dsh-effort-slider

A reasoning-effort slider that replaces the shipped model selector in the dsh web GUI composer. One control owns both halves of the model choice: the model list and that model's reasoning effort, which becomes a glowing rail with a starfield behind it.

中文：把 dsh 输入框里的模型选择器换成一条"星光滑轨"——模型列表与思考强度由同一个控件管，切到不支持当前档位的模型时自动落到最接近的一档，不再报错。

dsh plugin add @lolannn/dsh-effort-slider          # from npm
dsh plugin add git+https://github.com/suyunxing/dsh-effort-slider   # from source

Screenshots

The rail is drawn from the model's own reasoning.efforts, so the tick labels change with the model — off / low / high / max on one, low / medium / high / xhigh / max on another — and the knob lands on the level the session is actually using.

What it does

• Takes over conversation.input.model (a single slot, so this shadows the shipped selector rather than sitting next to it) at priority: -1 — a single slot renders the lowest priority first.
• Reads the model catalogue at runtime: the model list comes from catalog.groups (grouped by provider) and the levels are exactly the reasoning.efforts the current model declares (typically off / low / high / max). No hard-coded list.
• Shows the Session's real selection, not the deployment default, by reading the durable modelSelection projection.
• Picking a level or a model calls the session selectModel remote with the explicit { provider, model, reasoningEffort } triple, so the choice applies to the Session and is saved as the default in the background.
• Switching to a model that does not declare the level in hand is resolved on the client first: the closest level that model does declare is sent instead, and the status line says which substitution happened. A model with no reasoning at all is switched to with the field omitted, so the Host applies that model's own default. Nothing errors, nothing is silently dropped.
• The rail is animated: drifting particles, soft ribbon waves, iridescent four-point stars, a breathing star core, and a three-armed galaxy that collapses out of the rail at the top level. Reaching the top fires a shockwave burst, a star-warp starfield and a light-blade flash.
• The composer itself carries the light: a halo wraps the whole input card, with the same hue as the rail. Brightness, ring alpha and glow radius all follow the level, so off is a faint rim and the top level breathes; opening the panel lifts it. The pill only keeps an outline while the halo is on.
• The halo is a setting, not a hard rule. A 外圈光环 switch sits under the rail; turn it off and the glow returns to the pill alone, the way it was before the halo existed. The choice is stored in localStorage under effort-slider:halo, so it survives a reload.
• Models without reasoning efforts are tagged 无档位 in the model list and degrade gracefully once selected: the rail is dimmed and the status line says so instead of pretending to switch.

Surfaces

Slot	Content
conversation.input.model	the trigger (模型 · 档位) and, while open, the popup: model row, grouped model list, effort rail, level scale, status line
conversation.input.model.label	untouched — the shipped non-interactive label stays owner-managed

The popup renders through react-dom's createPortal into document.body, because the composer row can live inside a transformed/overflowing ancestor that would clip a position: fixed child.

Interaction

• Click the trigger, then drag the rail or press ← → ↑ ↓ Home End.
• The rail snaps to the nearest level on release (elastic overshoot).
• The level names under the rail are buttons too — same edge math as the thumb, so each label sits exactly under its stop.
• Click the 模型 row to expand the model list; picking a model carries the current effort across, remapping it when that model lacks the level (see below).
• Outside click or Esc closes. Direction keys are ignored while a text field has focus, so typing in the composer never changes the effort.

Files

• index.js — Host half. The bundle row exists so the browser half has an entry to attach to; it declares no service, tool or config.
• client.js — the whole feature: stylesheet injection, catalogue + projection mapping, canvas engine, controller and the slot registration.
• icon.svg — original artwork (pinched four-point star over a lit rail).

Notes

• All canvas work is requestAnimationFrame-driven and stops on unmount.
• prefers-reduced-motion: reduce disables the burst, the warp and the quake.
• Only transforms, opacity and canvas are animated; no layout thrash.
• The canvas is sized from the card (getBoundingClientRect + ResizeObserver), not from the viewport.
• Opening the popup while already on the top level arms the burst guard instead of firing it, so the celebration only plays when you actually move into the top level.
• The popup opens upward, but flips downward and caps its height when the trigger sits too low for that (landscape phone, small window).

Why a model switch remaps the effort

The Host validates the triple, it does not normalize it. In @deepseek-ai/dsh-llm → resolveCallWithInfo:

const reasoning = info.reasoning;
const requested = defaulted.reasoningEffort;
if (reasoning === void 0) {
  if (requested !== void 0) throw new LlmError(`... does not support reasoning effort "${requested}"`, "UNSUPPORTED_REASONING_EFFORT");
} else {
  const effective = requested ?? reasoning.defaultEffort;
  if (effective !== void 0 && !reasoning.efforts.some((e) => e.id === effective)) throw new LlmError(..., "UNSUPPORTED_REASONING_EFFORT");
}

The comment above it is explicit: "no clamping or aliasing is performed." A model with three rungs and a model with seven therefore reject each other's levels, and the session remote wraps that rejection as RemoteError("session/model-unavailable", …) — which used to surface as a red 切换失败：provider "…" model "…" does not support reasoning effort "…".

So resolveEffort(model, wantedId) in client.js maps the level onto the target model before the request is built:

Case	What is sent	Status line
target declares the level	it, unchanged	nothing
target declares other levels	the closest one	该模型没有「medium」档，已自动切到最接近的「low」
target declares no reasoning	the field is omitted	「<model>」没有思考强度档位，已按模型默认运行

"Closest" is measured on one shared scale rather than per-provider lists, because every adapter names its rungs from the same vocabulary: off · none · minimal · low · medium/med · high · xhigh · max · ultra. A tie resolves to the stronger rung — losing reasoning power is the worse of the two surprises (high → xhigh rather than high → medium when both are one step away). A rung whose id is not on that scale is placed by its position in its own list, and when the id being carried is itself unknown the target's defaultEffort (then its strongest rung) is used instead, because its distance cannot be measured.

The last real choice is remembered in lastEffort, so max → 无档位模型 → back returns to max instead of that model's default: the intent survives the models that cannot express it.

Why every visible label is passed through safeLabel

dsh-pocket installs a document-level click guard for phones: any click on a <button>/<a> whose text content looks like a file path is cancelled with preventDefault() + stopImmediatePropagation() and replaced by a toast ("手机上无法直接打开电脑上的文件"). Its matcher includes /[\w.\-]+\/[\w.\-]+\.\w{1,12}/, which a model id like workbuddy-ai/deepseek-v4.1-flash matches — so the trigger became unclickable on mobile while working fine on the desktop.

Every label rendered as text therefore runs through safeLabel(), which turns / and \ into ·. aria-label and title keep the raw identifier, because they are attributes and the guard only reads textContent.

Where the "current" model comes from

modelCatalog().default is the deployment default and can differ from what a Session really runs (a Session-local override). Spreading it into a commit silently switched the model while the user only meant to change the effort.

So the controller follows the Session's durable projection — the same source the shipped selector reads:

sessions.binding(sessionId).session.projections.faceOf("modelSelection")
// -> ObservableSnapshot<{ lastUsed, next }>, next = pending ?? lastUsed

model/selection carries the user's intent immediately (pending), and request/header records what a request actually used (lastUsed). Two details matter:

• The face is subscribed before the first getSnapshot(); reading a projection nobody observes can return no value at all.
• The catalogue default is a fallback only, used until the projection has a value.

Verification

• conversation.input.model resolves to this plugin (priority: -1, active) and the shipped registrant goes inactive.
• Trigger reads 选择模型与思考强度 — <model> · <effort>; changing the level propagates through selectModel into the projection and back into the label.
• The popup's model list exposes every routable model, grouped per provider.
• Switching medium (on a seven-rung model) to an off/low/high/max model sends low and reports the substitution; switching to a no-reasoning model sends no reasoningEffort at all; switching back restores medium.
• Checked in a real client at 360×780, 360×400 (short) and 1280×800: a real click opens the popup, it stays inside the viewport, the level labels sit under the thumb stops, the selected model is marked, and opening does not replay the top-level burst.