LLM Gateway Cost & Governance Desk

Overview

Kelly LLM Gateway is a Busabase Cloud App-in-Skill. Its canonical product surface is the AirApp in Busabase, not a separate local-data product. The same Hono source supports an explicitly requested local preview with OAuth connection bootstrap. It gives a platform team an operator dashboard over a shared LLM gateway: several consuming services (e.g. Support Bot, Search Ranking, Content Summarizer, Internal Copilot) routed through one gateway to a mix of internal and external models. It aggregates per-service/per-model call volume, cost, and error rate: an Overview (spend trend, rollout/anomaly summaries), a Cost Breakdown table, a Canary Rollout status board, and an Anomaly list.

This is deliberately generic and brand-free: no real company, product, or model name appears anywhere in the code, config, or seed data — only role-based service names ("Support Bot") and generic provider/model labels ("Provider A / Model Large", "Internal Model v2").

This is a direct-manipulation operator dashboard, not a review-then-approve queue: there is no AI-authored draft to approve and no separate execute/ decisions step. Anomalies are computed by a documented, deterministic rule-based function (content/kelly-llm-gateway-app/app/js/gateway-model.js, ported from the retired content/kelly-llm-gateway-app/server/anomalies.ts); the human platform operator makes every promote/ rollback/hold and acknowledgement decision directly in the UI, writing straight onto the route's own Busabase record — the same way kelly-lead-funnel's kanban stage moves work.

Default behavior is AirApp-first. Unless the user explicitly asks only for explanation, give the user the clickable AirApp URL. Start localhost only when local preview/debugging is explicitly requested; it uses the same Busabase resources and never offers another data provider. Use chat-only mode only when the user says "纯聊天", "chat only", "不要打开 UI", or similar.

Mandatory Dependencies

1. Read and follow $kelly-app-skill-creator for product behavior, visual

quality, responsive layout, and the complete canonical content/kelly-llm-gateway-app/ artifact.

1. Read and follow $busabase for connection, target Space, node discovery,

ChangeRequests, review, and merge behavior.

1. Read and follow $busabase-app-creator for resource modeling, AirApp

runtime limits, security, validation, and deployment.

If a dependency is unavailable, preserve this skill's artifact and product contracts, stop before the unavailable Busabase operation, and report the exact missing dependency. Do not invent a second data backend.

App UI Screenshots

<table> <tr> <td width="50%"><img src="assets/screenshots/overview.webp" alt="Gateway overview"></td> <td width="50%"><img src="assets/screenshots/cost-breakdown.webp" alt="Gateway cost breakdown"></td> </tr> <tr> <td><strong>Overview</strong><br>Total daily spend trend, a canary-rollout summary, and a top anomalies preview.</td> <td><strong>Cost Breakdown</strong><br>Sortable service × model table: calls, cost, error rate, canary %, status.</td> </tr> <tr> <td width="50%"><img src="assets/screenshots/rollouts.webp" alt="Gateway rollouts"></td> <td width="50%"><img src="assets/screenshots/anomalies.webp" alt="Gateway anomalies"></td> </tr> <tr> <td><strong>Rollouts</strong><br>Canary-rollout status board with rollback readiness and promote/rollback/hold actions.</td> <td><strong>Anomalies</strong><br>Deterministic cost/error spikes vs each route's own rolling baseline, with acknowledgement.</td> </tr> </table>

Boundary

• Deterministic, rule-based anomaly detection only

(content/kelly-llm-gateway-app/app/js/gateway-model.js). NEVER call an LLM to detect, rank, or auto-resolve an anomaly.

• The AirApp reads and writes its own Busabase Bases only; it never calls a

live gateway API or touches a real routing config. There is no execution/ merge step beyond the direct write itself — a human still applies the decision in the real system of record.

• Routes/services/models are never created by the AirApp; they enter

Busabase through an external process (a future gateway usage-API adapter), the same way kelly-lead-funnel's leads enter through an upstream sourcing process the app doesn't control. The AirApp only ever updates an existing route's rollout status or anomaly-ack fields.

• Generic, brand-neutral tool: never hardcode or reference a specific real

company, gateway, or model name in code, templates, or docs.

• Treat cost/usage data as sensitive. Never commit a local credential file or

Busabase secrets.

Busabase Resources

Four Bases under one application Folder (kelly-llm-gateway), declared in content/kelly-llm-gateway-app/app/js/config.js and the generated template sidecars under content/:

• routes: one row per service→model route — canary status, canary-pct,

rollback-ready, a human decision note, a 14-day daily usage series (JSON array of {date, calls, cost, errors}), and per-anomaly-kind acknowledgement state (cost-spike-ack/error-spike-ack, JSON or empty). Today's totals and each route's own rolling cost/error baseline are never stored — they are pure/derived from daily and recomputed on every read.

• services: one row per consuming service routed through the gateway.
• models: one row per backing model/provider behind the gateway.
• settings: sanitized config (base currency, anomaly thresholds, non-secret

gateway region/base URL/credential-env-var name), one row keyed by kind.

Resources provision lazily through an idempotent Busabase ChangeRequest the first time the app runs in a Space; see references/gateway-schema.md for exact field shapes.

Anomaly Detection

content/kelly-llm-gateway-app/app/js/gateway-model.js (computeAnomalies) flags a cost spike and/or an error spike per route by comparing today's cost/errorrate against that route's own rolling baseline (mean of the preceding days in daily, excluding today). Default thresholds: costspikethresholdpct: 50, errorspikethreshold_pct: 100 (i.e. cost ≥1.5x baseline, or error rate ≥2x baseline), each overridable via the settings Base. Severity is high at 2x the threshold, otherwise watch. No randomness, no ML — the same snapshot always produces the same anomalies.

Direct Rollout & Anomaly Actions

All human actions write straight onto the route's own Busabase record through busabase-sdk, exactly like kelly-lead-funnel's kanban stage moves — there is no approval queue and no separate decisions bucket:

• Promote: sets status: stable, canary-pct: 100, `rollback-ready:

false`.

• Rollback: sets status: rollback, rollback-ready: false.
• Hold: sets status: hold only.
• Acknowledge an anomaly: writes {note, acknowledged_at} onto the

route's cost-spike-ack or error-spike-ack field, whichever the anomaly's kind refers to.

From a standalone local preview the write merges immediately (trusted operator); from the deployed AirApp it creates a pending ChangeRequest for the trusted process to merge, per the AirApp boundary in $busabase-app-creator.

Demo Mode

• ?demo=1 opens a deterministic, fully offline mock gateway (4 services, 5

models, 8 service/model routes, 14 days of history) with computed spend trend, rollout status, and anomalies for documentation and screenshots.

• ?demo=spend, ?demo=rollouts, and ?demo=anomalies select named mock

scenes (their initial route).

• lang=en or lang=zh forces UI chrome language for screenshots.
• Demo mode never reads or writes Busabase and never claims a real

connection; demo rollout/ack actions only update the in-memory snapshot already rendered in the browser tab, never Busabase.

UI language: support English and Chinese chrome with Auto default.

Local App

Default behavior is AirApp-first — give the user the clickable AirApp URL. Start pnpm --dir content/kelly-llm-gateway-app dev only when local preview/debugging is explicitly requested.

Views

• #/overview: total daily spend trend (14 days), a canary-rollout summary,

and a top anomalies preview.

• #/spend: sortable service × model cost-breakdown table (calls, cost, error

rate, canary %, status).

• #/rollouts: canary rollout status board — canary %, rollback readiness,

and promote to 100% / rollback / hold actions with a note.

• #/anomalies: cost/error anomalies vs each route's own rolling baseline,

with acknowledgement.

• #/settings: sanitized setup summary — data provider, config path, gateway

region/base URL, credential env var name, and onboarding state. Never expose secret values.

Completion Criteria

Finish only when:

• the skill contains the complete canonical content/kelly-llm-gateway-app/ project and

pnpm --dir content/kelly-llm-gateway-app dev remains supported;

• all persistent config, state, and domain data use busabase-sdk and the

declared resource map — no local JSON, browser storage, or provider choice;

• Vault values and API credentials never reach browser-visible surfaces;
• local setup offers Cloud/custom URL OAuth plus the explicit Demo path,

while a deployed AirApp uses its ambient session;

• Overview, Cost Breakdown, Rollouts, Anomalies, and Help & Settings render

on desktop and phone widths;

• pnpm --dir content/kelly-llm-gateway-app run check and node --test pass.

Stop Conditions

Stop before consequential Busabase mutation when the target Space is ambiguous, the current user lacks permission, or a same-slug resource is not application-owned. Never call a live gateway API or change a real routing config from the AirApp.