@steve31415/llm-failover

LLM API client with automatic provider failover (Anthropic, Gemini, OpenAI).

Structured output (jsonSchema)

Callers author one portable JSON schema; each provider adapter translates it into that provider's native dialect, so you never write provider-specific schemas:

• Anthropic — forced tool_use; schema passed through as input_schema.
• Gemini — responseSchema, with additionalProperties recursively stripped (Gemini rejects it).
• OpenAI — response_format: json_schema with strict: true, and the schema normalized into OpenAI's strict dialect: every object gets additionalProperties: false and lists all its properties in required (OpenAI strict mode requires both). A property that was optional in the source schema is emulated as a nullable union (type: ["T", "null"], plus null added to any enum) so the model can still emit null to mean "absent". Without this normalization OpenAI 400s on any schema with optional fields, silently killing it as a failover target.

Publishing

Published via npm Trusted Publishing. Push a v* tag to trigger the GitHub Actions workflow, which builds, tests, and publishes automatically.

# 1. Bump version in package.json
# 2. Commit the version bump
# 3. Tag and push:
git tag v<version>
git push && git push origin v<version>

Plasticine Way: ec0578a (2026-06-21)