@deepseek-ai/dsh-client-ui-provider-quota

English | 中文

Session-header model-usage panel. The browser half contributes one header action into conversation.session.header.actions and presents DeepSeek, KIMI, GLM, and GPT in one compact two-by-two Chinese panel. Startup performs no provider request: the first panel open paints instantly from the host cache and a stale entry revalidates in the background, while Host-side caching, a manual refresh control, and an open-state five-minute poll keep it current. The node half serves the aggregated snapshot at GET /plugins/ui-provider-quota/api/usage (same-origin, so the browser needs no vendor keys and hits no CORS wall), resolving keys read-only from the environment, ~/.dsh/.env, ~/.claude/multi-gateway/config.json, and ~/.zcode/v2/config.json. GPT reads the account already signed in through the installed Codex app-server. A broken or prematurely closed app-server pipe stays contained as that provider's error instead of terminating the Web Host, and the snapshot returns neither the email address nor any credential. The selected DeepSeek key in ~/.claude/multi-gateway/config.json takes precedence over keys for other accounts. Snapshots cache in memory for five minutes and mirror to ~/.dsh/provider-quota-cache.json, so restarts answer from disk instantly and an expired entry is served stale-while-revalidating; ?force=1 collects against the vendors on the spot.

Channel shapes

• DeepSeek — official balance API, rendered as one account-balance value; internal topped-up/granted fields are not exposed in the panel.
• Kimi — sk-kimi-* keys query the Kimi Code /usages endpoint (weekly plus 5-hour rolling windows, booster-wallet facts); plain Moonshot open-platform keys query the money balance endpoint instead.
• Z.ai / GLM — the console monitor quota endpoint; quota only, never money.
• GPT — the installed Codex app-server's signed-in account snapshot. The panel renders only the main account's 10,080-minute weekly subscription window and deliberately does not merge model-specific five-hour buckets.

The plan labels also come from the signed-in accounts: KIMI exposes its official Allegro name, GLM exposes Pro, and GPT exposes Pro; the UI adds the user's 20X entitlement label to GPT.

Each quota window renders as an explicitly labelled progress meter: the full capacity is a visible gray track, and only the consumed share is blue. A zero value therefore stays entirely gray. The exact reset timestamp is shown below each meter in Beijing time. When a provider returns zero usage without a reset timestamp, the panel says that no reset is currently scheduled instead of inventing a time.

The four providers share one panel plane rather than separate rounded cards. Inset horizontal and vertical gray rules divide the 2×2 regions, while the short internal rule in KIMI and GLM continues to separate the five-hour and weekly windows.

Brand images use KIMI's official square app icon, the native icon extracted from the installed Z Code desktop app, and pinned MIT-licensed Lobe Icons assets for DeepSeek and OpenAI. All four brands use one 34×34px clipping frame with an 8px radius. Brand-specific dark or light fills preserve identity, while the shared frame owns the final silhouette.

The header action uses the project's native monochrome data icon. Its vector geometry and current-color rendering match the adjacent agent-mode and chevron controls. It is rendered proportionally at 14px so its outline stays lighter than the label; provider warnings remain inside the panel where their scope is clear.

Model Experience

None, as the node half registers no prompt, tool, message, or provider request and the browser half is display-only.

KV Cache effect

None.

Known Limitations and Deferred Work

• Kimi and Z.ai endpoints are undocumented — both were reverse-engineered from official clients and verified live, but either vendor can change them without notice; the panel reports per-provider errors instead of failing.
• Quota ≠ money for Kimi and Z.ai — only DeepSeek (and Moonshot-platform keys) expose real balances; the Kimi booster wallet reports usage facts only when enabled on the account.
• Key discovery is per-machine — resolution walks a fixed chain of local config files; a key stored nowhere in that chain shows as "no key found".
• GPT is desktop-account scoped — GPT usage requires an installed codex binary with a signed-in ChatGPT account. API-key-only Codex configurations do not expose ChatGPT subscription rate limits.