[juliusmarminge]

Problem

While a turn streams, providers that send message.updated on every tick (ACP agents including the dev Replay provider, Cursor, OpenCode) give the client's thread projection a new messages array each tick. Two selectors in packages/client-runtime/src/state/threadDetail.ts, the thread queue workflow and the checkpoint summaries, re-derived and published a new value with identical contents on every tick. Every subscriber that reads only these selectors (the queue sheet on mobile, the diff panel and chat view on web) re-rendered on every streamed update for nothing. The issue measured streaming on iPad with the Replay provider and names this per-update thread publication as one of the causes.

Change

Both selectors now keep the value they last published when the new value is equal, and publish a new one only when something they derive actually changes.

• Queue workflow: compares activeRun, the four flags, and each queued entry's run and message by identity. A queued entry's text, attachments and context all come from those two.
• Checkpoint summaries: if the checkpoints array is the same, the summaries are the same, except for which assistant message each one points at. So a new value is published only when the checkpoints change or a summary's assistantMessageId does.

Scope and approval

Fixes #27325

Triaged via the automated issue flow. This PR covers the per-update republication of these two selectors. The issue's other items are left for follow-up: the work-row shimmer, full markdown re-parse on each update, Liquid Glass compositing, the route-level thread state subscription, and idle display links.

Verification

• packages/client-runtime/src/state/threadDetail.test.ts, derived thread state through a streamed message: the tests apply streamed message.updated text ticks and check that the queue workflow and checkpoint summaries keep their identity. They also check that a queued message edit, a new assistant message for a checkpointed run, and a new checkpoint each still publish a new value. Removing the assistantMessageId comparison makes the checkpoint test fail.
  • Before (tests on fa4e4677c0, fix not applied):

    FAIL … > keeps the queue workflow through streamed text and follows a queued message edit
    AssertionError: expected { activeRun: null, …(5) } to be { activeRun: null, …(5) } // Object.is equality
    FAIL … > keeps checkpoint summaries through streamed text and follows a new checkpoint
    AssertionError: expected [ { …(9) } ] to be [ { …(9) } ] // Object.is equality
    Tests  2 failed | 14 passed (16)
    

  • After (this branch, threadDetail.test.ts and threadWorkflows.test.ts): Tests 38 passed (38).
• vp run check:changed passes.
• No visible surface changes, so there are no captures. These tests show the selectors no longer republish. They do not measure power. On mobile, the composer hook still subscribes to the whole thread, so it still re-renders on each update.
• cannot-verify ios-device: the issue's power measurement needs a physical iPad with battery telemetry. The mechanism is shared client-runtime logic and is covered by the tests above.

Claude Opus 5.5 (Claude Code) and GPT-6.1 Sol (Codex), via the T3 Code issue flow.

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Approved at 093810b

Macroscope's review found this PR approvable — This change avoids redundant queue and checkpoint subscriber updates while assistant text streams. Its effects are confined to client-side derived-state publication, with no changes to product defaults or persisted data.

No code changes detected at dd1e609. Prior analysis still applies.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review.

Walkthrough

Queue workflow and checkpoint summary atoms retain cached values when their derived state is unchanged. Tests cover identity retention during streamed updates and updates after queued-message edits, new assistant messages, and captured checkpoints.

Changes: Thread state identity

Layer / File(s)	Summary
Retain unchanged derived state packages/client-runtime/src/state/threadDetail.ts, packages/client-runtime/src/state/threadDetail.test.ts	The queue workflow atom compares active runs, workflow flags, and queued run and message identities before replacing its cached value. The checkpoint summary atom retains its array when checkpoint references and assistant message IDs remain unchanged. Tests cover identity retention and state updates.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix · Severity of issue fixed: Medium

Suggested reviewers: t3dotgg

Merge Risk: Minimal · up to dd1e6

The selector change is mergeable after normal checks. Its effect on iPad power remains unmeasured, but no correctness regression is established.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title clearly and concisely describes the main performance change: streamed text no longer republishes thread queue and checkpoint selector values.
Description check		The description includes all required sections. It explains the problem, the implementation, scope and issue reference, focused verification results, limitations, and follow-up work.
Linked Issues check		Direct issue [#27325] identifies per-update thread-state publication as a streaming cost and supports focused fixes. This PR retains the queue workflow when activeRun, flags, and queued run/message …
Out of Scope Changes check		The diff changes only the queue workflow and checkpoint-summary selectors in packages/client-runtime/src/state/threadDetail.ts and adds focused tests in `packages/client-runtime/src/state/threadDeta…

Full details: Linked Issues check

Explanation

Direct issue [#27325] identifies per-update thread-state publication as a streaming cost and supports focused fixes. This PR retains the queue workflow when activeRun, flags, and queued run/message identities are unchanged. It retains checkpoint summaries when checkpoints and derived assistantMessageId values are unchanged. Tests cover streamed text, queued message edits, new assistant messages, and new checkpoints. The issue does not require this PR to implement its other contributors or provide physical-device power measurements.

Full details: Out of Scope Changes check

Explanation

The diff changes only the queue workflow and checkpoint-summary selectors in packages/client-runtime/src/state/threadDetail.ts and adds focused tests in packages/client-runtime/src/state/threadDetail.test.ts. The changes support [#27325]. The diff does not change the work-row shimmer, markdown parsing, Liquid Glass compositing, route-level subscriptions, or display-link behavior.

Finishing Touches

Generate docstrings

• Commit to this branch
• Create a new PR

Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Usage-based review receipt

• Mode: Continue automatically
• Reviewed files: 2
• Charged: $0.50
• View usage details

Note

This review was completed with usage-based billing: files reviewed beyond your plan's included limits are billed at $0.25/file. View usage-based billing.

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: fix/issue-27325
• File: .coderabbit.config.ts
• Commit: dd1e609
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Team
• Run ID: ba81c48e-3f74-4cc2-9cb2-4c45b96b856d

Commits

Reviewing files that changed from the base of the PR and between 093810b and dd1e609.

Files selected for processing: 2

• packages/client-runtime/src/state/threadDetail.test.ts
• packages/client-runtime/src/state/threadDetail.ts

Limit details: You’ve used all 10 included reviews currently available.

[dbxe]

Note

🤖 Claude Opus 5.5 responding on behalf of Daniel Briskin

This covers one slice of #27325 (the queue and checkpoint selectors), so could the body say Refs #27325 instead of Fixes? The larger items are still open: the work-row shimmer, re-parsing the whole message on every update, and the route-level thread subscription. I'm running this branch through the same iPad bench tonight and will post the number here.

[dbxe]

Note

🤖 Claude Opus 5.5 responding on behalf of Daniel Briskin

Bench result: no measurable power change while a turn streams, which is what I expected for two selectors.

• Setup: the issue's iPad bench, streaming phase only, 2 ABBA/BAAB blocks. The PR's threadDetail.ts diff was applied onto the issue's baseline dac8b11ddc, because the bench backend is pinned there.
• Result: −5 / −11 mW per block against the baseline, about 1059 mW. That reads same: the bench resolves about 25 mW.

That doesn't argue against merging. It just means the streaming cost lives in the other items. I'm measuring candidates for the markdown re-parse and the route and composer subscriptions next.

[juliusmarminge]

The automated pipeline stopped on this pull request and will not merge it:

someone else pushed to fix/issue-27325 (head is now 093810b); the pipeline does not build on commits it did not review

It stays open for a person to finish or close.

[juliusmarminge]

Evidence

Proven in the real web app using Replay (dev), Chromium, dark theme, and a 1440×1000 viewport. Both revisions started with fresh synthetic data.

Steps: open a completed checkpoint and the diff panel; start another streaming turn; queue and edit a follow-up; approve the Replay action; observe streamed text; finish the turn and expand the new checkpoint diff. The scenario applies an identical synthetic file change on both sides because Replay describes edits without executing them.

	Before	After
Commit	c90b67859a	dd1e609b95
Redundant notifications during streamed text	Queue: 4; checkpoint summaries: 4	Queue: 0; checkpoint summaries: 0
Streaming profile
Queue edit
Completed checkpoint
Recording	Before interaction	After interaction

The live selector subscriptions counted new identities with unchanged serialized contents when assistant text changed. Queued-text edits still published, and checkpoint summaries increased from one to two on both revisions. The React commit observer recorded 14 DiffPanel commits during the measured streaming phase on base and 8 on head; those totals include legitimate lifecycle and tool updates and are not a timing benchmark.

Scenario source · Before profiling data · After profiling data · Sanitized runner records and original recordings

Recordings remove startup footage and unused canvas padding; interaction speed is unchanged. The profiling overlay is observational instrumentation. No visual redesign, CPU improvement, or battery improvement is claimed.

AI-assisted; verified with the tests and tools listed above.

Not shown:

• The native mobile queue sheet was not profiled; this comparison exercised the shared queue selector through the real web client.
• Queue attachment and attribution edits were not exercised; the meaningful queue-update control edited queued message text.
• Reassociation of an already completed checkpoint with a later assistant message was not exercised; Replay produced a new checkpoint and its assistant-message association.
• CPU, battery, and real-provider timing effects were not measured.