feat: wire Stream Event v1 transport and renderer log

[sjr666666]

Summary

Wires the Stream Event Protocol v1 across transport and renderer (issue #27), additive to the existing AgentEvent path:

Transport (#43)

• apps/electron/src/main/kernelHost.ts: attach() also subscribes RunManager.subscribeStream and forwards { sessionId, event } over a new IPC channel agent:stream (cleaned up on re-attach). Legacy agent:event untouched.
• apps/electron/src/preload/index.ts + index.cjs: expose kernel.onStreamEvent(cb) -> unsubscribe, mirroring onAgentEvent.

Renderer data layer (#43)

• packages/ui/src/atoms/streamAtoms.ts (new): pure reduceStreamLog keeps per-run events ordered by sequence, dedupes replays (idempotency requirement) into a drops counter, and flags gaps/out-of-order delivery as anomalies. Exposed as Jotai streamLogAtom / applyStreamEventAtom.
• KernelBridge subscribes onStreamEvent into the log; the legacy agent:event → applyAgentEventAtom path is untouched.
• FinagentClient interface + fallback noop + finagentClient.ts wiring.

Type fix (#43)

• packages/core/src/stream-events.ts: StreamEvent is now a true discriminated union (distributive mapped type) so payload narrows by type — the feat(core): add Stream Event Protocol v1 types (ADR 0001) #41 definition failed tsc narrowing. Shape unchanged.

No visual change — renderer data layer only; rendering from text_delta (and the explicit cancelled state) is the follow-up, best after the protocol open questions settle.

Verification

• Protocol tests: stream-events.test.ts + stream-event-adapter.test.ts → 10 passed
• Renderer: streamAtoms.test.ts (5) + sessionAtoms.test.ts (5) passed
• Kernel E2E regression: agent-kernel.test.ts → 6 passed
• tsc --noEmit for electron/core/ui/shared → no new errors beyond the known pre-existing i18next resolution issue (local bun 1.4 layout)

Dependency

Builds on #42 (RunManager stream channel) → #41 (protocol types). The discriminated-union fix lives here so #41 merges as-is.

Open questions

Same as ADR 0001 (resume data source, status.phase, messageId vs runId).

[helsome]

这组 #41 → #42 → #43 的方向是对的，但当前还不建议把协议定型进 main，有两个架构阻塞需要先解决：

1. messageId === runId 不应作为 v1 固化。#34 已明确要求 message / generation / run 是不同身份：message 有稳定 id，assistant generation 绑定具体 run。现在把两者合并会直接给 edit/regenerate/fork 留下错误语义。建议让 message 级事件携带真实 messageId；run 级事件可只依赖 runId，或 messageId 可选，但不要把两者定义成同一身份。
2. ADR 写了 reconnect=lastSequence replay，但当前 stack 只有 parallel event channel + renderer log/dedupe，没有 replay source/API/buffer，也没有“明确无法恢复”的失败路径。也就是说 contract 文档承诺了实现尚不存在的行为。

另外 #43 自己说明 #41 的 StreamEvent 不是正确 discriminated union，要到 #43 才修，所以不要单独 merge #41/#42 后在 main 留一个已知有缺陷的协议类型。

建议：在本 stack 顶部统一修掉 identity contract，并至少实现一个明确的 reconnect 行为（内存 replay 也可以，或明确 terminal failure），然后补 Issue #27 要求的真实 Copilot E2E：model + tool + citation，分别验证 cancel 和 transport interruption。完成后这组三层可以按依赖顺序合。

[sjr666666]

Thank you for the detailed review. Both blockers are now addressed at the top of this stack (new commits just pushed):

1. Identity contract (#34) — messageId is decoupled from runId:

• Envelope: messageId?: string; message-level events (message_started / text_delta / message_completed / cancelled) carry the real assistant messageId, run-level events (run_started / run_completed / tool_* / citation_added / status / error) carry only runId.
• RunManager now pre-assigns one stable assistant message id per run and threads it through the message-level events (and reuses it as the persisted assistant message id, so stream ids and storage ids are the same object identity).
• Idempotency key is now runId + sequence only.
• ADR 0001 updated; open question 3 marked resolved.

2. Reconnect — implemented a concrete in-memory replay plus an explicit unrecoverable path:

• StreamEventHistory (packages/shared/src/kernel/stream-history.ts): per-run in-memory tail, hardened for eviction.
• RunManager.replayStream(runId, lastSequence) returns the contiguous tail; when the run is unknown or the tail is non-contiguous (eviction) it returns recoverable: false.
• Exposed over IPC: runs:stream-replay (main handler + preload ts/cjs + FinagentClient contract).

Verification: adapter (7, incl. messageId injection) + history (5) + kernel E2E regression (6) all pass; electron/core/ui/shared tsc show no new errors beyond the pre-existing local i18next resolution issue.

On the real Copilot E2E (model + tool + citation, cancel and transport interruption): I'll wire this next against the eval fixture infrastructure so it runs without external API keys — will follow up in this thread before we merge.

作者已解决原 review 的两个实质 blocker（message/run identity 与 reconnect replay）。重新评估后，真实完整 E2E 不再作为当前合入前置条件；剩余仅需解决与最新 main 的 Git 冲突并通过基础 CI。

[helsome]

重新审查后通过代码方向。你已经解决了上一轮的两个核心 blocker：messageId / runId 身份拆分，以及明确的 in-memory replay / unrecoverable reconnect 路径。完整真实 Copilot E2E 可以作为后续增强，不再作为当前合入前置条件。现在只剩 Git 层面的 dirty：请把最终 #43 rebase 到最新 main，保留 Run Budget / Provider Router 等主干后续改动，并让基础 CI（unit/typecheck/secret scan）跑通；完成后可直接合。

审核标准已更新：Ready PR 在 APPROVE 前必须补齐可复现测试报告（至少 Bun 版本、OS、实际命令与结果）。原审批基于旧标准，先撤销，代码方向本身仍认可。

[helsome]

代码方向仍然认可，上一轮 identity/reconnect blocker 已修；这次 Changes Requested 只是同步当前统一验收规则。请在 clean rebase 最新 main 时，把测试报告补成可复现格式：Bun 版本、OS/平台、实际 focused test/typecheck 命令与 pass/fail；现有 protocol/renderer/kernel 测试范围够了，不需要补完整真实 Copilot E2E。rebase + 基础 CI + 报告字段齐全后即可恢复 APPROVE。

[helsome]

这轮复查有两个更新：

• 你在 feat(shared): persist stream event log for cross-restart replay (#75) #76 补充的同一套 stack 验证环境（Bun 1.4.2 / Windows 10.0.26200）与现有 protocol/renderer/kernel 测试范围足够；e0d46c4 对应 PR checks 也已 success，所以不再要求扩大测试范围。
• 但刚刚 feat(ui): add typed financial answer blocks for Copilot (#31) #81 已合入 main（6a9a288），feat: wire Stream Event v1 transport and renderer log #43 现在再次 dirty。请只做一次 clean rebase 到当前 main 并跑现有 PR checks。

合并策略调整为：#43 clean 后直接以 stack 顶部合入 #41→#42→#43 的完整协议增量，不再分别 squash #41/#42，避免重复历史；随后再处理 #76 的 persistence-only delta。