Aller au contenu principal

AI SDK (reference substrate)

ai-sdk is the chunk-shape substrate this guide pins. The reason the pin exists, and what the SDK provides for free, is named in foundations.

This page is the implementor's annex to AI SDK's own docs. The SDK's documentation already covers usage end-to-end; only the things the RFC needs that aren't (or aren't prominently) in AI SDK's docs live here. Keep this short on purpose — when the SDK adds a concept, do not re-document it; link out.

Token usage — the cache normalization rule​

The guide's per-turn token breakdown (session / context-window tracking) records five components: input, output, reasoning, cache_read, cache_write. The AI SDK's usage object surfaces these via inputTokens, outputTokens, reasoningTokens, cachedInputTokens (read), and the provider-specific cache-write field.

The non-obvious detail every implementor hits:

AI SDK v6's inputTokens already includes cache-read and cache-write. To isolate the non-cached input portion the recorder MUST subtract:

input = inputTokens - cache_read - cache_write

Without the subtraction, the persisted prompt_tokens double-counts the cached portion and the session's running total drifts from the provider's billing view. The rule is silent in AI SDK's docs because the SDK treats inputTokens as the unified count; the RFC's per-component breakdown forces the split.

The other four components map straight through:

RFC columnAI SDK source
prompt_tokensinputTokens − cache_read − cache_write (the rule above)
completion_tokensoutputTokens
reasoning_tokensreasoningTokens (0 when the model has no reasoning channel)
cache_readcachedInputTokens
cache_writeProvider-specific. Anthropic exposes cacheCreationInputTokens; OpenAI does not (treat as 0).

The recorder taps onStepFinish (per the AI SDK API) and applies the normalization on each step's usage delta.

Tool-loop helper​

AI SDK ships an opinionated tool-loop primitive — see ai-sdk-core / tool-loop-agent. Implementors taking the adapter path (per foundations / native vs adapter path) can lean on it directly: it handles tool calls, stop conditions, multi-step continuations.

The RFC does not normatively require it. Implementors who take the native path must reproduce the equivalent control flow themselves (tool-call dispatch, stop-condition evaluation, multi-step continuation). Both paths emit the same UIMessageChunk stream downstream.

What the RFC adds on top​

A few load-bearing decisions the SDK is deliberately silent on:

  • Persistence. AI SDK has no opinion on storage. The RFC pins the three-table SQLite shape — see persistency.
  • Compaction. AI SDK does not auto-compact. The RFC's overflow / tail / prune machinery lives in session / compaction.
  • Subagents. AI SDK's tool layer does not include a task tool. The RFC adds it — see subagents.
  • ACP outward wire. AI SDK is host-internal. The RFC's outward wire is ACP — see acp.
  • Skills, MCP, watchdog, sandbox. All RFC additions on top of the substrate.

When in doubt: AI SDK is the wire vocabulary; the RFC is the runtime contract that uses it.

See also​