linkingoscar/dsh-billing-glass

dsh-billing-glass — Liquid-Glass Billing Overlay

English | 中文

社区第三方插件，与 DeepSeek 官方无隶属关系 · Unofficial community plugin, not affiliated with or endorsed by DeepSeek.

A billing overlay plugin for the DeepSeek Harness Web GUI: a liquid-glass capsule pinned to the bottom-right corner showing the provider balance; click to expand a full billing card (session cost, daily spend, token-bucket breakdown, provider list). DeepSeek-first, with an extension point for more API providers.

Features

Tested with Harness 0.2.0-rc.2 (developer prerelease, commit 639ed015397290b3745d163aafe02ffee4aa3f84). API-key usage retains its existing pricing. Harness's deepseek-account quota is separate and is reported as an unrecognized provider; it is never estimated using API-key prices, even on a shared hostname.

• Always-on glass capsule: status dot + provider name + balance at a glance; click to expand. The expanded card's header is draggable (position stored in localStorage; it can be dragged anywhere on the page with no bottom dead zone). The card disables horizontal overflow — no horizontal scrollbar needed to reach content.
• Fresh data: the client polls every 10s; the DeepSeek balance cache TTL is 10s (other providers 60s); official daily spend is cached for 5 minutes; window focus, tab visibility change and card expansion refresh immediately, and the manual refresh button force-refreshes the current provider balance via POST.
• Liquid-glass material: backdrop-filter frost + translucent theme colors + specular highlight border + refraction sheen layer + soft float shadow; follows the --dsw-* light/dark theme automatically.
• Per-message cost badge: each assistant message's action bar shows a small cost chip (hover for the input/cache/output token split and model). On Harness 0.1.6, the billing card follows the main conversation while message chips refresh independently in every open pane, including historical sessions.
• Settings card: a "Billing capsule" page in the settings panel (Harness v0.1.0-rc.7+; hidden automatically on older hosts): toggle the glass capsule / per-message cost chips, and reset the card position in one click. Changes apply instantly and stay in the local browser only.
• Model series tag: the capsule and card show a short alias for the current model — DeepSeek Pro/Flash/Flash-Vision, Moonshot K2.5/K3, GPT 5.6-Sol/5.6-Terra/5.6-Luna, Claude Opus-4, Gemini 2.5-Pro, Qwen 3.7-Max, GLM 5.2; long tags ellipsize instead of stretching the card.
• Spend ledger & stats: an append-only JSONL ledger (storages/billing-glass-ledger.jsonl, idempotent, debounced appends, periodic compaction; the legacy billing-glass-ledger.json is migrated automatically; bad lines and partial tails are detected at startup, repaired and surfaced as a degraded warning). The expanded card shows today / this month / all-time totals, bucketed by the browser's IANA timezone so a UTC server does not shift the day boundary. costUsd is the only aggregation base; display uses costNative + nativeCurrency — the ambiguous single cost field is gone.
• Session cost: every assistant/message is priced against the official price policy, including the September 10, 2026 04:00 UTC V4.1 Flash price change. deepseek-flash and the two legacy Flash names use the new peak/valley prices from that instant; earlier calls retain their historical policy. Pro prices remain unchanged. Live and replay share one canonical attribution pipeline (header > source) and are merged with messageId dedupe; full persistent-log replay (including pre-install history) + live fallback. Harness v0.1.3+ uses read-only SessionHandles and revision caching, including the {eventState, events} result in v0.1.5; older hosts retain raw-log replay. Backends exposing neither read handles nor per-session raw artifacts degrade to live-only. Historical requests also restore the session's provider and model badge after a host restart; new live requests take precedence. Hover ⓘ for the tokens × unit price = subtotal formula.
• Historical price snapshots: the first price applied to each message is persisted with its bucket subtotals and catalog source. Replaying an old session reuses that snapshot, so a later catalog upgrade cannot silently rewrite historical spend.
• Unknown models fail closed: a model missing from the catalog (catalog lag, alias rename, brand-new model) is never silently priced at zero — the message is marked unpriced, the card and spend stats expose unpricedCalls: N, and the ledger stores priced: false.
• Daily spend (official only): the row appears only when DEEPSEEK_PLATFORM_TOKEN is configured. Without it the row is hidden — a balance-delta estimate would be confused by top-ups/refunds, so it is no longer a display source.
• Pricing sync check (button): the card's "plan" row has a ↻ verify pricing button that pulls the official pricing source of only the currently displayed provider and compares it against the built-in policy chain → ✅ synced / ⚠ drift found (details listed, page snapshot saved to storages/billing-glass-pricing-snapshot.html for the assistant to analyze) / unparseable (page changed; guided back to chat). 60s debounce.
• Multi-provider auto-switching: the card follows the provider in current use — the session's latest request/header provider > the harness-configured active provider (Settings → Models, agent-default-model) > the registry's first entry (DeepSeek-first). Click any provider in the bottom list to inspect it manually, click again to restore auto-follow; the configured active provider carries an "active" badge, providers without a configured key a "not configured" marker.
• Plan / fee system: each provider declares its plan (token pay-as-you-go / subscription); the "plan" row shows the billing mode + the current price tier (standard / peak / valley). Catalog models whose token prices are all zero are shown as subscription/quota plans and remain unpriced, never as "free" metered usage.
• Daily spend (official, optional): without DEEPSEEK_PLATFORM_TOKEN the row is hidden. With it, the card shows the exact official consumed value: