Skip to main content

v2.1.0 / v2.1.1

Released: May 28, 2026

v2.1.1 is a patch that bumps the version to keep toolpack-sdk aligned with the @toolpack-sdk/agents peer-dependency correction. No functional changes.

Highlights​

  • Vertex AI migrated to @google/genai — replaces the deprecated @google-cloud/vertexai package (deprecated June 2025, removal June 2026); public API unchanged
  • Automatic rate-limit retry — generate() retries transient 429 errors with configurable backoff
  • Per-mode response_format — modes can force json_object output, ideal for evaluator/parser agents
  • Resilient tool execution — a single tool failure can no longer orphan sibling tool calls

Breaking changes​

None. The vertexai provider now depends on @google/genai internally, but VertexAIConfig (projectId, location, googleAuthOptions) is unchanged.

toolpack-sdk​

Per-mode response format​

const evaluatorMode = createMode({
  name: 'evaluator',
  response_format: 'json_object',  // 'text' (default) | 'json_object'
});

Applied to generate() and stream() requests for the active mode. Tool-call rounds are unaffected.

Rate-limit retry​

Retries RateLimitError (HTTP 429) up to 3 times by default. Honors a provider-supplied retryAfter value, falling back to a [10s, 30s, 60s] backoff schedule. Configurable via maxRetries, backoffMs, isRetryable, and onRetry.

Resilient tool execution​

executeBatchWithLimit now uses Promise.allSettled — a single tool failure never drops sibling results. Every tool_call in an assistant turn always gets a matching tool_result.

Fix: Gemini/Vertex AI JSON mode with tools​

responseMimeType: 'application/json' is now applied only when no tools are present. Gemini/Vertex do not support JSON mode and function declarations simultaneously.

Install​

npm install toolpack-sdk@2.1.1
npm install @toolpack-sdk/agents@2.1.1