dsh-models-radar

Usage

The Model Radar page

1. Open Settings from the bottom-left of the DSH Web GUI.
2. Choose Model Radar.
3. Pick the DeepSWE or Pompeii channel at the top.
4. Select a model tier from the capability overview or the tier selector.

Every page activation refreshes within the freshness windows: cached data costs no upstream requests and only expired datasets are refetched. Clicking any overview row switches the tier used by the efficiency badges, trend, and task diagnostics below.

Capability overview

Each base model shows its currently strongest tier by default; expanding a row reveals the full reasoning-effort ladder. The IQ progress bar uses a fixed 0–110 absolute scale:

IQ	Capability band
< 70	Developing
70–84.9	General
85–94.9	Steady
95–99.9	Excellent
≥ 100	Leader

IQ trend

The last-24h and last-7d tabs are time-sliced views of the same hourly series, each with its own y-axis scaling and full stats (net change, low, average, high). The curve is colored by capability band with a matching translucent area fill; endpoint and hover markers use the band color.

Cost × IQ comparison

Three tabs (composite / time / price) plot every tier on a log cost axis × linear IQ axis: color = base model (site palette), shape = reasoning effort (off=× · low=○ · medium=△ · high=□ · xhigh=◇ · max=⬡ · ultra=★), same-base tiers joined by ladder lines in effort order. Upper-left = more efficient. The model chip row multi-selects, synced across the active tab; the codex-run DSV4 bases are hidden by default, matching the site.

Task diagnostics

DeepSWE uses the upstream's real majority-vote verdicts; Pompeii keeps continuous F1 semantics. Filtering, counting, and sorting all happen locally in the browser — switching filters adds no API requests.

Task titles and language badges come from the site's task catalog (the /table endpoint, 60-minute window, proxied through the Host which keeps only the tasks array): a title with a source repo is an outbound link (GitHub for DeepSWE, dataset page for Pompeii), and the badge follows the site's vocabulary — Py/JS/TS/Go/Rust, unknown languages shown verbatim. A failed catalog fetch only drops the badges and links; the channel view is unaffected.

Capability popover

Click the readout in the composer tool row (left of the model selector) to open the popover: a full base overview (for comparison) on top and the viewed tier's details below (efficiency badges, dual-window trend, task composition). The viewed tier follows the session model by default; clicking an overview row or using the trend card's tier selector views another tier temporarily until the session model changes or the popover closes.

The composer capability capsule

The compact capsule reads the model selected for the session's next request, not a guess from the last completed reply. Display:

SWE IQ 90.2

Match order:

1. Exact model@reasoningEffort match
2. The same base model's highest-IQ tier, prefixed with ≈
3. Hidden entirely when the base is absent from the DeepSWE leaderboard

Switching models in the composer updates the capsule immediately. The readout polls the host every 15 minutes (the shortest freshness window) — one local request per tick and at most one upstream fetch per channel per window; on failure the last successful value is kept.

Prefer no capsule by the composer? The「Show live capability readout」switch on the Model Radar card under Settings → Plugins → Configurable plugins hides it entirely; while hidden the capsule renders nothing and stops background polling. The preference persists in Host settings (shared across browsers, immune to clearing browser storage; a legacy localStorage choice is migrated once on upgrade).