🤖 feat: boundary-windowed chat loading + metadata-only workspace activity

[ThomasK33]

Summary

Overhaul chat subscription architecture to load only the current compaction epoch on startup, scope full transcript streaming to the active workspace, and add cursor-based "Load More" pagination for older history.

Background

Previously, every workspace started a full onChat subscription that replayed from the penultimate compaction boundary, meaning all workspaces eagerly loaded two epochs of history regardless of whether they were visible. This caused unnecessary data transfer at startup and steady-state bandwidth waste.

This PR introduces three key changes:

1. Latest-boundary replay — initial load starts from skip=0 (latest boundary only) instead of skip=1
2. Active-workspace scoping — only the currently-displayed workspace gets a full onChat stream; all others use a lightweight metadata-only activity feed for sidebar indicators
3. Load More pagination — a cursor-based ORPC endpoint + frontend button to page backwards through older compaction epochs on demand

Implementation

Backend (agentSession.ts, historyService.ts, ORPC schemas/router)

• emitHistoricalEvents() now calls getHistoryFromLatestBoundary(workspaceId, 0)
• New workspace.history.loadMore endpoint returns a single older boundary window with cursor-based pagination
• New getHistoryBoundaryWindow() helper scans boundaries to return exactly one epoch window at a time

Frontend aggregator (StreamingMessageAggregator.ts)

• Renamed pruneBeforePenultimateBoundary → pruneBeforeLatestBoundary
• Live compaction now prunes everything before the incoming boundary's sequence, keeping only the current epoch

Frontend store (WorkspaceStore.ts, WorkspaceContext.tsx)

• addWorkspace() no longer starts runOnChatSubscription() — subscription is managed by ensureActiveOnChatSubscription()
• Only one full onChat stream active at a time, switched via setActiveWorkspaceId()
• Activity subscription (workspace.activity.list/subscribe) provides streaming/recency/model fallbacks for non-active workspaces
• Per-workspace pagination state initialized on caught-up, exposed as hasOlderHistory/loadingOlderHistory in WorkspaceState

Frontend UI (ChatPane.tsx)

• "Load older messages" button above transcript when hasOlderHistory is true
• Disabled with "Loading..." text during fetch

Validation

• make typecheck ✅
• make lint ✅
• make fmt-check ✅
• bun test across 3 targeted test suites (150 tests, 427 assertions) ✅
• bun test src/node/services/agentSession (43 tests) ✅

Risks

• Non-active workspace sidebar indicators now depend on the activity subscription. If the activity feed lags or disconnects, sidebar status (streaming, model, recency) may briefly show stale values. Fallback logic merges aggregator + activity data, preferring the fresher value.
• Load More edge cases — heavily corrupted histories with missing historySequence metadata may cause pagination to stop earlier than expected (safe degradation rather than crash).

---

📋 Implementation Plan

Plan: boundary-windowed chat loading + metadata-only workspace activity

Context / Why

We want chat startup to feel snappier by reducing unnecessary replay volume and decoupling sidebar status from full transcript streams.

Requested outcome:

1. Initial open should stream only from the latest compaction boundary (current epoch), not the penultimate boundary.
2. Add Load More pagination so each click reveals the previous compaction epoch window.
3. Stop relying on full workspace.onChat streams for every workspace; use a metadata-only subscription for cross-workspace status (new activity / stream finished), while keeping full chat streaming focused on the active workspace.

This reduces startup data transfer, avoids replaying large historical tails by default, and preserves user-controlled access to older history.

Evidence

• src/node/services/agentSession.ts
  • emitHistoricalEvents() currently calls historyService.getHistoryFromLatestBoundary(workspaceId, 1) (penultimate boundary replay) before caught-up.
  • onChat replay modes already exist (full / since / live) with cursor-based reconnect safety checks.
• src/node/services/historyService.ts
  • getHistoryFromLatestBoundary(workspaceId, skip) already supports boundary-window selection (skip=0 latest, skip=1 previous, etc.), which can back a Load More flow.
• src/browser/utils/messages/StreamingMessageAggregator.ts
  • Live compaction currently prunes to the penultimate boundary (pruneBeforePenultimateBoundary) and includes an explicit TODO for paginated older-history support.
• src/browser/stores/WorkspaceStore.ts
  • addWorkspace() immediately starts runOnChatSubscription() for every workspace during syncWorkspaces().
  • runOnChatSubscription() already uses since cursor reconnects via aggregator.getOnChatCursor().
• src/common/orpc/schemas/api.ts + src/node/orpc/router.ts + src/common/orpc/schemas/workspace.ts
  • A metadata-only feed already exists: workspace.activity.list + workspace.activity.subscribe with WorkspaceActivitySnapshot { recency, streaming, lastModel, lastThinkingLevel }.
  • Frontend currently does not consume this feed.

Storage layout assessment: split chat.jsonl into epoch files?

Recommendation: not in this iteration (keep single chat.jsonl + boundary/cursor pagination).

Why:

• Current hot-path replay already avoids full-file parse by reading from a boundary offset (HistoryService.findLastBoundaryByteOffset + readHistoryFromOffset).
• The bigger wins are at subscription scope/pagination level (active workspace streaming + metadata feed), not physical file sharding.
• File-sharding would require touching many persistence assumptions:
  • HistoryService mutation APIs (appendToHistory, updateHistory, truncateHistory, migrateWorkspaceId) currently treat history as one atomic file.
  • Subagent transcript archival/indexing and ORPC transcript reads store fixed chat.jsonl paths.
  • CLI/debug tooling and path conventions assume ~/.mux/sessions/<workspace>/chat.jsonl.

If we eventually shard by compaction epoch, what must be added?

• A manifest/index file (ordered shard list + min/max historySequence per shard).
• Atomic write protocol for shard rollover at compaction boundaries.
• Cross-shard lookup path for updateHistory(historySequence) and delete/truncate operations.
• Compatibility readers for legacy single-file sessions.
• Reassembly utility for any workflows that still require single-stream JSONL exports.

Estimated additional product LoC beyond the current plan: ~400–700 LoC (+ substantial test churn).

Recommended approach (A): active onChat + boundary pagination + activity metadata feed

Net LoC estimate (product code): ~260–360 LoC

1) Scope full onChat streaming to the active workspace only

Keep addWorkspace() for registration/aggregator creation, but manage exactly one full chat stream (the workspace currently displayed).

Files/symbols:

• src/browser/stores/WorkspaceStore.ts
  • addWorkspace, removeWorkspace, syncWorkspaces
  • new fields: activeWorkspaceId, activeOnChatWorkspaceId
  • new methods: setActiveWorkspaceId, ensureActiveOnChatSubscription
• src/browser/contexts/WorkspaceContext.tsx
  • call workspaceStore.setActiveWorkspaceId(currentWorkspaceId) in an effect

// shape only
setActiveWorkspaceId(workspaceId: string | null): void {
  this.activeWorkspaceId = workspaceId;
  this.ensureActiveOnChatSubscription();
}

private ensureActiveOnChatSubscription(): void {
  if (this.activeOnChatWorkspaceId === this.activeWorkspaceId) return;
  if (this.activeOnChatWorkspaceId) this.stopOnChat(this.activeOnChatWorkspaceId);
  if (this.activeWorkspaceId) this.startOnChat(this.activeWorkspaceId);
  this.activeOnChatWorkspaceId = this.activeWorkspaceId;
}

Defensive points:

• Assert we never keep >1 active onChat subscription.
• On workspace removal, tear down active stream if that workspace was active.
• Preserve existing reconnect semantics (mode: "since" cursor, full fallback) inside runOnChatSubscription().

2) Replay only from the latest compaction boundary on initial load

Switch onChat full replay baseline from penultimate boundary (skip=1) to latest boundary (skip=0).

Files/symbols:

• src/node/services/agentSession.ts
  • emitHistoricalEvents() history load call + comments
• src/browser/utils/messages/StreamingMessageAggregator.ts
  • replace penultimate pruning logic with latest-boundary pruning behavior

// agentSession.ts
const historyResult = await this.historyService.getHistoryFromLatestBoundary(
  this.workspaceId,
  0 // latest boundary only
);

// StreamingMessageAggregator.ts (shape)
if (this.isCompactionBoundarySummaryMessage(incomingMessage)) {
  this.pruneBeforeBoundarySequence(incomingMessage.metadata?.historySequence);
}

This keeps live behavior aligned with fresh loads: once a new boundary arrives, older epochs are pruned by default.

3) Add explicit “Load More history” API with cursor pagination

Expose a non-stream endpoint that pages older compaction epochs via a stable cursor (not page index).

Files/symbols:

• src/common/orpc/schemas/api.ts
• src/node/orpc/router.ts
• src/node/services/workspaceService.ts
• (optional helper extraction) src/node/services/historyService.ts

Recommended request/response shape:

// shape only
workspace.history.loadMore: {
  input: {
    workspaceId: string;
    cursor: {
      beforeHistorySequence: number; // oldest sequence currently loaded in UI (exclusive upper bound)
      beforeMessageId?: string;      // defensive anchor for mismatch detection
    } | null;
  };
  output: {
    messages: MuxMessage[]; // previous boundary window only (older segment)
    nextCursor: {
      beforeHistorySequence: number;
      beforeMessageId?: string;
    } | null;               // null => no older history available
    hasOlder: boolean;
  };
}

Why cursor (vs skip index):

• Robust against dynamic history changes (compaction/edit/delete) between requests.
• No client/server page-index drift.
• Naturally supports prepend pagination by “load messages older than current oldest loaded row”.

4) Implement Load More UX in chat transcript

Add a top-of-transcript control that prepends exactly one older boundary window per click.

Files/symbols:

• src/browser/stores/WorkspaceStore.ts
  • new per-workspace pagination state: { nextCursor, hasOlder, loading }
  • new action: loadOlderHistory(workspaceId)
• src/browser/components/ChatPane.tsx
  • render “Load More” control above message list

// shape only
async loadOlderHistory(workspaceId: string): Promise<void> {
  const page = this.historyPagination.get(workspaceId);
  if (!page?.hasOlder || page.loading) return;

  const result = await this.client!.workspace.history.loadMore({
    workspaceId,
    cursor: page.nextCursor,
  });

  // Prepend older slice while preserving current window + live tail.
  this.assertGet(workspaceId).loadHistoricalMessages(result.messages, false, { mode: "append" });

  this.historyPagination.set(workspaceId, {
    nextCursor: result.nextCursor,
    hasOlder: result.hasOlder,
    loading: false,
  });
  this.states.bump(workspaceId);
}

Implementation note: if append cannot safely preserve strict chronological order for prepends, add a dedicated prependHistoricalMessages() path in the aggregator and assert sequence monotonicity after merge.

5) Use metadata-only activity feed for non-active workspace status

Leverage existing workspace.activity.list/subscribe as the metadata channel for unread + streaming indicators.

Files/symbols:

• src/browser/stores/WorkspaceStore.ts
  • add activity snapshot map + runActivitySubscription()
  • merge activity fallback into getWorkspaceState() / sidebar derivation
• src/browser/contexts/WorkspaceContext.tsx
  • no new API plumbing required if store owns the subscription lifecycle

// shape only fallback in getWorkspaceState()
const activity = this.workspaceActivity.get(workspaceId);
const canInterrupt = activeStreams.length > 0 || activity?.streaming === true;
const currentModel = aggregator.getCurrentModel() ?? activity?.lastModel ?? null;
const recencyTimestamp = aggregator.getRecencyTimestamp() ?? activity?.recency ?? null;

This preserves sidebar responsiveness (unread dot, streaming state, model tooltip) without paying transcript-stream costs for every workspace.

6) Tests to update/add

Files:

• src/node/services/agentSession*.test.ts (or add targeted replay test)
• src/browser/utils/messages/StreamingMessageAggregator.test.ts
• src/browser/stores/WorkspaceStore.test.ts
• src/browser/components/ChatPane*.test.tsx (or nearby transcript tests)

Test cases:

1. Full replay now starts at latest boundary (skip=0) and emits expected caught-up metadata.
2. Live boundary arrival prunes to latest-boundary window (no penultimate epoch retained).
3. loadOlderHistory() advances cursor page-by-page, prepending one boundary window per click until nextCursor=null / hasOlder=false.
4. Non-active workspaces keep unread/streaming indicators via activity snapshots without full onChat streams.
5. Active workspace switch cleanly moves the sole full onChat subscription and keeps since reconnect behavior intact.

Why this approach over sequential full subscriptions?

Sequential full subscriptions reduce startup spikes but still replay transcript data for every workspace. Using one full stream (active workspace) plus metadata-only activity for the rest cuts both startup and steady-state bandwidth much more aggressively while keeping sidebar signal quality.

Validation plan

• bun run test src/node/services/historyService.test.ts
• bun run test src/node/services/agentSession*.test.ts (targeted replay-focused cases)
• bun run test src/browser/utils/messages/StreamingMessageAggregator.test.ts
• bun run test src/browser/stores/WorkspaceStore.test.ts
• bun run test src/browser/components/Messages/MessageRenderer.test.tsx and/or ChatPane tests (if Load More UI is touched there)
• make typecheck

---

Generated with mux • Model: anthropic:claude-opus-4-6 • Thinking: xhigh • Cost: $0.95

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f2200cbcfc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

Addressed both review comments:

1. Pagination cursor refresh on live compaction — historyPagination is now recomputed in processStreamEvent when a compaction-boundary summary message arrives, so loadOlderHistory() uses the post-prune cursor.

2. Stale activity fallback for active workspace — canInterrupt, currentModel, and currentThinkingLevel now only fall back to activity snapshots for non-active workspaces. The active workspace trusts the live aggregator exclusively.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 48b6215b01

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

Addressed both new review comments:

1. Inactive workspaces now prefer activity snapshots — For non-active workspaces, canInterrupt, currentModel, and currentThinkingLevel now prefer the activity snapshot over potentially stale aggregator state (since inactive workspaces don't receive stream-end events).

2. Fixed zero-based pagination threshold — hasOlder now uses historySequence > 0 instead of > 1, correctly recognizing that sequence 0 is valid and sequence 1 means there's still an older row.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5f09caa963

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 019062e776

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a6687897b9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4103704408

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 405e73e2ae

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 94634553b3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e676f97ebf

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb3a0eaa28

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4e59cbc0f1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

@codex review

Addressed the remaining background-compaction callback feedback and added regression coverage. Please take another look.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".