover implementation, real Bun checks, Code/Security/UX, typed nested All/Any and Otherwise, one repeat and the separate Human limit, native threads/requests and explicit approval. Both 2/3 reports and 3/3 reports with one execution still running remain pending; only persisted settlement/trace advances the join. Early changes does not hide or cancel siblings.
• Actual changed-head recovery refuses stale approval, verifies the new head on rerun, retains old generation/history and resolves a later interrupted UX branch as unresolved. Reload/disconnect/restart preserves the entire Run and attention; mutations wait for fresh reconciliation. Exact deep links, read-only controls, narrow evidence and keyboard selection were exercised. A normal exact-page reopen recovered the subscription; uninterrupted automatic recovery is not claimed.
• Native queue evidence shows two Code requests plus one Security request in one attention run, oldest-first native navigation and independent resolution. Viewing and snoozing leave reports/gates/requests unchanged. Counts are distinct runs, with uncapped native totals and bounded progressive access to request IDs.
• Stable reviewer allocation prevents a removed Security identity being silently reassigned to UX. Retained predicates/bindings stay dangling until explicitly repaired; actual validation refuses invalid Save/Start without creating a run or launch. The accepted identity regression and legacy-numbered-ID proof remain attributed to their original executions.
• Current compatibility validation uses the complete compiled authenticated Attention/Decisions/Run principals, focused parallel/attention/run backend files and the inherited legacy-cache regression, with scoped prepared types. Three compiled authenticated principals, 21 owning backend cases and one real-core automatic-queue snooze case pass. Final prepared production/principal types pass; two narrowed-scope setup attempts are retained as failures before the correct actual scope was included. The inherited legacy-cache case initially failed at its run-only attention-link expectation. After the test-only canonical gate-node adaptation, the complete selected case passes, including all later raw-store/history/refusal/replay/restart/zero-launch assertions. All 55 assertion sites remain, every other test byte is preserved, and focused test-scope types/format/lint/diff checks pass. This fresh one-case run is separate from the prior 25 passes; seven filtered declarations are not new passes. The original failure and narrowed-type setup failures remain recorded. Prior 159 units/75 selected backend/10 authenticated cases remain separate historical aggregate evidence, including their recorded setup failures and corrected reruns.

The physical walkthrough uses deterministic external provider/forge replay with real authorization, core native lifecycle, services, SQL, workspaces and checks. Source-identical physical authoring used canonical import/edit/clone; fresh-from-empty authoring is separately covered by compiled real-client tests. No live provider/forge, hardware touch, second physical device, relay/tunnel, production bundle, benchmark or positive human-comprehension claim. Captures distinguish pre-settlement versus settled gates and stale versus later fresh reconnect state. The legacy-cache correction is inherited from #27; no new physical upgrade-client walkthrough is claimed.

Known broad generic PluginPage type/fixture and historical heap failures remain baseline debts, not green results. The bounded 100-run × 64-item attention snapshot lacks a worst-case benchmark; progressive request access is bounded at 32 IDs per attempt. Existing optional wording for more than 100 runs in one project and final widening-test coverage remain nonblocking limitations.

Merge Danger

Door: two-way. Authoring/attention UI is reversible; native responses, gate decisions and recovery use existing authoritative persisted commands.

Blast Radius: workflows. Parallel controls, report/settlement evidence, bounded attention reads/counts and generic host contributions. Native ordering and response protocols, destination permissions and separate core/plugin RPC handler layers remain intact. No new concurrency/budget, scheduler clock or client routing engine.

Implementation: Claude Opus 5.5 / Claude Code and OpenAI GPT-6.1-Sol / Codex. Orchestration, integration and primary client verification: GPT-6.1-Sol / Codex in T3 Code. Independent review: GPT-6-Astra / Codex with Gilfoyle; historical rounds also used Claude Opus 5.5 / Claude Code.

[Igloczek]

Current primary evidence for #28 at 8b6376be01ffe2d062f27cc8b12fbed3ff4defb4. Independent GPT-6-Astra/Gilfoyle R7 approves source and all eleven acceptance criteria. These are original captures; external provider/forge responses are declared replay while authorization, native commands, workspaces, checks and persistence are real. The before image is from reviewed lower layer #26; the ordinary deadline capture predates the allocator fix and its unchanged controls are identified separately. Remaining captures are from the current-head local/HTTPS acceptance pass. No live provider/forge, hardware-touch or production-build claim.

Before: parallel authoring unavailable in the reviewed lower layer #26.

After: actual installed same-skill reviewers, canonical policy and exact fractional deadlines (ordinary authoring capture at 948a6eb; these controls are unchanged in 8b6376b).

After: current remove/add allocation uses a new identity; an incomplete temporary reviewer blocks Save.

Three reports with only two settled executions remains pending.

One attention run contains all three native requests, including the later Code queue item.

The recorded Graph shows Repeat1/1 then At limit Human after exactly two implementation visits.

An outdated reviewed head refuses approval and offers the existing rerun action.

Restart/reopen preserves the interrupted review, earlier generation and eligible recovery.

Narrow Evidence sheet keeps current review and selected historical evidence distinct.

Original short rerun recording, including initial client reload before the new generation.

pr35-rerun-review-8b6376b.webm

Implementation: Claude Opus 5.5 / Claude Code and GPT-6.1-Sol / Codex. Primary verification and publication: GPT-6.1-Sol / Codex in T3 Code. Independent review: GPT-6-Astra / Codex with Gilfoyle.