P0: R7 verify integrated Dev MCP reliability stack on canonical main #52

Owner scope amendment — macOS only (2026-09-11)

依 Owner 最新裁決，本工作線相關功能設計、實作、交付與驗收 只處理 macOS。本裁決取代下文較早的跨平台／三平台要求。

• Windows、Linux 支援、移植、修復與平台驗證不在本次工作範圍；不再以三平台全綠作為本工作線的 acceptance criterion。
• 驗收保留 macOS 上必要的 source tests、獨立 package/install、真實呼叫入口、operation／副作用／讀回及適用的恢復與終態證據；ownership、lease、CAS、安全與獨立審查要求保持有效。
• 下方歷史測試、PR、平台結果與問題紀錄保留作為歷史證據，不代表目前仍有跨平台交付承諾。後續設計及驗收計畫以此 macOS-only 範圍為準。
• 本次更新僅修正 Issue 與規劃範圍；實作、部署與重啟仍依 Owner 的暫停指令保持暫停。
• 此文字更新不會自行修改 GitHub CI 設定或 branch protection；既有檢查設定如與新範圍不符，須另行明確調整，不得假稱已通過或繞過保護。

Status

READY_NOW — verification/evidence gate only.

Priority

P0 — this is the last source/integration verification gate before the remaining live runtime and multi-session acceptance work can be trusted.

Fresh evidence boundary — 2026-09-04

Canonical James3014/devspace main:

• commit: 296a3066fc620169f23beeeb5c9c41d9c9370d75
• tree: 4d47c8e5e58b8302c076c1b97c990c4005b11843
• latest merge: PR test(cli): stop mock agy from truncating its own stdout #45 (test(cli): stop mock agy from truncating its own stdout)

The current integrated line includes the accepted R4-R6 reliability stack:

• fix(mcp): make cutovers crash-consistent across sessions #40 / Issue P0: Bound MCP session lifecycle and make Dev MCP cutovers multi-session safe #31 crash-consistent cutover source
• fix(agents): bind execution idle policy per turn #41 / Issue P1: Align idleTimeoutMs tool contract with enforced runtime semantics #32 execution-idle policy
• fix(workspaces): isolate shared conversation mutations #42 / Issue P1: Isolate mutating ChatGPT conversations and expose cross-session capacity contention #33 cross-conversation mutation isolation/capacity
• fix(mcp): complete bounded cutover control plane #43 bounded cutover control plane
• test(agents): make capacity continuation rejection deterministic #44 continuation-capacity verifier determinism repair
• test(cli): stop mock agy from truncating its own stdout #45 mock Agy stdout-flush verifier determinism repair

Individual PR verification is not enough to establish the exact current-main integrated claim.

Goal

Produce one revision-bound integrated verification receipt for the exact current canonical main before further live reliability closure.

Required verification

Use a fresh clean clone/managed worktree from the then-current canonical main; rebind exact HEAD before running.

At minimum verify the integrated production seams for:

• capability manifest and promotion-lineage preservation;
• durable cutover state, controller, HTTP/MCP control plane, restart/reconcile-only behavior;
• MCP transport/session lifecycle and cutover admission;
• execution idle policy and provider activity capability;
• cross-conversation shared-checkout mutation fencing;
• global local-agent capacity diagnostics and continuation admission;
• local-agent execution contracts and durable session/reconciliation;
• production server.ts registration/schema consistency;
• Codex Goal / trust-bootstrap regressions materially affected by the integrated stack;
• CLI/Agy regressions affected by PR test(cli): stop mock agy from truncating its own stdout #45.

Required baseline commands/suites should include the repository's exact focused scripts plus:

• npm run typecheck
• npm run build
• git diff --check

Where concurrency/timing regressions were previously flaky, run the relevant focused witness repeatedly enough to distinguish deterministic product failure from verifier-fixture races. Do not weaken assertions to obtain green.

Negative controls

Must retain evidence that:

1. stale/diverged promotion or capability removal fails closed;
2. same-old-server or wrong source/build/capability cutover finish fails closed;
3. uncertain restart does not authorize a second restart or blind replay;
4. provider without trustworthy activity cannot gain a terminating idle policy silently;
5. two active ChatGPT conversations on one physical checkout cannot both perform conflicting consequential mutation;
6. local slot exhaustion is not mislabeled as provider rate-limit evidence;
7. worker/transport error never silently becomes VERIFIED/ACCEPTED.

Failure handling

If this gate exposes a real product defect, keep this Issue open while the defect is repaired when the repair is required to satisfy this exact integrated verification contract. If fresh evidence proves a distinct gap outside this contract, reconcile it into an existing exact Issue or create a focused follow-up before mutation.

Do not start chat-only repairs.

Dependencies / frontier

• R4-R6 source Issues P0: Bound MCP session lifecycle and make Dev MCP cutovers multi-session safe #31/P1: Align idleTimeoutMs tool contract with enforced runtime semantics #32/P1: Isolate mutating ChatGPT conversations and expose cross-session capacity contention #33 have merged replacement implementations.
• P0: Prevent live Dev MCP runtime lineage drift and capability regression #30 host-visible schema freshness may proceed in read-only/host-lifecycle parallel because it does not require source mutation.
• Live concurrent-session acceptance in P0: Bound MCP session lifecycle and make Dev MCP cutovers multi-session safe #31 must not be considered complete until this integrated-current-main gate is green and P0: Prevent live Dev MCP runtime lineage drift and capability regression #30 host schema binding is current.
• Harden Dev MCP reliability across host binding, worker startup, and timeout reconciliation #15 final real-task reliability trial remains downstream.

Non-goals

• deploying/restarting live Dev MCP;
• refreshing the ChatGPT host tool catalog;
• closing P0: Prevent live Dev MCP runtime lineage drift and capability regression #30/P0: Bound MCP session lifecycle and make Dev MCP cutovers multi-session safe #31/Harden Dev MCP reliability across host binding, worker startup, and timeout reconciliation #15;
• adding new reliability features;
• changing Nexus authority or G9/G10 semantics.

Claim ceiling

R7_INTEGRATED_CURRENT_MAIN_VERIFIED only for the exact tested commit/tree and verification environment.

This Issue does not establish live runtime reliability or multi-session production readiness.