upstream merge 2026-04-27 PR-C: chat + plans + version bump (4 commits)#447
Merged
upstream merge 2026-04-27 PR-C: chat + plans + version bump (4 commits)#447
Conversation
…rim fix to v2 (superset-sh#3716) * docs(chat): v2 chat architecture proposals + reference research Plans-only PR. Adds the chat-architecture research and proposals for the v2 chat rearchitect, plus a pragmatic v1->host-service migration plan that unblocks the greenfield work. - v2-chat-greenfield-architecture.md: tRPC subscriptions over WS, EventLog interface, P0-P4 race-fix + LocalEventLog (SQLite), P5a (Postgres) / P5b (Durable Objects), P6 (handoff via git courier), P7 (mid-turn handoff, deferred). - v1-to-v2-fast-migration.md: ship-this-month plan. Move ChatRuntimeService into host-service, collapse two polling queries into one atomic getSnapshot, gate behind per-workspace flag, canary, GA, delete legacy paths. - t3code/opencode/background-agents reference docs: research notes informing the design. * docs(chat): expand fast-migration plan with audit + verified bugs Adds an Implementation Audit section (file inventory, what's solid, stubs, gaps, verified bugs), a Fixes-at-a-glance priority table, and restructures Phased Migration so P0 captures the two HIGH defects concretely: - Runtime leak on session delete (no disposeRuntime / endSession) - Cross-workspace sessionId race in runtimeCreations map P1 folds in the dual-poll race fix via a single getSnapshot query, the fps:60 -> default 4fps polling cadence drop, and the missing cloud lastActiveAt update on host send (selector ordering). P4 enumerates parity gaps with v1 file references; Stop and Notification hook events explicitly deferred per scope decision. * fix(chat): runtime disposal, cross-workspace race, snapshot collapse Lands the P0 + P1 fixes from plans/v1-to-v2-fast-migration.md. Host-service runtime (packages/host-service/src/runtime/chat/chat.ts): - runtimeCreations map now stores { workspaceId, promise } and validates the workspace at the in-flight path. Concurrent requests with the same sessionId across different workspaceIds no longer silently share the wrong-workspace runtime. - Adds disposeRuntime(sessionId) - aborts harness, drops from map, idempotent. Closes the runtime leak on session delete. - Adds getSnapshot(input) returning { displayState, messages, observedAt } from a single observation. Reuses the existing display-state shaping logic. Host-service router (packages/host-service/src/trpc/router/chat/chat.ts): - New chat.getSnapshot query. - New chat.endSession mutation. - chat.sendMessage fires-and-forgets ctx.api.chat.updateSession with lastActiveAt: new Date() after success so the v2 session selector keeps reordering after activity. Cloud router (packages/trpc/src/router/chat/chat.ts): - updateSession now accepts an optional lastActiveAt: z.date(). V2-workspace ChatPane: - useWorkspaceChatDisplay: replaces dual getDisplayState + listMessages queries with a single getSnapshot query. Derives state from snapshot.displayState / snapshot.messages. - ChatPaneInterface: drops fps: 60 override (~120 RPCs/sec/pane) in favor of the hook's default fps: 4 (matches v1). - useWorkspaceChatController: handleDeleteSession now also calls workspaceTrpc.chat.endSession after the cloud delete to tear down the host-service runtime. P4 parity gaps (slash commands, file search, hooks, title generation, MCP, model auth state) are deferred per scope. * feat(chat): enable v2 chat pane Drops the last v1-IPC coupling from TiptapPromptEditor by switching the file-search and slash-command-preview calls to prop injection. v1's ChatInputFooter wires them to chatServiceTrpc (legacy IPC, unchanged); v2's wires searchFiles to workspaceTrpc.filesystem.searchFiles so the @-mention popup works for both local and remote host-service workspaces. The chip-anchored SlashCommandPreviewPopover becomes opt-in via prop and stays off in v2, which already mounts its own SlashCommandPreview using workspaceTrpc.chat.previewSlashCommand. usePaneRegistry now mounts the real <ChatPane> instead of the "temporarily disabled" placeholder, threading sessionId / launchConfig through ctx.actions.updateData. * feat(chat): wire v2 ModelPicker auth status to host-service Adds host-service auth surface and connects the v2 ModelPicker to it, so the picker can show real Anthropic/OpenAI authentication state instead of hardcoded `true`. Host-service: - types.ts: HostServiceRuntime gains an `auth: ChatService` field. - app.ts: instantiate a long-lived ChatService singleton (auth is per-machine, not per-workspace) and expose it via runtime.auth. - New trpc/router/auth/auth.ts router with 17 procedures mirroring v1's chatServiceTrpc.auth.* surface (getAnthropicStatus, startAnthropicOAuth, completeAnthropicOAuth, etc.). - Wired into appRouter as `auth`. ChatService delegates to mastra's createAuthStorage() under the hood, so this is a thin proxy — no auth logic is duplicated. Mastra's storage files are the same paths host-service already reads from when resolving credentials at runtime, so v1 settings/models writes and v2 reads stay in sync via disk. Renderer: - v2 ModelPicker now queries workspaceTrpc.auth.getAnthropicStatus and getOpenAIStatus; previously hardcoded both to authenticated. Refetches on selector open. The "open settings" action keeps navigating to /settings/models which is the existing v1 page (still works because it writes via Electron-main IPC to the same on-disk storage). Remote-host auth UI is a follow-on; for remote workspaces the status correctly reflects the remote machine's credentials, but the /settings/models flow only authenticates the local machine. * fix(chat): pass observer/reflector model to mastra runtime Mastra defaults observational-memory observer + reflector models to google/gemini-2.5-flash. Without GOOGLE_GENERATIVE_AI_API_KEY in the host-service process env, the post-turn reflection step fails with "Could not find API key process.env.GOOGLE_GENERATIVE_AI_API_KEY". v1's ChatRuntimeService already handles this in packages/chat/src/server/trpc/service.ts:88 — derive the OM model from whichever provider the user is authenticated with (Anthropic -> Haiku 4.5, OpenAI -> GPT-4.1 nano, Google env -> Gemini). Mirror that helper in host-service and pass the result as initialState {observerModelId, reflectorModelId} to createMastraCode. * feat(chat): wire v2 slash commands to host-service Replaces the host-service slash-command stubs with v1's actual logic (scans .claude/commands and .agents/commands in the workspace cwd plus the global ~/.claude/commands). v1's getSlashCommands and resolveSlashCommand are pure cwd-scoped functions — no Electron deps — so they import cleanly into host-service. @superset/chat: - server/desktop now re-exports getSlashCommands, resolveSlashCommand, and SlashCommand from the slash-commands subpath so host-service can consume them via the package's public surface. Host-service: - runtime/chat/chat.ts: getSlashCommands/resolveSlashCommand/preview- SlashCommand now look up the workspace's worktreePath and delegate to the v1 helpers (via a small resolveWorkspaceCwd helper). Returned command kinds are "builtin" | "custom" matching v1. v2 client: - useSlashCommandExecutor's default switch case now calls workspaceTrpc.chat.resolveSlashCommand for non-builtin commands. Successful resolutions substitute the rendered prompt as nextText via resolveSlashPromptResult; unhandled commands fall through with the original text. Built-in commands (/new /clear /stop /model /mcp) still resolve client-side without a roundtrip. * fix(chat): getSnapshot regression + disposeRuntime hardening Addresses bugs introduced in c183954 (getSnapshot collapse + runtime disposal). Pre-existing v1 issues (authStorage.reload guarding, OpenAI provider ID compat) are intentionally left out — they exist in v1's service.ts too and should be fixed there separately. getSnapshot regression - getSnapshot duplicated displayState shaping inline and diverged from getDisplayState in two ways: it skipped the answeredQuestionIds filter (so answered ask_user questions could shadow sandbox questions the same turn — exactly the bug the existing filter was added to prevent), and the sandbox question shape differed (no top-level description; reason baked into the "Yes" option label). - Extract a private buildDisplayState(runtime) helper used by both getDisplayState and getSnapshot so they cannot drift again. Both now apply the answeredQuestionIds filter and emit the same sandbox shape. disposeRuntime hardening - Accept and validate workspaceId. Mirrors getOrCreateRuntime; prevents cross-workspace teardown if a sessionId leaks. Router updated to pass it through. - If a creation is in-flight for this sessionId, await it before cleanup. Otherwise dispose returns no-op while createRuntime is mid-way and the just-created runtime gets inserted into this.runtimes after we delete from it — leak. - Disconnect the MCP manager (await runtime.mcpManager?.disconnect()). Mastra's McpManager owns child processes / server connections that leak until process exit otherwise. HookManager has no teardown API, so leaving it untouched is correct. All cleanup steps are best-effort (try/catch) so a single failure doesn't block the rest of the teardown. * fix(chat): slash commands work without an active session Slash commands are workspace-scoped (they only need cwd to scan .claude/commands and .agents/commands), but the host-service procedures required a sessionId via z.uuid() and the v2 client gated everything on having a session. Result: in a fresh chat with no session yet (the common case), the autocomplete was empty, the preview popover never fired, and pressing Enter on a custom slash command sent the literal "/foo args" text instead of resolving it. Host-service: - New workspaceSlashInput schema (workspaceId only) replaces sessionInput on getSlashCommands / resolveSlashCommand / previewSlashCommand. - ChatRuntimeManager methods drop the sessionId parameter (it was unused anyway — they only consume workspaceId to resolve cwd). v2 client: - ChatPaneInterface.tsx: getSlashCommands.useQuery no longer gated on sessionId; passes only workspaceId. Autocomplete now fires on fresh sessions. - SlashCommandPreview: drops sessionId prop and the !sessionId early return; previewSlashCommand mutation passes only workspaceId. - useSlashCommandExecutor: removes the !sessionId early return that blocked all non-builtin slash commands when no session exists. The switch's per-case guards handle session-dependent built-ins (/mcp still self-guards on sessionId since getMcpOverview is unchanged). Default case (custom commands) calls resolveSlashCommand with workspaceId only. - ChatInputFooter: drops the now-unused sessionId prop. * fix(chat): optimistic message lost on first send in v2 When v2 chat migrated from chat.listMessages to chat.getSnapshot in c183954, sendMessageToSession kept writing optimistic user messages to the listMessages cache via setData. The display now reads from snapshot.messages, so the optimistic message landed in a dead cache and the first user message disappeared until the next snapshot poll fetched it back from the harness (~250ms+ at 4fps polling). Most visible on the first message of a fresh chat (page is empty before it), but every send had the gap. Update sendMessageToSession to write to chat.getSnapshot.setData. Two cases: - Existing snapshot in cache: spread, append optimistic message. - Fresh session, no snapshot yet: seed with a minimal displayState (the next poll fills it in) and an array containing just the optimistic message so it renders immediately. Rollback path mirrors: filter the optimistic id out of snapshot.messages on send failure. * perf(chat): cut idle-poll rerender churn in v2 ChatPane Two small changes that eliminate forced rerenders when nothing has actually changed in the snapshot. 1. Drop observedAt from chat.getSnapshot The server was stamping every snapshot response with observedAt: Date.now(). Nothing on the client read it. But because the timestamp differs every poll, React Query's structuralSharing couldn't preserve the response object's identity — every 250ms (4fps polling) produced a new reference, forcing every component reading `snapshot` to rerender even when displayState and messages were byte-identical. Removing observedAt lets structural sharing kick in: idle polls now produce identical objects and downstream consumers don't rerender. Also drops the field from the optimistic-seed object in sendMessageToSession so it stays in sync with the new shape. 2. Memoize the slashCommands select mapper ChatPaneInterface.tsx had an inline `select: (commands) => commands.map(...)` in the getSlashCommands.useQuery options. The arrow function was recreated every render → mapper re-ran → new array reference → every consumer of `slashCommands` (TiptapPromptEditor, executor, etc.) rerendered even on identical input. useCallback the mapper and pass the stable reference instead. Together these two cuts close the cascade: poll resolves with structurally-equal data → React Query preserves identity → consumers see same reference → no rerender. Subsequent message-arrival polls still rerender (because data actually changed) — that's correct. * fix(chat): first-message flicker — keep optimistic in local state, not RQ cache User saw the message flicker in/out on the first send: thinking, then the message, then nothing, then thinking again, then the full result. Cause was a stale poll clobbering the optimistic update. The existing-session send path already does this correctly — useChatDisplay holds an `optimisticUserMessage` in component state and merges it into the messages output, outside React Query's cache. Stale polls can't overwrite local component state, so the message stays visible until the harness's response includes it. The new-session path (sendMessageToSession, only fired on first send) took a different route: write the optimistic message into the chat.getSnapshot React Query cache. With 4fps polling running continuously, a poll often resolved during the mutation with the harness's pre-message state and clobbered the optimistic write — message vanished, then reappeared on the next poll once the harness had processed it. Switch the new-session path to the same pattern: set pendingUserTurn local state in handleSend before calling sendMessageToSession. getVisibleMessagesWithPendingUserTurn already merges that into the displayed list and clears it once the server's messages array contains a matching entry. sendMessageToSession reduces to a thin mutation wrapper — no cache writes, no rollback. The restart flow already used pendingUserTurn for the same reason; this just brings the new-session path into parity.
Move ad-hoc reference notes out of the tracked plans/ tree and into plans/local/, which is gitignored.
PR-C 取り込み中、6c2a7719a (v2 chat architecture proposals) の auto-merge 残りを処理: - ChatPaneInterface.tsx: chatServiceTrpc と AppRouter 両方の import が必要なため両方残す (前者は auth 系状態問い合わせ、 後者は inferRouterOutputs<AppRouter>["chat"]["getSlashCommands"] 型導出に使用) - ModelPicker.tsx: getAnthropicStatus / getOpenAIStatus を workspaceTrpc.auth へ寄せた upstream の方針を採用 (chatServiceTrpc 側は他箇所 useNextEditCompletion 等に残置) - chat.ts: auto-merge で混入した node:fs / node:os / node:path の未使用 import を削除 (slash command 経由で fs 操作不要) bun.lock 更新は workspace version 1.5.10 → 1.6.2 への追従 (c885eb2 取り込み結果)。
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
🧹 Preview Cleanup CompleteThe following preview resources have been cleaned up:
Thank you for your contribution! 🎉 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
upstream (superset-sh/superset) から fork (MocA-Love/superset) への 2026-04-27 バッチ取り込み 第3弾 (最終)。22 commits を 3 PR に分割した中の PR-C: chat / plans / version bump (4 commits + 1 fix commit)。
取り込み内容
2d41885e9plans/*.mdで他社チャット (Open-Inspect / OpenCode / T3 Code) アーキテクチャ調査ドキュメント追加dab3225e3plans/local/を gitignore に追加し、6c2a7719a で導入された Chat architecture reference docs (background-agents-chat-architecture-reference.md,opencode-electron-chat-architecture-reference.md,t3code-chat-architecture-reference.md) を削除3f6c39c60971cf80eb加えて fork 固有の解消 commit
fix(merge): clean up chat trpc + unused imports after #3716を含む。Fork 側の方針変更点
desktop version 1.6.2 へ追従
fork は従来
1.5.10固定でv1.5.10-fork.N運用していたが、upstream の major/minor 進行 (1.6 系) に合わせて追従。次のリリースタグからv1.6.2-fork.1で開始する (これまでのv1.5.10-fork.N系列は終了)。.gitignoreの fork 方針upstream
dab3225e3(superset-sh#3782) はplans/local/,.mcp.json,.cursor/mcp.jsonの3つを ignore に追加するが、fork は AGENTS.md ルールにより.mcp.jsonをコミット追跡対象として管理しているため、.mcp.jsonと.cursor/mcp.jsonの ignore は取り込まない。plans/local/のみ追加。chat trpc 経路の整理
6c2a771 で
ModelPickerの Anthropic / OpenAI 状態取得がchatServiceTrpc.auth.*からworkspaceTrpc.auth.*(host-service 経由) に移行された。fork も同方針を採用 (host-service 経由)。chatServiceTrpcはuseNextEditCompletion等の他箇所で引き続き使用しているため import は維持。ChatPaneInterface.tsxではchatServiceTrpcとAppRouter(host-service) の両方が必要 (前者は auth 系問い合わせ、後者はinferRouterOutputs<AppRouter>["chat"]["getSlashCommands"]型導出)。両 import を保持。Fork 固有機能ヘルスチェック
baseline と比較し全項目健在を確認:
ansi_up,@vscode/ripgrep,@xyflow/reactelectron-builder.tsのdmg.size = "4g"維持packages/db/drizzle/最大 idx 0039 維持 (PR-B から)packages/local-db/drizzle/最大 idx 0072 維持Test plan
bun install整合性 OKbun run typecheck全 28 task greenbun run lintbiome greenbun run --filter @superset/desktop build成功確認runtime/chat/chat.ts経由) が動作v1.6.2-fork.1タグを切れることを確認このバッチでの取り込み完了
合計 22 upstream commits を取り込み完了。