# dsh-budget: cost governance for DeepSeek Harness.

#

# Aggregated token/cost metering (per model, session, day), session/daily/monthly

# budget caps with threshold alerts (webhook + desktop notification) and

# alert / block / degrade over-limit policies, carbon footprint estimation,

# per-model latency statistics, a Settings budget tab, and the /budget command.

#

# Every key below is a Schemastery-validated config field; an invalid value

# fails loudly at load time. Empty caps mean "unlimited" for that scope.

- insert:

- id: budget

name: dsh-budget

inject: [sessions]

config:

# Per-model USD prices per 1M tokens, merged OVER the built-in table.

# Keys are model ids ("deepseek-chat", "gpt-4o", ...); each entry

# supports input/output and optional cacheRead/cacheWrite prices.

prices: {}

# Fallback price for models absent from both tables.

defaultPrice: { input: 1.0, output: 3.0 }

# Budget caps in USD per scope; omit a scope for unlimited.

budgets:

session: 10

daily: 50

monthly: 500

# Alert once usage reaches this fraction of a cap (0..1).

warnRatio: 0.8

# Behavior after a cap is crossed:

# alert - notify only (webhook/notification/session event)

# block - fail new model requests until unblocked by the user

# degrade - rewrite the model to the configured cheaper tier

overLimit: alert

# Degradation map: model id -> cheaper model id of the SAME provider.

# Example: { "gpt-4o": "gpt-4o-mini" }

degradation: {}

# Optional webhook URL for threshold alerts (POST JSON). Credentials in

# the URL are never logged; use an https endpoint you control.

webhookUrl: ''

# Webhook request timeout in milliseconds.

webhookTimeoutMs: 5000

# Master switch for threshold alerts; the Settings tab can override

# this at runtime until the plugin reloads.

alertsEnabled: true

# Minimum milliseconds between two alerts of the same scope.

alertCooldownMs: 3600000

# Browser desktop notifications for new alerts while the Settings

# budget tab is open (requires browser permission).

desktopNotifications: false

# Settings budget tab polling interval in milliseconds.

refreshIntervalMs: 5000

# Carbon footprint estimation (token-based bridge over the upstream

# GPU-hours model).

carbon:

enabled: true

# Electricity region key: global us eu china india uk france iceland.

region: global

# Power usage effectiveness (datacenter overhead multiplier).

pue: 1.58

# IT energy per token in kWh; default derives from an A100 (400 W TDP)

# serving ~25 tokens/s: 400 W / 3600 s / 25 tok/s = 4.4e-6 kWh, rounded.

energyKwhPerToken: 0.000007

# Per-model latency statistics over the llm/stream wrapper.

latency:

enabled: true

# Samples kept per model for percentile computation.

windowSize: 200

# Display currency (code, units per USD, decimals). Costs are computed

# in USD; this only converts for display.

currency: { code: USD, rate: 1.0, decimals: 2 }

# /budget command output language: en | zh.

outputLanguage: en

# Days of per-day usage history kept in the panel snapshot.

historyDays: 30

# Durable day/month persistence across restarts (storage domain).

# When the storage domain is absent the plugin degrades to in-memory

# aggregation (the pre-persistence behavior).

persistence:

enabled: true

# Milliseconds between persistence snapshots (1000..300000).

intervalMs: 10000