fix(session): recover transcript streams and virtualize activity feed (ZER-673)

[ZeroPointSix]

Summary

• Restore the session detail feed in the required order: transcript snapshot, tail replay, then live SSE.
• Capture the snapshot and unified SSE cursor atomically so the hand-off cannot lose events; stable entry_id values upsert streamed text and tool revisions.
• Render long session activity through react-virtuoso, expose end-to-end SSE latency, and surface cursor resets, history gaps, and stream lag as explicit recovery states.
• Clear all visible transcript state before a session route changes, so a failed or slow B snapshot never displays A's activity under B's metadata.

Review Contract

Goal

Deliver ZER-673's resilient session transcript recovery and long-session activity feed: snapshots must resume safely into replay/live SSE, and transcript rendering must remain responsive for 1,000+ entries.

Non-goals

This PR does not change agent execution semantics, transcript retention capacity, authentication policy, or the server-side protocol beyond exposing the atomic shared-stream cursor already required by the recovery hand-off. It does not introduce write or mutation controls to the read-only session feed.

Accepted Residual Risks

The activity feed remains dependent on the bounded server replay history. When a cursor reset, history gap, or stream lag is detected, the UI intentionally stops live processing and asks the operator to reload a full snapshot instead of silently reconstructing uncertain state. The current production bundle-size warning is pre-existing/non-blocking and is not addressed here.

Acceptance Criteria

• Session details restore in the order full transcript snapshot, transcript tail, then live SSE using the snapshot's generation-qualified stream cursor.
• A streamed transcript revision or tool lifecycle update replaces the matching stable entry_id instead of appending a duplicate row.
• cursor_reset, event-history gaps, and stream lag show an explicit recovery action rather than silently omitting history.
• Changing from session A to B clears A's transcript, latency observation, and recovery state before B's snapshot is displayed; a B snapshot failure cannot resurrect A's content.
• The activity list uses virtualization for long transcripts.
• Regression coverage includes transcript upsert/parsing, session switch reset behavior, and concurrent lifecycle publish with replay/subscribe completion under a timeout.

Follow-ups

• Observe recovery alerts and end-to-end SSE latency in staging with long-running real sessions.
• Consider a dedicated browser-level route-transition test once the web test environment includes DOM rendering support.
• Review client-bundle code splitting separately from this resilience and correctness scope.

Verification

• pnpm test — 43 tests passed.
• pnpm lint
• pnpm build
• cargo test -p openab-core --no-default-features transcript::tests
• cargo test -p openab-core --no-default-features session_event::tests
• cargo test -p openab-gateway --no-default-features session_admin::tests

Closes ZER-673.

[linear-code]

ZER-673

[ZeroPointSix]

Context fact card

Target

• Repository: ZeroPointSix/openab-plus
• PR: fix(session): recover transcript streams and virtualize activity feed (ZER-673) #45 fix(session): recover transcript streams and virtualize activity feed (ZER-673)
• Requested outcome: 审查 transcript 恢复链、长会话虚拟化、相关回归与验收完整性
• Base branch: main
• Head: d3be1a9b49ff18f45fd3f32ba0391263128ddd46

Verified facts

• PR 当前 open、非 draft、mergeable，27 个文件，+3679/-371。
• 实现新增 per-session transcript ring buffer、统一 session/transcript SSE cursor、snapshot/tail/live 恢复链，以及 react-virtuoso 活动流。
• transcript snapshot 会原子捕获 shared stream generation/next sequence；前端使用 stable entry_id upsert streamed text/tool revisions。

Relevant code and tests

• Core: session_event.rs, transcript.rs, adapter.rs。
• Gateway: session_admin.rs + admin integration tests。
• Web: useSessionTranscript.ts, transcript.ts, SessionActivityFeed.tsx, SessionDetailPage.tsx。
• 已有测试覆盖 transcript chunk/tool upsert、history overflow、cursor capture、前端 entry upsert/解析；当前未看到并发锁顺序或 session A→B 切换时旧 transcript 清理的覆盖。

CI/CD and validation entrypoints

• CI run #120: success。
• Admin Web run #62: success。
• Docker Smoke Test run #136: success。
• Docker Smoke Test (Unified) run #143: 当前仍在运行。
• Review Contract run #185: failure；PR 描述缺少 ## Review Contract 及 Goal / Non-goals / Accepted Residual Risks / Acceptance Criteria / Follow-ups 必需章节。

Assumptions not allowed

• 不能因为单线程单测通过就假设 lifecycle publish 与 SSE replay/subscribe 不会并发。
• 不能假设路由切换时新 session transcript 一定比 session metadata 更快返回。

Risks / blockers

• 已发现 SessionEventBus 的 stream-history / lifecycle-history 锁获取顺序反转，存在并发死锁风险。
• 正在确认 session 切换期间旧 transcript 暂留问题及其用户可见影响。
• Review Contract 当前硬失败；Unified Docker Smoke 尚未完成。

Proposed change scope

• 本轮仅审查，不修改实现；代码 blocker、验收缺口和 CI 状态会以中文 review 同步到本 PR。

[ZeroPointSix]

审查结论：当前不建议合并

整体方案方向正确：transcript snapshot + tail + live SSE 的恢复顺序、shared stream cursor、stable entry_id upsert，以及 react-virtuoso 长列表虚拟化都覆盖了 ZER-673 的主要目标；CI、Admin Web、普通 Docker Smoke Test 也已通过。

但当前有两个代码 blocker，外加一个合并门禁 blocker：

1. [高] SessionEventBus 存在锁顺序反转，可能并发死锁。

   • publish() 先通过 stream.publish_session() 获取 SessionStreamHistory mutex，释放后再获取 lifecycle SessionEventHistory mutex；更关键的是 publish 路径与 transcript/cursor 路径建立了 stream→history 的顺序。
   • subscribe_after() / replay_after() 当前先持有 lifecycle history mutex，再调用 self.stream.next_sequence() 获取 stream history mutex，即 history→stream。
   • 在并发 lifecycle publish 与 replay/subscribe 时形成 ABBA 锁顺序，存在请求/SSE 建连卡死风险。现有 session_event::tests 都是串行功能测试，没有并发锁顺序覆盖。
   • 建议不要在持有 lifecycle history 锁时调用 stream.next_sequence()；先获取 next sequence 后再锁 lifecycle history，或统一所有路径的锁顺序，并补并发回归测试（带 timeout）。

2. [中] session A→B 路由切换时会保留旧 transcript，可能短暂展示错误会话内容。

   • useSessionTranscript() 只有 sessionId 为空时才清 entries；非空 A 切到非空 B 时，effect 开始恢复 B，但旧的 entriesRef/current entries 保持到 B 的 transcript snapshot 返回。
   • SessionDetailPage 的 session metadata 与 transcript 是独立请求；如果 B 的 metadata 先完成，页面会渲染 B 的标题/元数据，同时活动流仍是 A 的内容，直到 transcript 请求完成。若 B transcript 请求失败，旧内容还可能一直留在 recovery banner 下。
   • 建议 sessionId 变化后、发起 snapshot 前立即清空 entries/latency，并增加 A→B 切换与 snapshot failure 的 hook/component 测试。

3. [门禁] Review Contract 当前失败。

   • run #185 明确报缺少 ## Review Contract、### Goal、### Non-goals、### Accepted Residual Risks、### Acceptance Criteria、### Follow-ups。
   • PR 描述目前只有 Summary / Verification，因此即使代码问题修复，也还不能满足仓库合并门禁。

CI 状态

• CI #120: success
• Admin Web #62: success
• Docker Smoke Test #136: success
• Review Contract #185: failure
• Docker Smoke Test (Unified) #143: in progress

验收缺口

• 需要补统一 stream 的并发 publish + subscribe/replay 无死锁测试。
• 需要补 session A→B 切换时旧 transcript 不可见、B snapshot 失败也不会保留 A 内容的测试。
• Review Contract 需补齐后重新过门禁。
• 最终合并前还需要等待 Unified Docker Smoke 完成并确认绿色。

因此当前结论是：先修上述两个代码 blocker + Review Contract，再复审；暂不合并。

[ZeroPointSix]

Execution summary

Problem

• 审查 ZER-673 transcript 恢复、统一 SSE cursor、长会话活动流虚拟化实现。

Root cause / verified facts

• 统一 stream 接入后，SessionEventBus 的 publish 与 replay/subscribe 路径存在 lifecycle history / stream history 锁顺序反转，带来并发死锁风险。
• useSessionTranscript() 在非空 sessionId A→B 切换时不会先清空旧 entries；metadata 与 transcript 独立加载时可能把 A 的活动流短暂显示在 B 页面，snapshot 失败时旧内容可能继续残留。
• 当前 Review Contract run #185 失败，PR 描述缺少全部必需 Review Contract 章节。

Solution

• 本轮仅审查，未修改代码。
• 建议统一 mutex 获取顺序，并增加并发 publish + subscribe/replay timeout 回归测试。
• 建议 sessionId 变化时在 snapshot 前立即清空 transcript/latency，并增加 A→B 与 snapshot failure 测试。
• 补齐 PR Review Contract。

Files reviewed

• Core: session_event.rs, transcript.rs, adapter.rs, ACP protocol/profile pool 相关变更。
• Gateway: session_admin.rs 与 admin API integration。
• Web: useSessionTranscript.ts, transcript.ts, SessionActivityFeed.tsx, SessionDetailPage.tsx 及相关测试。

Validation

• PR 自报：pnpm test 42 passed；pnpm lint；pnpm build；Core transcript/session_event tests；Gateway session_admin tests。
• GitHub Actions：CI #120 success，Admin Web #62 success，Docker Smoke Test #136 success。

CI/CD

• Review Contract #185: failure。
• Docker Smoke Test (Unified) #143: in progress（本次复核时）。

Remaining risks / follow-ups

• 两个代码 blocker 修复并补回归测试后需复审。
• Review Contract 必须转绿。
• Unified Docker Smoke 完成并转绿后再考虑合并。

当前结论：不建议合并。 Review 4922739281 已提交。

[ZeroPointSix]

已由 #46 的干净 ZER-404 W1–W5 只读工作台整合替代。#46 从最新 main 按依赖顺序重建，保留本 PR 的有效能力并补齐跨包冲突、活动流接线、恢复链与最新回归修复；原分支保留供历史追溯。