Move thread list + messages off Redux cache to direct RPC reads #1197

Status: Draft / scoping. Filed as a follow-up to #1196 to capture the structural shape of the fix; not scheduled.

Summary

Stop maintaining threadSlice.threads and threadSlice.messagesByThreadId as a parallel client-side cache. Replace consumers with direct RPC reads against openhuman.threads_list / openhuman.thread_messages (the Rust core already persists these as JSONL upsert logs in src/openhuman/memory/conversations/store.rs), wrapped in a stale-while-revalidate layer (react-query or equivalent). Keep selectedThreadId as a small persisted UI atom.

Problem / Context

The Rust core is already the source of truth — threads live in a JSONL upsert log under the workspace dir, and threads_list is exposed over JSON-RPC. The React app shadows that as a Redux slice (threads, messagesByThreadId, messages) for sync render-time access and optimistic updates. Maintaining two stores creates a steady stream of cache-coherence bugs:

• fix(threads): persist selectedThreadId across reloads #1168 / fix(threads): preserve selectedThreadId on cold-boot identity hydration #1196 — the persisted selectedThreadId was wiped by clearAllThreads() on the boot-time identity-observation path because the slice conflates selection with cache.
• Fix new signup showing existing/stale chat threads #1157 — pre-auth thread rows had to be explicitly scrubbed from the slice when identity hydrated.
• Multiple ad-hoc invalidation paths in CoreStateProvider and Conversations.tsx exist to keep the cache in sync.

Going cache-less makes this entire class of bug structurally impossible. The cost is a non-trivial refactor — see Scope.

Scope

In scope

• Introduce a query layer over coreRpcClient for thread reads (useThreads(), useThreadMessages(id)). Stale-while-revalidate semantics, error/loading states surfaced to consumers, automatic refetch on identity change.
• Migrate ~30 useAppSelector(s => s.thread.*) call sites in app/src/pages/Conversations.tsx, app/src/components/**, app/src/providers/ChatRuntimeProvider.tsx, etc.
• Migrate the streaming append path in ChatRuntimeProvider — today it dispatches addInferenceResponse / addMessageLocal thunks that mutate messagesByThreadId. Needs to push into the query cache instead (or hold the in-flight turn in component-local state and let the post-turn refetch reconcile).
• Migrate optimistic UI for rename / delete / new-thread to query-cache mutations.
• Reduce threadSlice to just { selectedThreadId, activeThreadId }. Drop threads, messagesByThreadId, messages, welcomeThreadId (already a no-op per feat(onboarding): replace welcome-agent bot with react-joyride walkthrough #1180), and the loadThreads / loadThreadMessages thunks.
• Remove resetThreadCachesPreservingSelection (added in fix(threads): preserve selectedThreadId on cold-boot identity hydration #1196 as a workaround) and the boot-time scrub branch in CoreStateProvider.tsx:259 — no in-memory cache means nothing to scrub.

Out of scope

• Any change to the Rust storage layer or RPC surface. The JSONL upsert log is fine as-is.
• Realtime push from core → app for thread list changes (currently a poll/refetch model). Could be revisited separately if latency becomes a UX concern.
• Replacing Redux as a whole — non-thread slices stay.

Tradeoffs

• Adds a query-layer dependency (likely react-query) — already a common pattern, but a new top-level concept in this app.
• Streaming UX is the trickiest piece: tokens arrive ~10–50/s and currently mutate the slice cheaply. Need to either (a) buffer in component state during the turn and refetch on chat_done, or (b) write directly into the query cache with manual invalidation. (a) is simpler but flickers if the refetch is slow; (b) preserves current behavior.

Acceptance criteria

• Query hooks land — useThreads() and useThreadMessages(id) exist, hit coreRpcClient, and have unit tests covering loading/error/refetch.
• All consumers migrated — no remaining useAppSelector references to state.thread.threads or state.thread.messagesByThreadId. Streaming append flows through the new layer.
• threadSlice shrinks — only { selectedThreadId, activeThreadId } remain. The boot-time scrub branch in CoreStateProvider is removed.
• Regression coverage — existing fix(threads): persist selectedThreadId across reloads #1168 / Fix new signup showing existing/stale chat threads #1157 / fix(auth): stop stale chat threads after signup #1192 behaviors still hold (reload resumes the user's thread; pre-auth rows don't leak across identity changes; signup doesn't show stale titles).
• Diff coverage ≥ 80% — the implementing PR meets the changed-lines coverage gate (Vitest + cargo-llvm-cov, enforced by .github/workflows/coverage.yml) when code changes are involved.

Related

• fix(threads): preserve selectedThreadId on cold-boot identity hydration #1196 — fix(threads): preserve selectedThreadId on cold-boot identity hydration (the patch this issue would obsolete)
• fix(threads): persist selectedThreadId across reloads #1168 — original selectedThreadId persistence work
• Fix new signup showing existing/stale chat threads #1157 / fix(auth): stop stale chat threads after signup #1192 — pre-auth stale-thread scrub that motivated clearAllThreads()
• src/openhuman/memory/conversations/store.rs — JSONL upsert log (source of truth)
• app/src/store/threadSlice.ts — slice to be shrunk
• app/src/providers/ChatRuntimeProvider.tsx — streaming append path