@anthropic-ai/sdk
Version:
The official TypeScript library for the Anthropic API
702 lines (661 loc) • 26.8 kB
text/typescript
/**
* OpenTelemetry tracing of the client's API calls.
*
* Two terms recur in this file, and wherever else the SDK documents its spans:
*
* - An "API call" is one invocation of an SDK method that sends a request to the API. It
* includes every retry, and reading the stream that the method returns. Its span, a `CLIENT`
* span named after the method, is the "call span".
* - An "attempt" is one HTTP request within an API call, with its response when it receives
* one. An API call that the SDK retries makes several attempts. This SDK emits no span per
* attempt. The application's HTTP instrumentation, when it has one, emits each "attempt span"
* as a child of the call span.
*
* Each SDK method starts the call span of the API call it makes (`startApiCallSpan`) and passes it
* down in the call's request options (`__span`), so every part of the call records on it as things
* happen. The client sends its trace context with each attempt (`injectTraceHeaders`), unless the
* `propagation` option is off, and ends it with the call's result: at once when the call fails or
* the response body has been parsed (`endApiCallSpan`), or, for a streamed body, when the
* application has read or closed the stream (`endApiCallSpanWhenStreamEnds`).
*/
import {
context,
propagation,
trace,
SpanKind,
SpanStatusCode,
type Attributes,
type Span,
type Tracer,
type TracerProvider,
} from '@opentelemetry/api';
import { APIConnectionError, APIConnectionTimeoutError, APIError, APIUserAbortError } from '../core/error';
import { Stream } from '../core/streaming';
import { isAbortError } from './errors';
import { buildHeaders, type HeadersLike } from './headers';
import type { BaseAnthropic } from '../client';
import type { APIResponseProps } from './parse';
import type { RequestOptions } from './request-options';
import { readEnv } from './utils/env';
import { loggerFor } from './utils/log';
const INSTRUMENTATION_SCOPE = 'com.anthropic.sdk.typescript';
// Attribute names of the OpenTelemetry semantic conventions, named as
// `@opentelemetry/semantic-conventions` names them. They are copied rather than imported: that
// package is not a dependency of this SDK.
export const ATTR_ERROR_TYPE = 'error.type';
export const ATTR_GEN_AI_CONVERSATION_ID = 'gen_ai.conversation.id';
export const ATTR_GEN_AI_PROVIDER_NAME = 'gen_ai.provider.name';
export const ATTR_HTTP_REQUEST_METHOD = 'http.request.method';
export const ATTR_HTTP_REQUEST_RESEND_COUNT = 'http.request.resend_count';
export const ATTR_HTTP_RESPONSE_STATUS_CODE = 'http.response.status_code';
export const ATTR_HTTP_ROUTE = 'http.route';
export const ATTR_SERVER_ADDRESS = 'server.address';
export const ATTR_SERVER_PORT = 'server.port';
export const ATTR_URL_FULL = 'url.full';
export const ATTR_URL_PATH = 'url.path';
export const ATTR_URL_QUERY = 'url.query';
export const ATTR_URL_SCHEME = 'url.scheme';
// This API's own name, in the same style.
export const ATTR_ANTHROPIC_SESSION_THREAD_ID = 'anthropic.session.thread.id';
/**
* The `openTelemetry` client option.
*/
export interface OpenTelemetryOptions {
/**
* The tracer provider that creates this client's spans, in place of the one
* registered with the OpenTelemetry API.
*/
tracerProvider?: TracerProvider | undefined;
/**
* Whether this client sends the trace context of its spans with its requests, as the headers
* that the propagator registered with the OpenTelemetry API writes: `traceparent`, and
* `tracestate` and `baggage` when the propagator writes them. `false` is for an application
* whose HTTP client instrumentation already propagates the trace context, so that a request
* does not carry it twice. It does not affect the spans.
*
* @default true
*/
propagation?: boolean | undefined;
/**
* This client's spans. `true` and `false` are short for `{ enabled: true }` and
* `{ enabled: false }`.
*/
traces?: boolean | OpenTelemetryTracesOptions | undefined;
}
/**
* The `traces` field of the `openTelemetry` client option.
*/
export interface OpenTelemetryTracesOptions {
/**
* Whether this client emits spans. A client that emits no span does not use the other settings.
*
* @default true
*/
enabled?: boolean | undefined;
/**
* Whether this client's spans carry the content of its requests and responses, such as
* prompts and completions. `metadata_only` leaves the content out.
*
* @default 'metadata_only'
*/
contentMode?: 'metadata_only' | 'content' | undefined;
/**
* The most content that one span carries, in bytes. The SDK measures the content attributes'
* values by their size in UTF-8. A span with more content loses some of it, and has
* `anthropic.content.truncated`. An integer of at least 1.
*
* @default 750000
*/
maxContentBytes?: number | undefined;
}
/**
* The `openTelemetry` request option.
*/
export interface OpenTelemetryRequestOptions {
/**
* An ID that links this API call's span to the spans of the other API calls in one
* conversation. The SDK sets it as `gen_ai.conversation.id`, and does not send it to the API.
* Without one, a call to a Session thread's path has the ID of that thread there.
*/
conversationId?: string | undefined;
}
/**
* The `openTelemetry` client option as a client keeps it, with every field of `traces` settled.
* It is itself a valid `openTelemetry` option.
*/
export interface ResolvedOpenTelemetryOptions {
/**
* The tracer provider that creates the client's spans, or `undefined` for the one registered
* with the OpenTelemetry API.
*/
readonly tracerProvider?: TracerProvider | undefined;
/** Whether the client sends the trace context of its spans with its requests. */
readonly propagation: boolean;
/** The settings of the client's spans. */
readonly traces: {
/** Whether the client emits spans. */
readonly enabled: boolean;
/** Whether the client's spans carry the content of its requests and responses. */
readonly contentMode: 'metadata_only' | 'content';
/** The most content that one span carries, which the SDK measures in UTF-8 bytes. */
readonly maxContentBytes: number;
};
}
/** The default `traces.maxContentBytes`. */
const DEFAULT_MAX_CONTENT_BYTES = 750_000;
/**
* Settles `propagation` and the `traces` field of `option`, the `openTelemetry` client option. The
* environment variables apply only to a client that is given no option. An option that is given is
* complete: a field that it leaves unset has the default, whatever the environment says. A content
* mode or a size that is not valid is the default one.
*
* A client calls this once, when it is constructed, and keeps the result, which a
* `withOptions()` copy without its own option takes as it is. So a later change to the environment
* affects neither the client nor such a copy. The result is frozen.
*/
export function resolveOpenTelemetryOptions(
option: OpenTelemetryOptions | false | undefined,
): ResolvedOpenTelemetryOptions {
// `false` is short for `{ traces: false }`.
const { tracerProvider, propagation, traces }: OpenTelemetryOptions =
option === false ? { traces: false } : option ?? optionFromEnv();
const { enabled, contentMode, maxContentBytes }: OpenTelemetryTracesOptions =
typeof traces === 'boolean' ? { enabled: traces } : traces ?? {};
return Object.freeze({
tracerProvider,
propagation: propagation ?? true,
traces: Object.freeze({
enabled: enabled ?? true,
contentMode: contentMode === 'content' ? 'content' : 'metadata_only',
maxContentBytes:
maxContentBytes !== undefined && Number.isInteger(maxContentBytes) && maxContentBytes >= 1 ?
maxContentBytes
: DEFAULT_MAX_CONTENT_BYTES,
}),
});
}
/** What the environment variables give, for a client that is given no `openTelemetry` option. */
function optionFromEnv(): OpenTelemetryOptions {
return {
propagation: readEnv('ANTHROPIC_OPEN_TELEMETRY_PROPAGATION')?.toLowerCase() !== 'false',
traces: {
enabled: readEnv('ANTHROPIC_OPEN_TELEMETRY')?.toLowerCase() !== 'false',
contentMode:
readEnv('ANTHROPIC_OPEN_TELEMETRY_TRACES_CONTENT_MODE')?.toLowerCase() === 'content' ?
'content'
: 'metadata_only',
// The variable holds decimal digits only: `1e3` and `0x10` are not sizes.
maxContentBytes: Number(
/^\d+$/.exec(readEnv('ANTHROPIC_OPEN_TELEMETRY_TRACES_MAX_CONTENT_BYTES') ?? '')?.[0] ?? NaN,
),
},
};
}
/**
* The tracer for a client with the settled `openTelemetry` option `option`, or `undefined` when
* the client's spans are disabled.
* `version` is the SDK's version. It is a parameter because some packages build
* `src/internal` without `src/version.ts`. The client of another platform's package gives
* `platform`, which the name of the tracer's scope then ends with, and that package's `version`.
*/
export function getTracer(
option: ResolvedOpenTelemetryOptions,
version: string,
platform?: string | undefined,
): Tracer | undefined {
if (!option.traces.enabled) {
return undefined;
}
const scope = platform !== undefined ? `${INSTRUMENTATION_SCOPE}.${platform}` : INSTRUMENTATION_SCOPE;
return (option.tracerProvider ?? trace.getTracerProvider()).getTracer(scope, version);
}
/**
* Runs `step`, a step of tracing that calls the application's OpenTelemetry code: its tracer
* provider, sampler, span processors or propagator. A failure of telemetry never fails an API call,
* so what `step` throws goes to the logger of `client`, when there is one, and the result is then
* `undefined`: the SDK goes on as if tracing were off for that step.
*/
export function guardTracing<Result>(
client: BaseAnthropic | undefined,
step: () => Result,
): Result | undefined {
try {
return step();
} catch (problem) {
if (client) {
loggerFor(client).debug('could not record telemetry', problem);
}
return undefined;
}
}
/**
* The call span of an API call, with its name and start time, which a `Span` does not expose. It
* travels in the call's request options, as `__span`.
*/
export interface ApiCallSpan extends ApiCallSpanSettings {
span: Span;
/** The span's name, which says which SDK method makes the API call. */
name: string;
/** When the span started, in epoch milliseconds. */
startTime: number;
/** How many attempts the call has made so far. */
attempts?: number;
/**
* The error of a stream helper that failed while it read the call's stream. Without it the
* stream's end, which the helper's exit causes, would read as a cancellation.
*/
failure?: unknown;
/**
* The size of the request's content on the span, which counts toward the content size limit, or
* more than that when the content was not measured.
*/
requestContentSize?: number;
}
/**
* What an SDK method says of its API call when it starts the call span, besides the span's name.
*/
export interface ApiCallSpanSettings {
/**
* The endpoint's path as the API's description gives it, with the names of its parameters in
* braces, for `http.route`.
*/
route?: string | undefined;
/**
* The names of the query parameters that the API's description defines for the endpoint. Only
* these are recorded in `url.query`.
*/
queryParams?: readonly string[] | undefined;
/** Those of `queryParams` whose value is content on this endpoint only. */
contentQueryParams?: readonly string[] | undefined;
/**
* What the method knows of the call when it makes it, to start the span with: the IDs in the
* request's path, each under its attribute's name, from the method's own arguments. The span of
* a later page of a list starts with them too.
*/
attributes?: Attributes | undefined;
/**
* The attribute for the ID of the object that the method creates. The span gets it from the
* call's parsed response ({@link setCreatedIDAttribute}), so a failed create has none.
*/
createdIDAttribute?: string | undefined;
/**
* The field of the parsed response that has the ID of the created object, when it is not `id`:
* the response to adding a member has the member's `user_id`, and no `id`.
*/
createdIDField?: string | undefined;
}
/**
* Starts the call span of the API call that an SDK method of `client` is about to make with
* `options`, as a child of the current span. `name` is the span's name, and `settings` is what
* else the method says of the call. The method passes the result down in the call's request
* options, as `__span`.
*
* When `options` already has a span, that span is the result: a helper that calls an SDK method
* starts the span itself, to record on it, and the method's API call uses that span instead of
* starting another.
*
* Returns `undefined` when the client's spans are disabled, when the span failed to start
* ({@link guardTracing}), or when the span neither records nor propagates a trace context: such a
* span has no effect, so the API call runs as if tracing were off. With `propagation` off, a span
* with a valid context counts as propagating: it is the active span during the attempts, which the
* application's HTTP instrumentation continues.
*/
export function startApiCallSpan(
client: BaseAnthropic,
name: string,
options?: RequestOptions | undefined,
settings?: ApiCallSpanSettings | undefined,
): ApiCallSpan | undefined {
if (options?.__span) {
return options.__span;
}
const tracer = client._tracer;
if (!tracer) {
return undefined;
}
return guardTracing(client, () => {
const startTime = Date.now();
// A Session thread is a conversation, so its ID in the path stands in for one the application
// did not give.
const threadID = settings?.attributes?.[ATTR_ANTHROPIC_SESSION_THREAD_ID];
const conversationId =
options?.openTelemetry?.conversationId ?? (typeof threadID === 'string' ? threadID : undefined);
const span = tracer.startSpan(name, {
kind: SpanKind.CLIENT,
// On the span from its start, with what the method gives, for a sampler to decide on.
attributes: {
[ATTR_GEN_AI_PROVIDER_NAME]: client._genAIProviderName,
...serverAttributes(client.baseURL),
...(conversationId != null ? { [ATTR_GEN_AI_CONVERSATION_ID]: conversationId } : undefined),
...(settings?.route != null ? { [ATTR_HTTP_ROUTE]: settings.route } : undefined),
...settings?.attributes,
},
});
if (
!span.isRecording() &&
!(client.openTelemetry.propagation ?
Object.keys(injectContext(span)).length > 0
: trace.isSpanContextValid(span.spanContext()))
) {
return undefined;
}
return { ...settings, span, name, startTime };
});
}
/**
* Sets on `call`, the call span of an API call whose method creates one object, the ID of that
* object from `data`, the call's parsed response: its `id`, or the field that the method named.
* Does nothing for any other call, or when `data` has no string there.
*/
export function setCreatedIDAttribute(call: ApiCallSpan, data: unknown): void {
if (call.createdIDAttribute === undefined || typeof data !== 'object' || data === null) {
return;
}
const id: unknown = (data as Record<string, unknown>)[call.createdIDField ?? 'id'];
if (typeof id === 'string') {
call.span.setAttribute(call.createdIDAttribute, id);
}
}
/**
* `server.address` and `server.port` for the server at `url`, a client's base URL.
*/
function serverAttributes(url: string): Attributes {
try {
const { hostname, port, protocol } = new URL(url);
return {
// `URL.hostname` keeps an IPv6 literal's brackets, which delimit it in a URL;
// `server.address` is the bare address.
[ATTR_SERVER_ADDRESS]: hostname.replace(/^\[(.*)\]$/, '$1'),
[ATTR_SERVER_PORT]:
port ? Number(port)
: protocol === 'http:' ? 80
: 443,
};
} catch {
return {};
}
}
/**
* Records on `call` that the client is about to make an attempt of the call: a request with
* `method` to `url`. The first attempt gives the span `http.request.method` and the `url.*`
* attributes, which a later attempt never replaces. Each later attempt, whatever its reason,
* counts in `http.request.resend_count`, which a call with one attempt does not have.
*
* A query with a parameter whose value is content ({@link CONTENT_QUERY_PARAMS}, and the call's own
* `contentQueryParams`) makes `url.query` and `url.full` content too, which the span has only when
* `recordsContent` is true.
*/
export function recordAttempt(
call: ApiCallSpan,
method: string,
url: string,
recordsContent: boolean = false,
): void {
call.attempts = (call.attempts ?? 0) + 1;
if (call.attempts > 1) {
call.span.setAttribute(ATTR_HTTP_REQUEST_RESEND_COUNT, call.attempts - 1);
return;
}
call.span.setAttribute(ATTR_HTTP_REQUEST_METHOD, method);
try {
const { protocol, host, pathname, search } = new URL(url);
const query = definedQuery(search.slice(1), call.queryParams, call.contentQueryParams);
call.span.setAttributes({
[ATTR_URL_SCHEME]: protocol.slice(0, -1),
[ATTR_URL_PATH]: pathname,
});
if (query?.content && !recordsContent) {
return;
}
call.span.setAttributes({
...(query !== undefined ? { [ATTR_URL_QUERY]: query.value } : undefined),
// With the recorded query, not the one sent.
[ATTR_URL_FULL]: `${protocol}//${host}${pathname}${query !== undefined ? `?${query.value}` : ''}`,
});
} catch {}
}
/**
* The query parameters whose value is content on every endpoint that defines them.
*/
const CONTENT_QUERY_PARAMS: readonly string[] = ['email', 'path_prefix', 'expected_content_sha256'];
/**
* The parts of `query`, a query string as sent, whose parameter is one of `defined`, as they were
* sent and in their order, or `undefined` when there is none. A part's parameter is its
* percent-decoded name, or that name without a final `[]`, which is how an array parameter `ids`
* is sent: a name with other brackets, such as `created_at[gte]`, has to be one of `defined` as it
* is. `content` says whether one of those parameters is one of {@link CONTENT_QUERY_PARAMS} or of
* `contentParams`.
*/
function definedQuery(
query: string,
defined: readonly string[] | undefined,
contentParams: readonly string[] | undefined,
): { value: string; content: boolean } | undefined {
if (!query || !defined?.length) {
return undefined;
}
let content = false;
const parts = query.split('&').filter((part) => {
try {
const name = decodeURIComponent(part.split('=', 1)[0]!);
const param = defined.includes(name) || !name.endsWith('[]') ? name : name.slice(0, -2);
if (!defined.includes(param)) {
return false;
}
content ||= CONTENT_QUERY_PARAMS.includes(param) || !!contentParams?.includes(param);
return true;
} catch {
return false;
}
});
return parts.length ? { value: parts.join('&'), content } : undefined;
}
/**
* Returns `headers` with the trace context of `span` injected, as the propagator the
* application registered writes it, in place of any such headers already present. The client
* sends these with each attempt, so a retry carries the context afresh.
*/
export function injectTraceHeaders(span: Span, headers: HeadersLike): Headers {
// `fetchOptions` can replace the built headers with a plain object.
const values = buildHeaders([headers]).values;
for (const [key, value] of Object.entries(injectContext(span))) {
values.set(key, value);
}
return values;
}
/**
* The result of an API call made with `options`, as far as the call got.
*/
export type ApiCallResult =
| {
/** `response` arrived. */
status: 'success';
options: RequestOptions;
response: Response;
/** The parsed body of `response`, or `undefined` when the body was not read. */
data: unknown;
}
| {
/** The call failed with `error`. */
status: 'error';
options: RequestOptions;
error: unknown;
/** The response, when one arrived before the call failed. */
response?: Response | undefined;
};
/**
* Ends `span`, the call span of an API call, now that the call has `result`. The span gets
* `http.response.status_code` when a response arrived, and its status: `ERROR` with `error.type`
* when the call failed, none when the application aborted it, and `OK` otherwise. The span of a
* streamed body that the application has yet to read ends through
* {@link endApiCallSpanWhenStreamEnds} instead.
*
* Does nothing to a span that is not recording, such as one that already ended.
*/
export function endApiCallSpan(span: Span, result: ApiCallResult): void {
if (!span.isRecording()) {
return;
}
setResponseStatusCode(span, result);
switch (result.status) {
case 'success':
span.setStatus({ code: SpanStatusCode.OK });
break;
case 'error':
setErrorStatus(span, result.error);
break;
}
span.end();
}
/** How the application's read of a streamed body ended. */
export type StreamEnd =
| { reason: 'completed' } // the server ended the body, and the application read all of it
| { reason: 'canceled' } // the application closed the iterator early, or aborted the request
| { reason: 'failed'; error: unknown }; // reading threw
export interface StreamObserver<Item> {
/** Called with each item, before the application receives it. */
onItem?(item: Item): void;
/** Called once: when the read ends, or when the request is aborted, whichever is first. */
onEnd(end: StreamEnd): void;
}
/**
* Reports the application's read of `stream` to `observer`, by replacing the iterator of `stream`
* in place. `signal` is the abort signal of the request that `stream` reads. Only the first
* iterator is observed. What the callbacks throw is dropped.
*/
export function observeStream<Item>(
stream: AsyncIterable<Item>,
signal: AbortSignal,
observer: StreamObserver<Item>,
): void {
// Unbound, so that the abort listener, which `signal` keeps, does not keep an unread `stream`.
const iterate = stream[Symbol.asyncIterator];
let ended = false;
const end = (how: StreamEnd): void => {
if (ended) return;
ended = true;
signal.removeEventListener('abort', cancel);
try {
observer.onEnd(how);
} catch {}
};
const cancel = () => end({ reason: 'canceled' });
if (signal.aborted) return cancel();
signal.addEventListener('abort', cancel, { once: true });
let observed = false;
stream[Symbol.asyncIterator] = function (this: AsyncIterable<Item>) {
const source = iterate.call(this);
if (observed) return source;
observed = true;
// From here the iterator sees every way the read can end. `Stream` also aborts its own
// controller when it fails or is closed, which must not count as the application's abort.
signal.removeEventListener('abort', cancel);
const iterator: AsyncIterableIterator<Item> = {
async next() {
let result: IteratorResult<Item>;
try {
result = await source.next();
} catch (error) {
end({ reason: 'failed', error });
throw error;
}
if (result.done) {
// `Stream` ends quietly when the application aborts the request.
end({ reason: signal.aborted ? 'canceled' : 'completed' });
} else if (!ended) {
try {
observer.onItem?.(result.value);
} catch {}
}
return result;
},
async return(value) {
cancel();
return (await source.return?.(value)) ?? { done: true, value };
},
async throw(error) {
cancel();
if (source.throw) return source.throw(error);
throw error;
},
[Symbol.asyncIterator]: () => iterator,
};
return iterator;
};
}
/**
* Ends `span`, the call span of an API call, when the application's read of its streamed response
* body `stream` ends. Server-sent events have no end event, so the server ending them is not `OK`;
* any other body is complete at its end.
*/
export function endApiCallSpanWhenStreamEnds(
span: Span,
stream: AsyncIterable<unknown>,
props: APIResponseProps,
): void {
if (!span.isRecording()) return;
span.setAttribute(ATTR_HTTP_RESPONSE_STATUS_CODE, props.response.status);
const isSSE = stream instanceof Stream;
observeStream(stream, props.controller.signal, {
onEnd(end) {
if (end.reason === 'failed') setErrorStatus(span, end.error);
else if (end.reason === 'completed' && !isSSE) {
span.setStatus({ code: SpanStatusCode.OK });
}
span.end();
},
});
}
/**
* Sets `http.response.status_code` on the call span from the response in `result`, or from its
* error when that is an error response.
*/
function setResponseStatusCode(span: Span, result: ApiCallResult): void {
const status =
result.response?.status ??
(result.status === 'error' && result.error instanceof APIError ? result.error.status : undefined);
if (status !== undefined) {
span.setAttribute(ATTR_HTTP_RESPONSE_STATUS_CODE, status);
}
}
/**
* Gives `span` the status of an operation that failed with `error`: `ERROR`, with `error.type`
* set to `type` or else derived from `error`. An abort by the application is a cancellation,
* not a failure, and leaves the status unset.
*/
export function setErrorStatus(span: Span, error: unknown, type?: string): void {
if (error instanceof APIUserAbortError || isAbortError(error)) {
return;
}
span.setStatus({ code: SpanStatusCode.ERROR });
span.setAttribute(ATTR_ERROR_TYPE, type ?? errorType(error));
}
/**
* The headers that carry `span`'s trace context, from the propagator the application registered.
*/
function injectContext(span: Span): Record<string, string> {
const carrier: Record<string, string> = {};
propagation.inject(trace.setSpan(context.active(), span), carrier);
return carrier;
}
/**
* The `error.type` of an operation that failed with `error`. An error response has the type of
* its error envelope when the API sent one, and its status code otherwise.
*/
function errorType(error: unknown): string {
if (error instanceof APIConnectionTimeoutError) {
return 'client_timeout_error';
}
if (error instanceof APIConnectionError) {
return 'connection_error';
}
if (error instanceof APIError) {
if (error.type) {
return error.type;
}
if (error.status !== undefined) {
return String(error.status);
}
}
return 'client_error';
}