fix(stream): stop fade-text effect from re-fading the whole message on tool-call rewinds

[silent-reader-cn]

Summary

With the fade-text effect setting enabled, Hermes long tasks visibly blink/flash the entire message on every tool call: all visible words fade in, disappear, then fade in again.

Root cause

Tool-call XML (<function_calls>…</function_calls>) appears inline in the assistant text stream. _stripXmlToolCalls strips the tail of an unclosed <function_calls opening tag (regex [\s\S]*$), which temporarily rewinds displayText to a strict prefix of what was already shown. That rewind triggered two cascading rebuilds:

1. _streamFadeNextText detected the prefix mismatch and reset the playout cursor to zero.
2. _smdWrite's self-heal then saw the new text no longer starting with the written text, cleared assistantBody.innerHTML='', and re-created every word as a fresh is-new span — replaying the fade animation on all visible text at once.

Every tool call in a long task hit this, producing the visible flashing loop.

Fix

• _streamFadeNextText: on a detected rewind, shrink the playout cursor to the common prefix instead of resetting to zero. The DOM syncs to the shrunken prefix this frame (dropping the rewind tail).
• _smdWrite + _anchorProseIncrementalNode: when a rebuild is unavoidable, compute the common prefix and set a _streamFadeSilentPrefixChars window.
• _streamFadeRenderer.add_text + _streamFadeAppendText: words inside the silent-prefix window are written as plain text nodes (no animation replayed); only the post-rewind tail fades in.

Result: already-visible text stays put on tool-call boundaries; only genuinely new words animate.

Tests

Added 3 regression tests in tests/test_smooth_text_fade.py:

• playout rewind keeps the common prefix visible (no reset-to-zero)
• rebuild mutes the common prefix (only tail animates)
• _smdWrite self-heal sets the silent-prefix window on rewind

Verified: tests/test_smooth_text_fade.py + test_smd_media_in_stream.py + test_anchor_prose_incremental_finalize.py + test_issue3397_transparent_stream_prose_segments.py — 62 passed. Full suite: 582 passed, 1 pre-existing Windows permission-test failure unrelated to this change.

[greptile-apps]

Greptile Summary

The PR prevents streamed assistant messages from replaying their fade animation when tool-call filtering rewinds visible text.

• Preserves the rendered common prefix when stream playout rewinds.
• Mutes rebuilt fade spans that correspond to previously visible rendered text.
• Adds regression coverage for main and anchor prose rendering, markdown/media offsets, and Windows command-length limits.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
static/messages.js	Updates stream rewind handling and rendered-prefix muting across the main and anchor prose fade paths.
tests/test_smooth_text_fade.py	Adds focused regression coverage for rewind playout, rendered-prefix muting, parser rebuilds, markdown/media offsets, and root isolation.
tests/test_smd_media_in_stream.py	Extends the extracted JavaScript test harness with the new fade-prefix state.

_{Reviews (4): Last reviewed commit: "Merge branch 'master' into fix/fade-text..." | Re-trigger Greptile}

[nesquena-hermes]

Summary

Reading all three changed files at PR head 56c9307, plus the corresponding origin/master files, the rewind detection itself is useful. I found one behavior blocker in the new suppression mechanism: the silent-prefix count lives in the raw source-text coordinate space, while both consumers live in rendered-text space. That mismatch means the PR can suppress the first genuinely new words after a rewind.

Code reference

static/messages.js:4350-4366 computes the budget before rebuilding, while renderer add_text receives only rendered text chunks:

_streamFadeSilentPrefixChars=_silentCommon;
let silentLeft=_streamFadeSilentPrefixChars||0;
if(silentLeft>0){
  silentLeft-=match[0].length;
}

The common length includes Markdown delimiters, link destinations, and MEDIA token bytes. Smd does not pass those bytes to add_text: it emits token callbacks or media nodes and passes only visible prose. The same mismatch exists in static/messages.js:4536-4573 and static/messages.js:4864-4889.

Diagnosis / recommendation

A concrete case is old prose alpha beta rewritten to alpha gamma. The source common prefix includes the four strong-emphasis delimiters, but the renderer consumes only alpha and the separating whitespace. The leftover budget then writes gamma as plain text. That contradicts the intended contract that only the already-visible prefix is muted and the changed tail fades. A common prefix containing a MEDIA token has the same problem because the token becomes a node and does not reach add_text at all.
Please scope suppression to the owning parser or renderer, as the existing media-tail state is scoped, and calculate the boundary in rendered-text space. One practical option is to snapshot the old node text before rebuild, rebuild, compare it with the new node text, then remove is-new only from spans covered by that rendered common prefix. Whichever approach is used, clear the state on reduced-motion, safe-renderer, empty-text, and parser-finalization paths. The current stream-global counter can otherwise survive a rebuild that has no fade consumer and affect a later append.

Verification

The new tests at tests/test_smooth_text_fade.py:589-688 use plain text, and the self-heal test stubs parser writing, so they cannot expose this coordinate mismatch. Add a real vendored-smd regression for strong emphasis, links or code, and MEDIA before the changed tail. Assert the old rendered prefix is plain while the first changed word has exactly one stream-fade-word is-new span. Also cover reduced motion during rebuild followed by motion being enabled, and two concurrent anchor/main parsers to prove suppression cannot cross owners. I did not execute PR-authored tests; this review used read-only source, diff, and assertion inspection.

[silent-reader-cn]

Addressed the coordinate-space blocker — rewind mute now runs in RENDERED-text space.

What changed (commit a6d37b22):

• Replaced the source-space _streamFadeSilentPrefixChars budget with _streamFadeMuteRenderedPrefix(): snapshot the OLD rendered node text before the rebuild, then after parser_write compare it against the NEW rendered text and strip is-new only from spans inside the rendered common prefix.
• Markdown delimiters / link destinations / MEDIA token bytes never reach the fade add_text hook, so they no longer inflate the mute budget — **alpha** beta → **alpha** gamma now fades gamma correctly (source-space counted 10 chars and muted it).
• Scoped to the owning node (main assistantBody and anchor .msg-body handled separately) — no stream-global counter that could survive a rebuild without a fade consumer.
• _rewindPrevRendered is a local variable consumed right after the rebuild; nothing persists across reduced-motion / safe-renderer / empty-text / parser-finalization paths.

Tests:

• test_stream_fade_mute_rendered_prefix_plain_words — plain-word prefix mute
• test_stream_fade_mute_rendered_prefix_markdown_and_media_bytes — the exact blocker regression (source-space budget bug)
• test_stream_fade_mute_rendered_prefix_scoped_per_root — concurrent main vs anchor nodes stay independent
• test_stream_fade_smd_write_self_heal_sets_silent_prefix_on_rewind — updated for the rendered-space snapshot flow

Gate: node -c OK; test_smooth_text_fade.py + test_smd_media_in_stream.py + test_svg_audio_video_rendering.py + test_insights.py → 83 passed, 18 subtests.

One note: the strong-emphasis/link/MEDIA-before-changed-tail behavioral test you asked for lives at the unit level (_streamFadeMuteRenderedPrefix with a fake DOM) rather than driving the real vendored smd parser — the full node-driven harness for this renderer doesn't mount a real DOM. Happy to add a vendored-smd round-trip if you want it.

[silent-reader-cn]

CI lint on a6d37b2 flagged tests/test_smooth_text_fade.py:723 F541 (f-string without any placeholders) — the f""" prefix in _fade_fake_node was only there to escape JS braces. Fixed in commit a72c661b by switching to a plain string with literal braces (equivalent output, passes the ruff E9+F+B gate).

Verified locally: ruff_lint.py --diff origin/master → 0 findings; test_smooth_text_fade.py → 21 passed; test_smd_media_in_stream.py → 31 passed (18 subtests).

[nesquena-hermes]

Shipped in exp-v0.52.186 🎉 Thanks @silent-reader-cn.

Gate at the rebased head: Codex SAFE TO SHIP — verified both rebuild paths (the anchor-prose _smdWrite and the _anchorProseSmdCache path) use a node-scoped RENDERED-text prefix compare (not source-byte), so markdown/link/MEDIA bytes can't over-mute the first genuinely-new word; normal append, reduce-motion, and URL-scheme filtering are all intact; the rewind work is O(prefix). Full 5-shard suite green (bar the 2 known box-baseline env flakes); focused 52/52 including the rewind_remount_mutes_common_prefix and rewind_keeps_common_prefix_visible behavioral oracles. Clean fix for a genuinely annoying full-message blink on tool calls.