dsh-reasoning-effort-slider

English | 中文

Turns the Effort row of DeepSeek Harness's model menu into a Codex-style reasoning-effort slider:

┌────────────────────────────────────────┐
│ Model       DeepSeek-V4.1-Flash 1017 › │
├────────────────────────────────────────┤
│ Effort                        Max    › │   ← native row; its value text is tinted by level
│ ●━━━━━━━━━━━━━━━━━━━━━━━━━◯           │   ← injected: 28px track + white knob
│ ·           ·           ·          ·   │   ← level ticks (reached ones brighten)
│                   Max                  │   ← current level name (live while dragging)
└────────────────────────────────────────┘

Nothing is added to the composer tool row. The collapsed model seat keeps the native model · effort ⌄ typography; only that effort text is tinted.

Interaction

Action	Behaviour
Press on the track	Jumps to the pressed cell
Drag	The knob follows the pointer: knob and fill track the cursor, while the level name, ticks and writes use the nearest level; on release it snaps onto the cell in 80 ms, without bounce
Release	Commits immediately (resting the pointer for ~160 ms also commits)
← → ↑ ↓	One level per press, committed immediately
Home / End	Lowest / highest level
Hover	Tooltip with the current level and “further right costs more tokens”

Writes are deduplicated and coalesced: a drag commits at most a few times, and picking the level already in effect writes nothing. A failure shows its reason under the track and clears after 4 s.

Colour and effects (driven by level position, independent of level count)

Position	Fill right end	Starfield	Level text
First level	blue #4d93f8	none	native (an Off level is never tinted)
Middle	blue → violet #9333ea	none	blue → indigo
Highest	deep violet #4c1d95	26 staggered stars + violet nova layer	violet #8b5cf6

• The fill is linear-gradient(90deg, blue → colour at the current position): always blue on the left, deepening to the right.
• The nova layer and the starfield are clipped to the filled part, so nothing glows to the right of the knob.
• The knob is a white circle as tall as the track (28 px) whose centre is inset by 14 px, so it stays a full circle at both ends.

Why it cannot break the native menu

• It only adds a data-dres-row attribute to the native effort row (taller, wrapping). Native classes, copy and chevron are untouched.
• Anything unexpected is fail-open: no aria-controls on the trigger, a menu that is not the model/effort pair, or a row outside this composer card means “do not inject” — the native menu keeps working.
• Growing that row triggers the menu's own ResizeObserver, which re-measures and clamps the panel back on screen.
• On close or unload the container is removed, the attribute dropped, and every tinted text colour restored.

Data path (native public surface only)

Purpose	API
Read levels	session.modelCatalog() → groups[].models[].reasoning.efforts
Read current value	the modelSelection session projection (useProjection), falling back to the catalog default
Write level	session.selectModel({ sessionId, provider, model, reasoningEffort }) — the same entry the native menu uses

Install

From GitHub:

dsh plugin --profile <profile> add github:sleepinggggggggggg/dsh-reasoning-effort-slider

From a local directory:

dsh plugin --profile <profile> add <absolute path to this directory>

On desktop, use Settings → Plugins → Add plugin with the same value, then reload the page (or restart the app).

Known limits

• Depends on the native menu's semantic structure (aria-haspopup / aria-expanded / aria-controls, two root rows with role="menuitem"). If a host release changes it, the plugin silently does nothing — it never breaks the menu — and needs a follow-up.
• The model menu has to be open to change the level; nothing is shown in the composer tool row.
• A model that advertises a single level (for example with thinking disabled) renders no control, and the native row keeps its original height (setActive drops the growth).
• The model catalog is cached for the page lifetime (a reconnect refetches it). Providers or models added in Settings need one page reload.
• The selection applies to the next request; a running step keeps the level it started with (same as the native menu).

Acknowledgements

The design decisions come from reading three sibling implementations. No code was copied; this repository is an independent implementation.

• HanaAyane/dsh-reasoning-effort — the white knob + track + ticks + level-name visual language, and the modelSelection projection + session.selectModel data path.
• Microqian2th/dsh-codex-effort-slider — the DOM anchor strategy for injecting into the native Effort row (aria-haspopup / aria-expanded / aria-controls / role="menuitem"), plus the blue → violet → deep violet palette and the starfield position curve.
• Bai-2008/a-better-dsh-codex-effort-slider — the trade-offs of taking over the model seat entirely, a route this plugin does not take.

Licence

MIT — see LICENSE.

Files

File	Purpose
index.js	Host half: empty apply(), only there so the bundle enters the profile
client.js	Everything: DOM bridge, portaled slider, colour/starfield logic, dictionaries
cordis.patch.yml	Inserts the bundle row
locale/{zh,en}.json	Plugin-manager copy and control strings
icon.svg	Icon shown in the plugin manager