[Feature]: Desktop GUI: Add option to keep reasoning panel expanded instead of auto-collapsing

[AlouhaBY]

Problem or Use Case

Is your feature request related to a problem?

When using thinking models (e.g., DeepSeek with reasoning_content), the Desktop GUI shows a reasoning panel that displays the model's thinking process. However, once the model finishes reasoning and starts outputting the final answer, the reasoning panel auto-collapses.

For users who want to follow the model's reasoning alongside the final response, this means they have to manually expand the panel every single turn. It's a friction point that discourages deep engagement with the model's thought process.

Proposed Solution

Add a config option under display: to control this behavior:

display:
  reasoning_auto_collapse: false   # default: false (current behavior)

When set to true, the reasoning panel should remain expanded even after the final response starts streaming, allowing users to read both simultaneously.

### Alternatives Considered

1. A /reasoning slash command parameter like /reasoning stay or /reasoning pin
2. A click-to-pin button on the reasoning panel itself in the Desktop GUI

## Current config (already optimal)
display:
  show_reasoning: true      # already set — shows reasoning
  reasoning_full: true      # already set — shows full content


These work correctly but don't control the auto-collapse behavior.

## Context

- **Platform**: Windows 10, Hermes Desktop GUI (Electron)
- **Model**: DeepSeek (which sends `reasoning_content` chunks before `content`)
- **Why it matters**: Users running thinking models often want to compare the model's reasoning path with its final output to better understand, debug, or learn from the model's decisions.

### Feature Type

Configuration option

### Scope

Small (single file, < 50 lines)

### Contribution

- [ ] I'd like to implement this myself and submit a PR

### Debug Report (optional)

```shell

[daerias]

Technical detail: two distinct bugs in one symptom

Debugged on Hermes Desktop 0.18.0 with DeepSeek V4 Pro (reasoning_effort: high). The reasoning panel doesn't just auto-collapse — it completely unmounts from the DOM after streaming.

Bug 1: Auto-collapse (cosmetic)

ThinkingDisclosure line 59:

const open = userOpen ?? pending

When streaming ends (pending → false), disclosure auto-collapses. Fix: read display.reasoning_full from config, default open to true when set.

Bug 2: Full DOM removal (blocker)

ReasoningAccordionGroup lines 160-168:

const hasContent = useAuiState(s =>
    s.message.parts
      .slice(Math.max(0, startIndex), endIndex + 1)
      .some(p => p?.type === "reasoning" && typeof p.text === "string" && p.text.trim().length > 0)
)

if (!hasContent) {
    return null  // ← Component unmounts entirely
}

After streaming completes, message.parts for the reasoning group lose their text content (parts are cleared/reorganized in the assistant-ui store), so hasContent becomes false and the entire ThinkingDisclosure unmounts. The reasoning data IS persisted in the session DB (reasoning + reasoning_content columns), but the runtime message-parts store drops it.

Suggested approach

1. Fix the message-parts store so reasoning part text survives past stream completion
2. Wire display.reasoning_full config to control default expanded state
3. Both fixes needed — fixing just the auto-collapse without fixing content retention would still show an empty panel

[DDmouth]

Same issue here — Windows 10, Hermes Desktop, DeepSeek models with reasoning_effort: high.

My specific scenario

When Hermes generates long analytical responses (multi-dimensional A-share stock analysis reports, research summaries), the assistant output body gets folded/collapsed:

1. During streaming: Reasoning panel auto-collapses when pending → false. Long output also clipped by ExpandableBlock (>200KB gate from PR fix(desktop): stop a single message from crashing or freezing the chat #47664) to 120px preview.

2. After completion: Reasoning panel fully unmounts from DOM (ReasoningAccordionGroup: if (!hasContent) return null). Cannot re-expand at all.

3. ExpandableBlock behavior: Very long final answer body remains behind 120px clamp. Requires manual expand every time.

Config

reasoning_full: true is NOT wired to desktop UI (message-parts.tsx:59 ignores it).

Fix scope

1. Wire display.reasoning_full to keep thinking expanded
2. Make ExpandableBlock remember expanded state per-message
3. Fix DOM unmount in ReasoningAccordionGroup — keep thinking content in DOM after streaming

Daily frustration for reasoning model users generating long-form content. +1.

[daerias]

Update on the root cause (from the #58931 duplicate):

This isn't just auto-collapse — the Thinking block disappears from the DOM entirely after streaming completes. Line 160-168 in message-parts.tsx:

const hasContent = useAuiState(s => ...)
if (!hasContent) return null  // Component unmounts

When hasContent flips to false post-stream, ThinkingDisclosure unmounts. The reasoning text is gone from the DOM — not hidden, not collapsed, removed. User can't re-expand what doesn't exist.

The display.reasoning_full config key (already present in config.yaml but unwired) would be the natural hook. If reasoning_full: true, the open state should stay true after pending finishes, preventing the unmount.

[yahiaaktham]

This is a highly needed feature for me, I do not want to keep expanding the thinking arrow in the desktop version.

[GottZ]

This was generated by AI during triage.

Summary

One PR addresses #53617. #63714 fixes the completion-time loss of reasoning content by removing an overly broad prefix-based deduplication, but it does not implement the requested persistent expanded state or wire display.reasoning_full into the Desktop UI.

Related pull requests

• fix(desktop): remove overly aggressive reasoning dedup in completeAssistantMessage #63714 best fix — (+16/-10) — partial fix: The diff preserves non-text reasoning parts when the final answer overlaps with them and retains streamed parts for an empty final response, addressing the root cause of the Thinking block disappearing after completion. The keep_open review on fix(desktop): remove overly aggressive reasoning dedup in completeAssistantMessage #63714 should be satisfied before merge by returning parts directly in the empty-final-text branch and adding regression tests for overlapping reasoning/final text and empty completion payloads; the PR still does not control post-stream expansion.

Suggested consolidation

Merge #63714 after the keep_open review's identity-preservation change and focused regression tests are added, because it is the only candidate and fixes the documented completion-time reasoning loss. Keep #53617 open for the remaining display.reasoning_full wiring and expanded-by-default behavior; there are no duplicate PRs to close.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I53617(["issue #53617 (open)"])
    I58931(["issue #58931 (open)"])
    P63714["PR #63714 (open)"]
    P63714 -->|best fix| I53617
    P63714 -->|best fix| I58931
    class I53617 open
    class I58931 open
    class I53617 target
    class P63714 open
    class P63714 best
    class P63714 best
    click I53617 "https://github.com/NousResearch/hermes-agent/issues/53617"
    click I58931 "https://github.com/NousResearch/hermes-agent/issues/58931"
    click P63714 "https://github.com/NousResearch/hermes-agent/pull/63714"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 1 pull request and 2 issues in this complex. Each diff was read against this issue; Assessment working set: 2 kB of PR diffs, 4 kB of issue/PR text, 6 kB of discussion (7 comments), 4 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.