Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import { describe, expect, it, vi } from 'vitest'
import { attachmentContextWindow } from '../attachmentContextWindow'
describe('attachment budget across provider failover', () => {
it('uses the exact fallback model catalog window, including a smaller non-default model', () => {
const catalog = vi.fn((provider: string, model: string) => {
if (provider === 'claude' && model === 'public-small-model') return 32_000
if (provider === 'deepseek' && model === 'public-large-model') return 256_000
return undefined
})
expect(
attachmentContextWindow(
128_000,
'openai',
[
{ provider: 'deepseek', model: 'public-large-model' },
{ provider: 'claude', model: 'public-small-model' },
],
catalog,
100_000
)
).toBe(32_000)
expect(catalog).toHaveBeenCalledWith('claude', 'public-small-model')
})
it('keeps the session model window for same-provider credential failover', () => {
const catalog = vi.fn(() => 1)
expect(
attachmentContextWindow(
128_000,
'openai',
[{ provider: 'openai', model: 'inert-other-model', credentialSlot: 'other-slot' }],
catalog,
100_000
)
).toBe(128_000)
expect(catalog).not.toHaveBeenCalled()
})
it('never widens the primary window and resolves missing metadata by the existing policy', () => {
expect(
attachmentContextWindow(
8_000,
'openai',
[{ provider: 'claude', model: 'public-model' }],
undefined,
100_000
)
).toBe(8_000)
expect(
attachmentContextWindow(
128_000,
'openai',
[{ provider: 'claude', model: 'public-model' }],
undefined,
100_000
)
).toBe(100_000)
})
it('rejects unusable catalog windows rather than enlarging or ignoring the bound', () => {
for (const value of [0, -1, 1.5, NaN, Infinity, Number.MAX_SAFE_INTEGER + 1]) {
expect(() =>
attachmentContextWindow(
128_000,
'openai',
[{ provider: 'claude', model: 'public-model' }],
() => value,
100_000
)
).toThrow('positive safe integer')
}
})
})