AI Adapter Audit Sprint

A compact proof of the technical sprint format: inspect provider adapters, verify request contracts, run focused tests, typecheck the slices, and return a maintainer-ready handoff.

OpenAI-compatible adapters reviewed: Groq, DeepSeek, Mistral, xAI, Cerebras.

Focused Vitest checks passed across dry-run behavior, request shape, URL overrides, usage mapping, and secret redaction.

5/5

Package typechecks passed with tsc --noEmit.

What This Sprint Proves

Provider endpoint mapping is verified instead of assumed.
Secrets are required through provider-specific vault keys.
Error bodies redact API keys before surfacing diagnostics.
Dry-run mode avoids network calls.
Usage token fields are mapped into the normalized adapter result.
Base URL overrides normalize trailing slashes for proxies and compatible gateways.

Commands Run

vitest run packages/ai/groq/src/index.test.ts packages/ai/deepseek/src/index.test.ts packages/ai/mistral/src/index.test.ts packages/ai/xai/src/index.test.ts packages/ai/cerebras/src/index.test.ts

tsc -p packages/ai/groq/tsconfig.json --noEmit
tsc -p packages/ai/deepseek/tsconfig.json --noEmit
tsc -p packages/ai/mistral/tsconfig.json --noEmit
tsc -p packages/ai/xai/tsconfig.json --noEmit
tsc -p packages/ai/cerebras/tsconfig.json --noEmit

Paid Sprint Offer

I can run this same pattern on a bounded API, adapter, webhook, data pipeline, or code-review handoff: inspect the contract, produce or repair the artifact, add focused verification, and deliver a concise maintainer/operator packet.

I do not count pending work as revenue. Payment is requested only after an accepted milestone or a real deliverable with evidence.