Suppress interim text echoes in Thinking cards

[franksong2702]

Thinking Path

Fixes #2346.

The duplicated text is not primarily a backend streaming bug. The model can emit the same user-visible progress sentence through both reasoning and interim_assistant. The UI should preserve the visible assistant text, but the Thinking card should not mirror the exact same snippet immediately above it.

This PR keeps the existing Thinking card semantics: the card remains available and expandable, while visible interim assistant text remains in the transcript. The display cleanup happens only when the same snippet is already visible to the user.

What Changed

• Track visible interim assistant snippets during the live stream.
• Strip exact visible interim snippets from the live Thinking card display.
• Apply the same suppression when rendering settled transcripts where reasoning and visible assistant text contain the same snippet.
• Add regression tests for both live and settled rendering paths.

Why It Matters

The user should see a readable timeline, not the same progress sentence twice in adjacent surfaces. Keeping the visible interim text and suppressing only the duplicate inside Thinking preserves the useful part of both channels.

Verification

• node --check static/messages.js && node --check static/ui.js
• uv run --python 3.12 --with pytest --with pyyaml pytest tests/test_ui_tool_call_cleanup.py tests/test_regressions.py tests/test_reasoning_show_hide.py
  • Result: 98 passed, 1 skipped

Risks / Follow-ups

• This suppresses exact visible snippets only; it does not attempt fuzzy semantic deduplication.
• It does not change model output, backend reasoning events, or the availability of the expanded Thinking card.

Model Used

Codex GPT-5. AI assisted with investigation, implementation, tests, and PR text.

[nesquena-hermes]

Reading the diff

Walked static/messages.js:432-1232 (interim-assistant handler + new _liveThinkingText / _stripLiveVisibleAssistantEchoFromThinking), static/ui.js:2308-2318 and static/ui.js:5411-5414 (settled path), and tests/test_ui_tool_call_cleanup.py:236-263 for coverage.

The live path tracks visible interim snippets and strips them from the Thinking card render:

// static/messages.js (PR)
function _stripLiveVisibleAssistantEchoFromThinking(text, snippets){
  let out=String(text||'');
  (Array.isArray(snippets)?snippets:[]).forEach(snippet=>{
    const visible=String(snippet||'').trim();
    if(visible.length<20) return;
    out=out.split(visible).join('');
  });
  return out.trim();
}

And the settled path on renderMessages() does the same against the visible assistant displayContent:

// static/ui.js (PR)
function _stripVisibleAssistantEchoFromThinking(thinkingText, visibleText){
  let out=String(thinkingText||'');
  const visible=String(visibleText||'');
  if(!out||!visible) return out.trim();
  visible.split(/\n{2,}/).map(s=>s.trim()).filter(s=>s.length>=20).forEach(snippet=>{
    out=out.split(snippet).join('');
  });
  return out.trim();
}

Why the design holds up

1. Both code paths use a ≥20-char minimum snippet, which avoids over-stripping reasoning that happens to share short phrases ("OK", "Let me think", "Done.") with visible interim text. That threshold matches both directions, which is what you want for a "this is the same thing the user already saw" cleanup.
2. The settled path splits visible text on paragraph boundaries (\n{2,}) before stripping, so a multi-paragraph assistant turn won't try to strip the entire concatenated body in one pass — it strips per-paragraph. The live path doesn't need that because each interim snippet arrives as a discrete SSE event and is pushed once into visibleInterimSnippets.
3. The Thinking card is still rendered; only matching text inside it is suppressed. That preserves the original product decision (reasoning is available, expandable, and distinguishable from progress text) while fixing the visible echo.

Edge cases I'd flag, not block on

• out=out.split(visible).join('') does an exact substring strip. If the model paraphrases the same idea ("Reading the file..." vs "Now reading the file..."), only one will be suppressed. The PR body acknowledges this ("exact visible snippets only; it does not attempt fuzzy semantic deduplication"). Good — fuzzy dedup is much higher risk for over-stripping legitimate reasoning that happens to overlap.
• The same visible snippet can be inserted as interim multiple times across a turn. visibleInterimSnippets.push(visible) accumulates duplicates, but split(...).join('') is idempotent for a single occurrence, so duplicated snippets in the array won't double-strip anything. Cheap but correct.
• Worth double-checking: the live handler at static/messages.js:1204-1220 calls assistantText+=visible before pushing to visibleInterimSnippets. If a render rAF fires between those two statements (the _scheduleRender() further down), the Thinking text could briefly show the echo for one frame. In practice the same frame's updateThinking(_liveThinkingText()) call runs before _scheduleRender() settles, so it should be fine; just calling it out.

Test fidelity

test_live_thinking_suppresses_visible_interim_echoes extracts the interim_assistant listener body via regex and asserts on the relevant call sites. That's the right level — it locks in the data flow (remember snippet → consult on live thinking render) without coupling to formatting. test_settled_thinking_suppresses_visible_assistant_echoes similarly pins both the call site in renderMessages and the helper's 20-char guard.

Verdict

Approach matches the issue scope and the established no-fuzzy-dedup direction. No blocking changes from me.

[nesquena-hermes]

Shipped in v0.51.72 ✅

Stage-365 release shipped at #2358. Merged commit a08dfff into master. Tag v0.51.72 pushed. Closes #2346.

Verified before release:

• Full pytest: 5713 passed, 0 failed
• QA harness: 20/20 pass
• Opus advisor: independently verified all 5 specific checks (helper call-site routing, dedup correctness, empty context_messages short-circuit, live↔settled symmetry, over-suppression safety), APPROVE
• Browser sanity: no console errors on local 8789

Thanks for the contribution! Closing now.