@jeremysnr/snug-anthropic

Fit Anthropic messages into a token budget. Automatically handles tool_use/tool_result pairing so you never get a 400 error from an orphaned tool message.

import { fitMessages } from '@jeremysnr/snug-anthropic';

const { messages } = fitMessages(conversationHistory, {
  budget: 100000,
  reserve: 4096,
});

const response = await client.messages.create({
  model: 'claude-opus-4-5',
  max_tokens: 4096,
  messages,
});

Messages are prioritised by recency — older messages are dropped first. tool_use and tool_result pairs are kept together automatically; if one half doesn't fit, both are dropped.

Install

npm install @jeremysnr/snug-anthropic

@anthropic-ai/sdk must be installed as a peer dependency.

API

fitMessages(messages, options)

| Option | Type | Default | Description | |--------|------|---------|-------------| | budget | number | — | Token limit | | reserve | number | 0 | Tokens to hold back for the model's reply |

Returns

{
  messages: MessageParam[];  // ready to send
  dropped: MessageParam[];   // what didn't fit
  tokensUsed: number;
  tokensRemaining: number;
}

Part of the snug ecosystem

• @jeremysnr/snug — zero-dependency core primitive
• @jeremysnr/snug-tiktoken — snug with tiktoken, model-agnostic
• @jeremysnr/snug-openai — snug for the OpenAI SDK

Licence

MIT