[lawrencecchen]

Summary

HOT PATH AUDIT area F, row H1 (AUDIT.md). Home shows each Chief reply whole by default, so a streamed talk token changes nothing on screen. Yet each token:

• rebuilt and compared the live run of every conversation (HomeLiveBook.shown), then assigned the whole live dictionary;
• woke every open Home view's store observation, and each view's inputs (they carried the reply text) differed, so each ran an O(transcript) projection pass on the main thread.

The fix:

• publishLive updates only the token's conversation (live[c], streams[c], new runShapes[c]).
• HomeRunShape: what a whole-message view draws of the running replies (who writes, the activity line, the thought headline), never the text. Whole-message views observe and compare it, so a talk token does not wake them. Typing dots, the activity line and the headline still wake them.
• Streaming mode (HomeReplyDisplay.streaming) reads the text as before.

New perf-hot path id home-stream-token (scripts/cmux-next/perf-hot-benches.json).

Testing

Red test first: HomeWholeMessageTokenCostTests (MessagesLab). It streams 1,500 tokens of the Chief's live run over a 2,000-message history with 1 and 4 views open. It expects 0 store-observation wakes and 0 projection passes, and prints HOMETOKEN lines (ms per token on the main thread).

Observed / Not run

• Observed: compile-check 21869a1f80d8 (red head 7e48dc2) and 5d2989681d5e (fix head f3fc589) passed.
• Not run yet: the red step at 572c1cd and the green step at f3fc589 (package-test-lane.sh suite Packages/Shared/CmuxMessagesLab HomeWholeMessageTokenCostTests). The branch waits for a Mac lane line (STATUS 9221). The existing Home streaming suites (HomeStreamingPerfTests, HomeStreamStressTests, HomeMidStreamSendTests, DragWhileStreamingTests) are not run yet either.

Changelog

• Changed: a Chief reply that streams no longer slows typing and scrolling in Home.

No capture: no visible change; only main-thread work per token goes down.

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Stops a streamed talk token from slowing typing and scrolling in Home. Home shows each Chief reply whole by default, so a token changes nothing on screen — yet before, every token rebuilt and compared the live state of every conversation, woke each open view's store observation, and ran an O(transcript) projection pass on the main thread.

Changes

• publishLive now updates only the token's conversation (live[c], streams[c], new runShapes[c]).
• Adds HomeRunShape (who writes, the activity line, the thought headline — never the text); whole-message views observe it, so a talk token leaves them asleep while typing dots, the activity line and the headline still wake them. Streaming mode reads the text as before.
• Empty shapes are stored as no entry, so thought tokens before a run has any reply text wake no view either.
• Adds perf-hot bench home-stream-token and HomeWholeMessageTokenCostTests (1,500 tokens, 2,000-message history, 1 and 4 views) expecting 0 wakes and 0 projection passes, plus a test for thought-before-reply tokens.

^{Written for commit 3b67a48. Summary will update on new commits.}

[lawrencecchen]

Proof steps (fleet):

• Red at 572c1cd (compiled at f3fc589): job e7dcb7a9fe2d, aws-m4pro-1, FAILED as expected. 1 view: wakes 1500, refreshPasses 1500 for 1500 tokens; 4 views: 6000 and 6000. Main thread per token, 4 views: p50 0.235 ms, p95 0.643 ms, max 118.6 ms.
• Green at f3fc589: job 31afca435e8d, aws-m4pro-3, all 6 suites passed (HomeWholeMessageTokenCostTests, HomeStreamPacerTests, HomeMidStreamSendTests, HomeSendOrderTests, DragWhileStreamingTests, HomeStreamingPerfTests). Wakes 0, passes 0. 4 views, per token: p50 0.052 ms, p95 0.064 ms, max 0.099 ms.
• Note: the store loads a 60-message window of the 2,000-message history, so the old per-token pass was O(60 rows) here. With more loaded rows it costs more.
• These ms are from unit-test runs on a shared worker; they are directional. The deterministic counts (wakes, passes) are the proof.

[lawrencecchen]

Review (hpa-stream-reviewer-1): FIX at f3fc589
reviewer-session: eff63ca9-f8cc-4bd1-b636-2e789d1a2119
author-session: a5c6b7c9dacc183ab

The design is correct, but one line adds wakes that the base did not have. It is a one-line fix plus a test.

F1 (major, perf regression on the same hot path): Packages/Shared/CmuxHomeCore/Sources/CmuxHomeCore/Store/HomeStore+Live.swift:61
if runShapes[conversation] != shape { runShapes[conversation] = shape.isEmpty ? nil : shape }

• The left side is HomeRunShape? and the right side is a non-optional HomeRunShape. When the stored value is nil and the new shape is empty, Swift compares .none != .some(HomeRunShape()). That is always true, so the code writes nil over nil.
• Observation does not de-duplicate writes, so each of these writes fires runShapes. runShapes is one observed property, so every whole-message view of every conversation wakes. Each wake runs a main-actor Task, a full HomeProjectionInputs (with rows.first, which is O(inbox)), and a new observation registration.
• Concrete case: the thought deltas at the start of a turn (kind != "talk" in HomeLiveMapping), before the run has an activity line or talk. HomeLiveBook.shown(in:) returns nil (it needs activity != nil || hasTalk) and streamRows is empty, so the shape is empty and nothing is stored. Every thought token then does a nil-over-nil write. On the base, these tokens changed neither live nor streams and woke no view. On this head, each one wakes every open whole-message Home view. A Chief turn usually starts with thinking, so this happens on most turns.
• Other calls that publish an empty shape do the same, for example a draft-book change that leaves no visible row.
• Fix: compare with the value that would be stored. Either let stored: HomeRunShape? = shape.isEmpty ? nil : shape; if runShapes[conversation] != stored { runShapes[conversation] = stored }, or if runShape(in: conversation) != shape { ... }.
• Regression test: in HomeWholeMessageTokenCostTests, or next to it, stream N thought deltas (isTalk: false, a new run, no activity yet) into a whole-message view and expect observedWakes to be 0. It fails on this head and passes after the fix.

What I checked and found correct

• publishLive for one conversation only: every change to liveBook or draftBook is to the conversation that is passed in.
  • apply uses event.conversation. A snapshot rebuilds only that conversation's runs.
  • settleLive(c) and liveBook.posted / draftBook.posted use message.conversation. draftBook.apply uses update.conversation, and typingOff(id, _) uses id.
  • The global touch() tick is compared only inside one conversation in shown(in:). drafts is still assigned whole.
  • Nothing else writes live, streams or the books.
• Nothing a whole-message view shows goes stale:
  • refresh reads liveRun only for activity and thought (the live line).
  • typists(in:) reads only live[id].participant, which is the shape's runner.
  • replyTypists reads only stream turn and author, which are the shape's replies. Their visibility (the whitespace filter in streamRows) is already in the shape.
  • core.streams and streamsDiffer return nothing in .wholeMessages.
  • The hand-off is still woken by transcriptVersion and by the shape's replies changing.
  • canRetryChiefTurn (HomeProjection+Retry.swift) and DebugHome read the store when they are called and are not observers.
  • Nothing in ios/ or elsewhere in Packages/ observes live or streams.
  • core.replies is set only at init, so the observe branch cannot go out of sync with HomeProjectionInputs.
• The test is sound:
  • The real view runs over the real store on the real chief-live-v1 path.
  • observedWakes is counted inside the real observation Task.
  • The red run (1500/6000 wakes and passes, the same 3 yields) shows that the yields let every wake run, so the green 0 is not a vacuous pass.
  • It asserts both wakes and projection passes.
• Repo rules:
  • runShapes is derived in the same publishLive from the same books as live and streams, so it is not a second owner of state.
  • There is no timer or debounce.
  • HomeRunShape is a value type with instances, not a caseless namespace.
  • The perf-hot marks are registered in perf-hot-benches.json.

Not run: I did not compile or run tests (no local builds). I relied on the fleet results that were given: red e7dcb7a9fe2d, green 31afca435e8d.

[lawrencecchen]

Review F1 fixed at 68ec51f.

• Red at 612ecea (the new test thoughtTokensBeforeAnyReplyWakeNoWholeMessageView, compiled at the head): job 645cb4522982 FAILED as expected: HOMETOKEN thought tokens 200 wakes 200.
• Green at 68ec51f: job c7276cf318b8, all 6 suites passed: thought tokens 200 wakes 0; talk tokens 1500: wakes 0, passes 0 (1 and 4 views).
• Compile-check bcfc25206872 passed at 68ec51f.
• Fix: publishLive compares the shape as stored (nil when nothing runs).

[lawrencecchen]

Review (hpa-stream-reviewer-1): SHIP at 68ec51f
reviewer-session: eff63ca9-f8cc-4bd1-b636-2e789d1a2119
author-session: a5c6b7c9dacc183ab
R9: pass
perf claim: measured c7276cf318b8 (green) vs e7dcb7a9fe2d (red, talk tokens) and 645cb4522982 (red, thought tokens): wakes and passes per talk token 1/view -> 0; wakes per thought token before any reply 1/view -> 0

Findings:

• F1 from my review of f3fc589 is fixed. HomeStore+Live.swift:61-63 now compares with the value it will store (stored = shape.isEmpty ? nil : shape). An empty shape over no entry no longer writes, so it does not wake runShapes observers.
• The new test thoughtTokensBeforeAnyReplyWakeNoWholeMessageView covers the real case. It sends thought deltas (isTalk: false) through the real chief-live-v1 path. The run then has no talk and no activity, so shown(in:) is nil, the stream rows are empty and the shape is empty. It counts observedWakes in the real observation Task.
  • Red job 645cb4522982: 200 wakes for 200 tokens on the base of the fix.
  • Green job c7276cf318b8: 0 wakes.
• The rest of the whole diff 0afe0a8..68ec51f is unchanged since my first review, and I still find it correct.
  • publishLive changes only the conversation that is passed in. Every change to the live book or the draft book is to that one conversation.
  • A whole-message view still shows everything it shows on base. That covers the typing dots, the live line (activity and thought), replyTypists, the hand-off through transcriptVersion and the shape's replies, Retry eligibility and DebugHome. The last two read the store when they are called and do not observe it.
  • runShapes is derived in the same function, from the same books as live and streams, so it is not a second owner of state.
  • There is no timer or debounce.
  • HomeRunShape is a value type with instances, not a caseless namespace.

Not run: I did not compile or run tests (no local builds). Fleet evidence: red e7dcb7a9fe2d and 645cb4522982, green c7276cf318b8 (6 suites), compile bcfc25206872.

[lawrencecchen]

Merged feat-cmux-next (after #33947 landed, lead decision 6) at dbcc0a5. The only conflict was scripts/cmux-next/perf-hot-benches.json; it now keeps both entries (acpmux-wait-wake from #33947, home-stream-token from this PR).

• Compile-check 3dbff13c8801 passed at the merged head.
• Green 2e215880271b had no verdict (worker cmuxs-Mac-mini-3: "Could not resolve host: github.com" at checkout). Rerun once with tag infra-rerun: d58773ffcb3f on aws-m4pro-3, all 6 suites passed (talk tokens: 0 wakes, 0 passes for 1 and 4 views; thought tokens: 0 wakes).

[lawrencecchen]

Review (hpa-stream-reviewer-1): SHIP at dbcc0a5
reviewer-session: eff63ca9-f8cc-4bd1-b636-2e789d1a2119
author-session: a5c6b7c9dacc183ab
R9: pass
perf claim: measured d58773ffcb3f/c7276cf318b8 (green) vs e7dcb7a9fe2d/645cb4522982 (red): wakes and passes per talk token 1/view -> 0; wakes per thought token before any reply 1/view -> 0

Findings:

• I reviewed this head again on its own; it is not a carried verdict. dbcc0a5 is a plain merge of 68ec51f (my SHIP) and origin/feat-cmux-next at 63381ba.
• The PR's own diff did not change. I compared git diff 63381baf630 dbcc0a52a51 with git diff 0afe0a889c45 68ec51fe45a. They are the same, with one exception: one hunk header in HomeProjection.swift moved from line 121 to line 126.
• The conflict resolution in perf-hot-benches.json is correct. The file is valid JSON, and it keeps feat's entries (including acpmux-wait-wake) and this PR's home-stream-token.
• The code that feat brought in does not affect the review. In the areas this PR touches, it adds Home approval cards (HomeProjection+Approval.swift, HomeApprovalRowProvider.swift and others). This code does not observe the store or read live, streams, typists or runShapes. It adds no new reader of the values that whole-message views stopped observing.
• My findings on 68ec51f still apply.

Not run: I did not compile or run tests (no local builds). Fleet evidence: compile 3dbff13c8801, green d58773ffcb3f (6 suites). Job 2e215880271b gave no verdict because of a DNS failure at checkout, and I do not count it.

[lawrencecchen]

Checks at dbcc0a5: changes and action export areas show fail, but both jobs ended cancelled ("The operation was canceled" during actions/checkout), so they are infrastructure, not this PR's code. The other checks are pending. Receipt: #33945 (comment)

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review dbcc0a5: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr33945 -f ref=cp/hpa-stream-home-token -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4169666218001 (f983340e52)
• cmux NEXT nightly 1.0.0-nightly.4169913779901 (821716ba6e)
• cmux NEXT nightly 1.0.0-nightly.4170266726001 (ad2c76731f)

[lawrencecchen]

Swift logic follow-up (tip watcher swift-logic row, STATUS 13027 "SWIFT LOGIC CULPRITS"; lead order).

This PR raised swift_logic_lines by +74 at its merge commit 552e353: scripts/cmux-next/swift-logic-lines.py before and after the merge, with the same table (the tip's, c6be943). The largest declarations of the rise:

• struct HomeRunShape in Packages/Shared/CmuxHomeCore/Sources/CmuxHomeCore/Store/HomeRunShape.swift: +38 lines counted as Swift logic (no row in plans/cmux-next/swift-logic-classes.tsv)
• extension HomeProjection in Packages/Shared/CmuxMessagesLab/Sources/MessagesLabHome/Cmux/HomeProjection+Observe.swift: +12 lines counted as Swift logic (no row in plans/cmux-next/swift-logic-classes.tsv)
• extension HomeStore in Packages/Shared/CmuxHomeCore/Sources/CmuxHomeCore/Store/HomeStore+Live.swift: +10 lines counted as Swift logic (no row in plans/cmux-next/swift-logic-classes.tsv)
• struct HomeProjectionInputs in Packages/Shared/CmuxMessagesLab/Sources/MessagesLabHome/Cmux/HomeProjectionInputs.swift: +8 lines counted as Swift logic (no row in plans/cmux-next/swift-logic-classes.tsv)

plans/cmux-next/layer-ownership.md says Rust owns every state, list and rule and Swift is the platform shell. Please move the new logic of this PR in a follow-up: to Rust (the crate that owns the area; the owner lists in STATUS 12147-12165 name the target per area), or at least out of the app into its package. If a declaration above is platform code (AppKit, Core Animation, OS calls, strings, tokens, DEBUG tooling), reply here and the table gets a PLATFORM row with your reason instead.

The swift-logic row is a report, not a gate: nothing blocks on this comment.