fix(web): keep long streams responsive

[wbxl2000]

Related Issue

No linked issue. This is a focused fix for a reproduced long-stream rendering stall on the current kap-server /api/v1 path.

Problem

When animation frames are suspended, high-frequency WebSocket render events accumulated without coalescing or a processing budget. A later lifecycle event synchronously drained the entire backlog, repeatedly cloning message state and producing a main-thread long task. The WebSocket remained healthy, but the transcript could stop visibly advancing until a refresh rebuilt it from the authoritative snapshot.

In an isolated kap-server session, 20,000 raw deltas became 1,544 Web render events; the old control-event drain processed all 1,544 at once, took 48.6 ms, and produced a 52 ms long task while still preserving the final 80,000-character text.

What changed

• Carry raw turn, offset, and text/thinking identity alongside projected assistant deltas.
• Coalesce only adjacent deltas with the same session, turn, message, content index, kind, and contiguous offset. Cap each group and split oversized frames so concatenation cost stays bounded.
• Process ordered groups in bounded slices with both animation-frame scheduling and a hidden-tab task fallback. Lifecycle, approval, snapshot, and other control events remain ordering barriers and are never dropped.
• Flush around authoritative resync snapshots and dispose both scheduled callbacks during HMR so stale modules cannot mutate old state.
• Discard only the removed session's pending render and control events during teardown, while preserving other sessions' queued order.
• Add public-path coverage for 10,000 deltas, hidden-tab fallback, control ordering, snapshot/resync replacement, text/thinking metadata, sequence watermarks, oversized frames, per-session teardown, and idempotent disposal.

A stronger 400,000-character isolated burst reduced 782 Web enqueue calls to 13 capped reducer groups, with 1.2 ms total enqueue time, 0.8 ms coalescing time, and a 0.4 ms maximum reducer drain. The final text matched the server SHA-256 exactly. The terminal Markdown commit for that intentionally extreme unbroken block still produced one 61 ms task; the streaming queue itself did not, so this change does not introduce a broader UI state rewrite.

Validation:

• pnpm --filter @moonshot-ai/kimi-web test (28 files, 499 tests)
• pnpm --filter @moonshot-ai/kimi-web typecheck
• pnpm --filter @moonshot-ai/kimi-web check:style (baseline findings only)
• pnpm --filter @moonshot-ai/kimi-web build
• kap-server WebSocket/snapshot targeted tests (5 files, 55 tests)
• git diff --check

No user guide update is needed because commands, configuration, and interaction semantics are unchanged; the changeset records the user-visible reliability fix.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue, or explained the problem above.
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.

[changeset-bot]

🦋 Changeset detected

Latest commit: e774b6f

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[wbxl2000]

@codex review

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@e774b6f

npx https://pkg.pr.new/@moonshot-ai/kimi-code@e774b6f

commit: e774b6f

[chatgpt-codex-connector]

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

Reviewed commit: eeded68f70

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[wbxl2000]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: ab9ad560a2

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".