@kindgi/adapter-model-shared

Shared plumbing for Kindgi's model adapters: what every adapter needs to turn a vendor's API into a @kindgi/capabilities ModelProvider the same way.

Today's engine for sending the requests is the AI SDK's provider packages (@ai-sdk/*), called at the provider-spec level (doGenerate). It never uses the ai package, its agent loop or Vercel's gateway, so no request goes anywhere but the vendor's own API. The engine sits behind one module, @kindgi/adapter-model-shared/ai-sdk, the only entry point whose types are a library's. Swapping it later touches that module and the adapters that hand their models in, nothing else.

@kindgi/adapter-model-shared/ai-sdk is for Kindgi's own adapters. It has no compatibility promise across a change of engine: an adapter you write yourself should implement ModelProvider from @kindgi/capabilities, which stays.

What every adapter gets

• Retries are Kindgi's. One send makes one HTTP attempt. The provider makes up to attempts in all (default 3) on a retryable failure: 408, 409, 429, 5xx, or no response.
  • It waits what the vendor asks (retry-after-ms, retry-after), or backs off exponentially with jitter.
  • A vendor asking for more than 60 s ends the call at once, its wait in the message (… (the vendor asks to wait 120 s)), rather than retrying early into a limit it named.
  • An answer the library can't read (a malformed 200, an empty body) is tried once more, not twice: each answer may be billed.
  • Each attempt is counted (@kindgi/capabilities/attempts): a result's attempts, or attemptsOf(error) for a failed call.
  • An abort, before the first attempt, in flight or during a wait, rejects with the caller's reason.
• Errors are typed. A failed call throws ModelProviderError with a kind:
  • auth: 401, 403;
  • rate-limited: 429;
  • unavailable: 408 (the vendor timing out), 409, 424 (Bedrock's model failure), 5xx, and an answer the library can't read;
  • context-too-long and content-filter, read from the vendor's words on a 400, 413 or 422 ("maximum context length", "prompt is too long", not any mention of "context");
  • invalid-request, including a request the library refuses before sending it;
  • network: no response at all (fetch failed, a connection reset or refused, DNS, a socket timeout). A host refused before any connection (a system-style code that isn't a dropped connection's, such as the runtime's egress refusal EKINDGIEGRESS) is network too, never retried.
  • A redirect the request won't follow (an adapter's credential goes to its endpoint only) is unavailable, never retried. It also has the HTTP status. Its message is the vendor's own words after the status (401 {"message":"…"}), as other adapters' failed calls read.
  • An adapter's own ModelProviderError (its sign-in failed, say) ends the call as it is, never retried, even when the library re-wrapped it as a connection failure.
  • Anything else (a bug, ours or the library's) propagates as it is, not retried and not dressed as the vendor's answer.
• A stop the library has no unified reason for is error, except running out of context (Bedrock's model_context_window_exceeded), which is length.
• The reasoning state survives a pause. A provider's opaque state (reasoning with its signature or encrypted content, each text part with its own text and id, a tool call's thought signature) is kept, in order, in the first tool call's signature.
  • It goes back on the next call to the same provider's same model, so a durable run can stop after a tool call, restart, and resume with no loss.
  • Another model's state is never sent, nor another provider's for a model of the same name.
  • When the trail's text no longer matches the answer's (a caller trimmed or redacted it), the trail's text goes back alone, where the first text part was, without the state that was about other words.
• Tool names: . → __ on send and back on receive, since vendors refuse dots. The system prompt names them as sent.
• Usage onto Kindgi's counters: prompt (cache reads and writes included), completion, cache read and write, and reasoning tokens. The adapter's own cost formula prices them, and rawUsage is the vendor's usage as sent.
• Settings a model refuses are dropped, with a warning: a temperature on a model registered sampling: false gives sampling-unsupported, never a 400. Lowest thinking is the engine's portable reasoning level.
• traceparent is sent as a header only when the call has one.
• For audit: the served model (servedModel) and the vendor's request id (providerRequestId).

Exports

@kindgi/adapter-model-shared, in Kindgi's terms:

• ModelProviderError (kind, status) and ModelProviderErrorKind;
• kindOf(failure) and modelProviderError(failure, cause): a failed response as a typed error;
• withRetries(send, policy), backoffMs(headers, attempt), vendorWaitMs(headers), MAX_VENDOR_WAIT_MS, RetryPolicy, RetryableFailure: the retry policy, library-free. A failure's maxAttempts caps its own attempts below the policy's; its kind says what it is when no status can.
• tokenCostUsd(model, usage, options?): a call's cost from the model's registered rates. It covers prompt and completion, cache reads and writes as multiples of the prompt rate, the long-context tier and the data-residency uplift. The arithmetic is the same as the OpenAI-compatible and Anthropic adapters' own: tests in those adapters hold them equal on the bundled presets, since this package never depends on an adapter. What a missing cache multiplier means is the adapter's to say: cacheReadMultiplier / cacheWriteMultiplier, default 1.

@kindgi/adapter-model-shared/ai-sdk, where an adapter hands its model in: createAiSdkModelProvider(options) returns a ModelProvider. Its options:

• metadata: the registration's ProviderMetadata;
• languageModel(name, fetch): the engine's model for a model name, sending with the fetch it's given (each HTTP attempt is counted through it);
• providerOptions?(model): provider options on every call (Azure: { azure: { store: false } });
• cost(model, usage): the adapter's cost formula;
• attempts?: HTTP attempts in all on a retryable failure (default 3);
• beforeAttempt?(signal): run before each HTTP attempt, inside the retries, with the call's abort signal. An adapter signs in here (a token, a key, the runtime's identity) when its library would report a failure as a plain error, and throws a ModelProviderError. What it returns is that attempt's own: the adapter's code the library calls during the attempt (a credential provider, a fetch wrapper) reads it with attemptPrepared(), and calls running at once never see each other's;
• explain?(error): a sentence after a failed call's message, where the vendor's words don't say what to check (Bedrock's 403). The kind, status and attempts stay as they are.

Pins

The engine's packages change often (two releases of one of them in a day during the evaluation). Every @ai-sdk/* dependency here, and in the adapters built on it, is pinned exactly and bumped on purpose, with the live cases re-run per bump. Kindgi keeps its own copy of the AI SDK's source and of the exact package files it pins.