MIGRATION

Migrating from 1.x to 2.0

2.0 removes every 1.x compatibility shim. Where a removed input is still passed, the library fails loudly with an AIError whose hint names the replacement, so an upgrade cannot go unnoticed at runtime; where a removed output is read, TypeScript reports the missing property.

Package name

The package is now llmwire. @tanvoid0/bot-client 2.x is a shim that depends on llmwire and re-exports every entry (@tanvoid0/bot-client/openai → llmwire/openai) plus the bot-client CLI, so an existing install keeps working; it is kept for one major. To switch: npm rm @tanvoid0/bot-client && npm i llmwire, then replace the import path. The CLI is npx llmwire.

Requests

1.x	2.0	What happens if you keep the old one
`history: [...]`	`messages: [...]` (same shape; `role: 'tool'` and image parts are new)	`INVALID_REQUEST`: "AIRequest.history was removed in 2.0"
`responseSchema` (Gemini only)	`schema` (JSON Schema object or any Standard Schema; every provider)	`INVALID_REQUEST`
`usageContext`	delete it; it was never read	`INVALID_REQUEST`
`prompt` required	`prompt` optional; one of `prompt` / `messages` required	`INVALID_REQUEST` when both are missing

Responses

1.x	2.0
`tokensUsed`, `promptTokens`, `completionTokens`	`usage.totalTokens`, `usage.promptTokens`, `usage.completionTokens`
`processingTime`	`durationMs`
`confidence`, `cost`, `modelCapabilities`, `suggestedImprovements`, `timestamp`	removed; they were never computed
—	new: `object` (when `schema` given), `toolCalls`, `steps`

Stream chunks

Chunks are a discriminated union. Switch on type; only text chunks are the answer.

// 1.x
for await (const chunk of aiFactory.processStream(req)) {
  out += chunk.text;
  if (chunk.done) console.log(chunk.usage);
}

// 2.0
for await (const chunk of aiFactory.processStream(req)) {
  if (chunk.type === 'text') out += chunk.text;
  if (chunk.type === 'reasoning') thinking += chunk.text;
  if (chunk.type === 'tool-call') calls.push(chunk.toolCall);
  if (chunk.type === 'done') console.log(chunk.finishReason, chunk.usage);
}

A custom provider that still yields { text, done } makes the factory throw INVALID_RESPONSE on the first chunk. Yield { type: 'text', text }, { type: 'reasoning', text }, { type: 'tool-call', toolCall } and a final { type: 'done', finishReason } instead. chunksOf(response) turns a one-shot AIResponse into that sequence.

Custom providers

1.x	2.0
`createResponse(success, data, error, model, usage)`	`this.ok(data, extra)` / `this.fail(this.error(code, message))`
`handleError(error)` (throws)	`this.toError(error, model)` (returns an `AIError`)
`buildChatMessages()` returned `{ role, content: string }[]`	returns `Message[]`; `content` may be a parts array, `role` may be `'tool'`; use `textOf` / `partsOf`
`AIFactory` from `.` constructs the five built-ins when given none	unchanged; `AIFactory` from `./core` has no defaults and needs `providers`

Removed types

ConversationHistory, AIProviderConfig, ProviderType, ProviderConfig, ContentGenerationRequest, AnalysisRequest, CodeGenerationRequest, ConversationRequest, PostProcessingOptions, ModelCapabilities, ProcessingMetrics, ChatMessage. None had a reader inside the library.