@stigg/agent-sdk-js 0.1.0 → 0.2.0

package/dist/bindings/shared/token-usage.d.ts

@@ -7,7 +7,7 @@
  * `fromFlatTokenUsage` is read by every vendor that counts that way, not by one of them.
  */
 import { type Measurement, type Metrics, type MetricsMapper, type TokenUsage } from '../../types';
-import type { BedrockLikeUsage, FlatTokenUsage, GeminiLikeUsage, LangChainLikeUsage, MistralLikeUsage, OllamaLikeUsage, OpenAILikeUsage, ResponsesLikeUsage } from './types';
+import type { BedrockLikeUsage, FlatTokenUsage, GeminiLikeUsage, LangChainLikeUsage, MistralLikeUsage, DatabricksLikeUsage, DatabricksQueryLikeUsage, OllamaLikeUsage, OpenAILikeUsage, ResponsesLikeUsage } from './types';
 export declare function fromFlatTokenUsage(usage: FlatTokenUsage | null | undefined): TokenUsage;
 export declare function fromOpenAIUsage(usage: OpenAILikeUsage | null | undefined): TokenUsage;
 /**
@@ -140,6 +140,21 @@ export declare const OLLAMA_METRICS: readonly ["prompt_eval_count", "eval_count"
 export declare const OPENAI_METRICS: readonly ["prompt_tokens", "completion_tokens", "input_tokens", "output_tokens", "cached_tokens", "total_tokens"];
 /** A gateway speaks the OpenAI shape and adds the price it charged, under its own field name. */
 export declare const GATEWAY_METRICS: readonly ["prompt_tokens", "completion_tokens", "input_tokens", "output_tokens", "cached_tokens", "total_tokens", "cost"];
+/**
+ * A gateway that forwards the upstream body untouched, and so reports in two vocabularies.
+ *
+ * Unlike LiteLLM, which normalises every upstream to the OpenAI wire, Cloudflare AI Gateway's
+ * provider-native routes pass the provider's own body straight through - so `/openai` returns
+ * OpenAI's counter names and `/anthropic` returns Anthropic's. One recipe serves both wires, so
+ * this is the union of what its surfaces can report, which is what `metrics` is defined to be.
+ *
+ * No `cost`, and that is the defining fact rather than an oversight. Cloudflare states no price on
+ * any response: its figure lives in the gateway's own log, is explicitly an estimate, and is zero
+ * on a cache hit. So the recipe carries the log id as a dimension and lets spend be reconciled out
+ * of band, exactly as the gateway reference prescribes for a price that is only available
+ * asynchronously.
+ */
+export declare const CLOUDFLARE_METRICS: readonly ["prompt_tokens", "completion_tokens", "input_tokens", "output_tokens", "cached_tokens", "total_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
 /**
  * Bedrock's five, under its own names and with no price among them.
  *
@@ -154,6 +169,93 @@ export declare const GATEWAY_METRICS: readonly ["prompt_tokens", "completion_tok
 export declare const BEDROCK_PATHS: readonly ["inputTokens", "outputTokens", "cacheReadInputTokens", "cacheWriteInputTokens", "totalTokens"];
 /** Those five. Flat, so the paths are the names. */
 export declare const BEDROCK_METRICS: readonly ["inputTokens", "outputTokens", "cacheReadInputTokens", "cacheWriteInputTokens", "totalTokens"];
+/**
+ * Databricks' six, under its own names - and its own names are two vendors' names at once.
+ *
+ * The first three are OpenAI's spelling, because the serving endpoints speak the OpenAI wire. The
+ * two cache counters are *Anthropic's* spelling, flat on the usage block rather than nested under
+ * `prompt_tokens_details` where an OpenAI-shaped reader would look for them, and `reasoning_tokens`
+ * is Databricks' own. Reported exactly as Databricks spells them, per the rule that a number has to
+ * reconcile against the vendor's own dashboard - renaming the cache pair to `cached_tokens` would
+ * make one account read as two vocabularies and find nothing where it looked.
+ *
+ * No `cost`, and that is the fact that most shapes this binding: Databricks states no price on any
+ * response, on either route. See `databricks-recipe.ts` for what that costs a run.
+ */
+export declare const DATABRICKS_PATHS: readonly ["prompt_tokens", "completion_tokens", "total_tokens", "reasoning_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
+/**
+ * The same counters as the Responses API spells them.
+ *
+ * Databricks serves both OpenAI surfaces, and they do not share field names: Chat Completions
+ * sends `prompt_tokens`/`completion_tokens` and Responses sends `input_tokens`/`output_tokens`.
+ * Reading a Responses body with the list above finds neither, so the call would report its total
+ * and nothing else - a number with no breakdown under it, which prices wrongly rather than
+ * visibly. Each surface reads its own list.
+ */
+export declare const DATABRICKS_RESPONSES_PATHS: readonly ["input_tokens", "output_tokens", "total_tokens", "reasoning_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
+/** The union across both surfaces: what a Databricks call can report, which is what `metrics` is. */
+export declare const DATABRICKS_METRICS: readonly ["prompt_tokens", "completion_tokens", "input_tokens", "output_tokens", "total_tokens", "reasoning_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
+/**
+ * Where the **first-party client's** counters are read and what each is reported as.
+ *
+ * The second map in this file, and it exists for precisely the reason the Mistral one does.
+ * `@databricks/sdk-modelservingquery` is generated, and its zod schema parses `prompt_tokens` off
+ * the wire and transforms it to `promptTokens` before the caller ever sees it - so following the
+ * key it was read from would report a camelCase feature id for a call that the *same workspace*
+ * reports in snake_case when the application holds an `OpenAI` client instead.
+ *
+ * That is not a cosmetic difference. It is one Databricks account producing two different feature
+ * ids depending on which package the application happened to import, splitting one bill across two
+ * meters - which is the failure `pickRenamedMetrics` was built to prevent. Three entries rather
+ * than six because `ExternalModelUsageElement` carries only three; the cache pair and
+ * `reasoning_tokens` have no field on it, and a counter with no field is absent rather than zero.
+ */
+export declare const DATABRICKS_QUERY_METRIC_NAME_MAP: {
+    readonly promptTokens: "prompt_tokens";
+    readonly completionTokens: "completion_tokens";
+    readonly totalTokens: "total_tokens";
+};
+/**
+ * Workers AI's three, and only three.
+ *
+ * OpenAI's spelling, flat, with no cache detail nested underneath - so this is a narrower list
+ * than `OPENAI_CHAT_METRICS` rather than a reuse of it. Declaring `cached_tokens` here would type
+ * a pricing table against a counter Workers AI never sends.
+ *
+ * No neurons either, which is the counter a reader might expect. Cloudflare bills Workers AI in
+ * Neurons, and publishes none of them on any response - binding or REST - so the conversion from
+ * tokens is a price list this SDK does not hold. Tokens are what the response states and tokens
+ * are what this reports.
+ */
+export declare const WORKERS_AI_PATHS: readonly ["prompt_tokens", "completion_tokens", "total_tokens"];
+/** Those three. Flat, so the paths are the names. */
+export declare const WORKERS_AI_METRICS: readonly ["prompt_tokens", "completion_tokens", "total_tokens"];
+/**
+ * Databricks' usage block -> the neutral digest.
+ *
+ * `input` is `prompt_tokens` verbatim, *not* reduced by the cache reads the way `fromOpenAIUsage`
+ * reduces its own. OpenAI documents its cached tokens as counted inside `prompt_tokens`, so
+ * subtracting there is undoing a documented overlap. Databricks documents no such thing, and the
+ * spelling it chose for the pair is Anthropic's - whose cache counters sit *beside* the input
+ * count rather than inside it. So the Anthropic convention is the better guess and the overlap is
+ * left alone; the stated total is read rather than computed, which keeps the one number that gets
+ * invoiced right whichever way the overlap actually runs. The README's overlap table says the
+ * direction is unverified for Databricks, because it is.
+ */
+export declare function fromDatabricksUsage(usage: DatabricksLikeUsage | null | undefined): TokenUsage;
+/**
+ * The same block as read through the first-party client -> the neutral digest.
+ *
+ * Its own reader rather than a call into `fromDatabricksUsage`, because the field names genuinely
+ * differ: the generated client camelCases the wire on the way in, so a snake_case reader finds
+ * none of these and returns an all-zero digest without throwing. That is the shape of failure the
+ * invariants call out - an extractor that meters every call at nothing and trips no test.
+ *
+ * No cache arithmetic to get wrong here, which is the one simplification over the wire-shaped
+ * reader: `ExternalModelUsageElement` declares no cache counters at all, so there is no overlap of
+ * unknown direction to leave alone. The stated total is still read rather than recomputed.
+ */
+export declare function fromDatabricksQueryUsage(usage: DatabricksQueryLikeUsage | null | undefined): TokenUsage;
 /** The leaf a dotted path reports under - `prompt_tokens_details.cached_tokens` -> `cached_tokens`. */
 export declare const metricName: (path: string) => string;
 /**
@@ -188,6 +290,9 @@ export declare const METRIC_PATHS: {
     readonly typesafe: readonly ["input_tokens", "output_tokens"];
     readonly gemini: readonly ["promptTokenCount", "candidatesTokenCount", "cachedContentTokenCount", "thoughtsTokenCount", "toolUsePromptTokenCount", "totalTokenCount"];
     readonly bedrock: readonly ["inputTokens", "outputTokens", "cacheReadInputTokens", "cacheWriteInputTokens", "totalTokens"];
+    readonly databricks: readonly ["prompt_tokens", "completion_tokens", "total_tokens", "reasoning_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
+    readonly databricksResponses: readonly ["input_tokens", "output_tokens", "total_tokens", "reasoning_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"];
+    readonly workersAi: readonly ["prompt_tokens", "completion_tokens", "total_tokens"];
     readonly ollama: readonly ["prompt_eval_count", "eval_count"];
 };
 /**

package/dist/bindings/shared/token-usage.js

@@ -1,6 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.normalizedMetrics = exports.addMetrics = exports.METRIC_PATHS = exports.metricNamesMatchPaths = exports.COMPUTED_TOTALS = exports.metricName = exports.BEDROCK_METRICS = exports.BEDROCK_PATHS = exports.GATEWAY_METRICS = exports.OPENAI_METRICS = exports.OLLAMA_METRICS = exports.OLLAMA_COMPONENT_METRICS = exports.TYPESAFE_METRICS = exports.TYPESAFE_COMPONENT_METRICS = exports.MISTRAL_METRICS = exports.MISTRAL_METRIC_NAME_MAP = exports.OPENAI_RESPONSES_METRICS = exports.OPENAI_CHAT_METRICS = exports.GEMINI_COMPONENT_METRICS = exports.GEMINI_METRICS = exports.ANTHROPIC_METRICS = exports.ANTHROPIC_COMPONENT_METRICS = exports.TOTAL_TOKENS = exports.addTokenUsage = void 0;
+exports.normalizedMetrics = exports.addMetrics = exports.METRIC_PATHS = exports.metricNamesMatchPaths = exports.COMPUTED_TOTALS = exports.metricName = exports.WORKERS_AI_METRICS = exports.WORKERS_AI_PATHS = exports.DATABRICKS_QUERY_METRIC_NAME_MAP = exports.DATABRICKS_METRICS = exports.DATABRICKS_RESPONSES_PATHS = exports.DATABRICKS_PATHS = exports.BEDROCK_METRICS = exports.BEDROCK_PATHS = exports.CLOUDFLARE_METRICS = exports.GATEWAY_METRICS = exports.OPENAI_METRICS = exports.OLLAMA_METRICS = exports.OLLAMA_COMPONENT_METRICS = exports.TYPESAFE_METRICS = exports.TYPESAFE_COMPONENT_METRICS = exports.MISTRAL_METRICS = exports.MISTRAL_METRIC_NAME_MAP = exports.OPENAI_RESPONSES_METRICS = exports.OPENAI_CHAT_METRICS = exports.GEMINI_COMPONENT_METRICS = exports.GEMINI_METRICS = exports.ANTHROPIC_METRICS = exports.ANTHROPIC_COMPONENT_METRICS = exports.TOTAL_TOKENS = exports.addTokenUsage = void 0;
 exports.fromFlatTokenUsage = fromFlatTokenUsage;
 exports.fromOpenAIUsage = fromOpenAIUsage;
 exports.fromResponsesUsage = fromResponsesUsage;
@@ -10,6 +10,8 @@ exports.fromGeminiUsage = fromGeminiUsage;
 exports.fromMistralUsage = fromMistralUsage;
 exports.fromOllamaUsage = fromOllamaUsage;
 exports.normalizeUsage = normalizeUsage;
+exports.fromDatabricksUsage = fromDatabricksUsage;
+exports.fromDatabricksQueryUsage = fromDatabricksQueryUsage;
 exports.withComputedTotal = withComputedTotal;
 exports.pickMetrics = pickMetrics;
 exports.pickRenamedMetrics = pickRenamedMetrics;
@@ -404,6 +406,25 @@ exports.OPENAI_METRICS = [
 ];
 /** A gateway speaks the OpenAI shape and adds the price it charged, under its own field name. */
 exports.GATEWAY_METRICS = [...exports.OPENAI_METRICS, 'cost'];
+/**
+ * A gateway that forwards the upstream body untouched, and so reports in two vocabularies.
+ *
+ * Unlike LiteLLM, which normalises every upstream to the OpenAI wire, Cloudflare AI Gateway's
+ * provider-native routes pass the provider's own body straight through - so `/openai` returns
+ * OpenAI's counter names and `/anthropic` returns Anthropic's. One recipe serves both wires, so
+ * this is the union of what its surfaces can report, which is what `metrics` is defined to be.
+ *
+ * No `cost`, and that is the defining fact rather than an oversight. Cloudflare states no price on
+ * any response: its figure lives in the gateway's own log, is explicitly an estimate, and is zero
+ * on a cache hit. So the recipe carries the log id as a dimension and lets spend be reconciled out
+ * of band, exactly as the gateway reference prescribes for a price that is only available
+ * asynchronously.
+ */
+exports.CLOUDFLARE_METRICS = [
+    ...exports.OPENAI_METRICS,
+    'cache_read_input_tokens',
+    'cache_creation_input_tokens',
+];
 /**
  * Bedrock's five, under its own names and with no price among them.
  *
@@ -424,6 +445,143 @@ exports.BEDROCK_PATHS = [
 ];
 /** Those five. Flat, so the paths are the names. */
 exports.BEDROCK_METRICS = exports.BEDROCK_PATHS;
+/**
+ * Databricks' six, under its own names - and its own names are two vendors' names at once.
+ *
+ * The first three are OpenAI's spelling, because the serving endpoints speak the OpenAI wire. The
+ * two cache counters are *Anthropic's* spelling, flat on the usage block rather than nested under
+ * `prompt_tokens_details` where an OpenAI-shaped reader would look for them, and `reasoning_tokens`
+ * is Databricks' own. Reported exactly as Databricks spells them, per the rule that a number has to
+ * reconcile against the vendor's own dashboard - renaming the cache pair to `cached_tokens` would
+ * make one account read as two vocabularies and find nothing where it looked.
+ *
+ * No `cost`, and that is the fact that most shapes this binding: Databricks states no price on any
+ * response, on either route. See `databricks-recipe.ts` for what that costs a run.
+ */
+exports.DATABRICKS_PATHS = [
+    'prompt_tokens',
+    'completion_tokens',
+    'total_tokens',
+    'reasoning_tokens',
+    'cache_read_input_tokens',
+    'cache_creation_input_tokens',
+];
+/**
+ * The same counters as the Responses API spells them.
+ *
+ * Databricks serves both OpenAI surfaces, and they do not share field names: Chat Completions
+ * sends `prompt_tokens`/`completion_tokens` and Responses sends `input_tokens`/`output_tokens`.
+ * Reading a Responses body with the list above finds neither, so the call would report its total
+ * and nothing else - a number with no breakdown under it, which prices wrongly rather than
+ * visibly. Each surface reads its own list.
+ */
+exports.DATABRICKS_RESPONSES_PATHS = [
+    'input_tokens',
+    'output_tokens',
+    'total_tokens',
+    'reasoning_tokens',
+    'cache_read_input_tokens',
+    'cache_creation_input_tokens',
+];
+/** The union across both surfaces: what a Databricks call can report, which is what `metrics` is. */
+exports.DATABRICKS_METRICS = [
+    'prompt_tokens',
+    'completion_tokens',
+    'input_tokens',
+    'output_tokens',
+    'total_tokens',
+    'reasoning_tokens',
+    'cache_read_input_tokens',
+    'cache_creation_input_tokens',
+];
+/**
+ * Where the **first-party client's** counters are read and what each is reported as.
+ *
+ * The second map in this file, and it exists for precisely the reason the Mistral one does.
+ * `@databricks/sdk-modelservingquery` is generated, and its zod schema parses `prompt_tokens` off
+ * the wire and transforms it to `promptTokens` before the caller ever sees it - so following the
+ * key it was read from would report a camelCase feature id for a call that the *same workspace*
+ * reports in snake_case when the application holds an `OpenAI` client instead.
+ *
+ * That is not a cosmetic difference. It is one Databricks account producing two different feature
+ * ids depending on which package the application happened to import, splitting one bill across two
+ * meters - which is the failure `pickRenamedMetrics` was built to prevent. Three entries rather
+ * than six because `ExternalModelUsageElement` carries only three; the cache pair and
+ * `reasoning_tokens` have no field on it, and a counter with no field is absent rather than zero.
+ */
+exports.DATABRICKS_QUERY_METRIC_NAME_MAP = {
+    promptTokens: 'prompt_tokens',
+    completionTokens: 'completion_tokens',
+    totalTokens: 'total_tokens',
+};
+/**
+ * Workers AI's three, and only three.
+ *
+ * OpenAI's spelling, flat, with no cache detail nested underneath - so this is a narrower list
+ * than `OPENAI_CHAT_METRICS` rather than a reuse of it. Declaring `cached_tokens` here would type
+ * a pricing table against a counter Workers AI never sends.
+ *
+ * No neurons either, which is the counter a reader might expect. Cloudflare bills Workers AI in
+ * Neurons, and publishes none of them on any response - binding or REST - so the conversion from
+ * tokens is a price list this SDK does not hold. Tokens are what the response states and tokens
+ * are what this reports.
+ */
+exports.WORKERS_AI_PATHS = ['prompt_tokens', 'completion_tokens', 'total_tokens'];
+/** Those three. Flat, so the paths are the names. */
+exports.WORKERS_AI_METRICS = exports.WORKERS_AI_PATHS;
+/**
+ * Databricks' usage block -> the neutral digest.
+ *
+ * `input` is `prompt_tokens` verbatim, *not* reduced by the cache reads the way `fromOpenAIUsage`
+ * reduces its own. OpenAI documents its cached tokens as counted inside `prompt_tokens`, so
+ * subtracting there is undoing a documented overlap. Databricks documents no such thing, and the
+ * spelling it chose for the pair is Anthropic's - whose cache counters sit *beside* the input
+ * count rather than inside it. So the Anthropic convention is the better guess and the overlap is
+ * left alone; the stated total is read rather than computed, which keeps the one number that gets
+ * invoiced right whichever way the overlap actually runs. The README's overlap table says the
+ * direction is unverified for Databricks, because it is.
+ */
+function fromDatabricksUsage(usage) {
+    if (!usage)
+        return (0, types_1.emptyTokenUsage)();
+    return {
+        // Read, never computed: Databricks states a total, and a stated total beats a reassembled one
+        // precisely because the overlap above is unknown.
+        ...withTotal(n(usage.total_tokens)),
+        // Both spellings, because one Databricks workspace serves both OpenAI surfaces: Chat
+        // Completions sends `prompt_tokens`/`completion_tokens` and Responses sends
+        // `input_tokens`/`output_tokens`. A body carries one pair or the other, never both, so the
+        // fallback cannot double-count - and reading only the chat pair left `normalizedMetrics` and
+        // every custom `TokenUsage` mapper seeing zeroes for a Responses call.
+        input: n(usage.prompt_tokens ?? usage.input_tokens),
+        output: n(usage.completion_tokens ?? usage.output_tokens),
+        cacheRead: n(usage.cache_read_input_tokens),
+        cacheCreate: n(usage.cache_creation_input_tokens),
+    };
+}
+/**
+ * The same block as read through the first-party client -> the neutral digest.
+ *
+ * Its own reader rather than a call into `fromDatabricksUsage`, because the field names genuinely
+ * differ: the generated client camelCases the wire on the way in, so a snake_case reader finds
+ * none of these and returns an all-zero digest without throwing. That is the shape of failure the
+ * invariants call out - an extractor that meters every call at nothing and trips no test.
+ *
+ * No cache arithmetic to get wrong here, which is the one simplification over the wire-shaped
+ * reader: `ExternalModelUsageElement` declares no cache counters at all, so there is no overlap of
+ * unknown direction to leave alone. The stated total is still read rather than recomputed.
+ */
+function fromDatabricksQueryUsage(usage) {
+    if (!usage)
+        return (0, types_1.emptyTokenUsage)();
+    return {
+        ...withTotal(n(usage.totalTokens)),
+        input: n(usage.promptTokens),
+        output: n(usage.completionTokens),
+        cacheRead: 0,
+        cacheCreate: 0,
+    };
+}
 /** The leaf a dotted path reports under - `prompt_tokens_details.cached_tokens` -> `cached_tokens`. */
 const metricName = (path) => path.split('.').pop();
 exports.metricName = metricName;
@@ -470,6 +628,9 @@ exports.METRIC_PATHS = {
     typesafe: TYPESAFE_PATHS,
     gemini: GEMINI_PATHS,
     bedrock: exports.BEDROCK_PATHS,
+    databricks: exports.DATABRICKS_PATHS,
+    databricksResponses: exports.DATABRICKS_RESPONSES_PATHS,
+    workersAi: exports.WORKERS_AI_PATHS,
     ollama: exports.OLLAMA_COMPONENT_METRICS,
 };
 /**

package/dist/bindings/shared/types.d.ts

@@ -31,6 +31,47 @@ export type OpenAILikeUsage = {
         cache_write_tokens?: number | null;
     } | null;
 };
+/**
+ * Databricks model serving `usage` - OpenAI's three counters, with Anthropic's spelling for cache.
+ *
+ * The mixed vocabulary is Databricks' own and is reported verbatim rather than normalised to one
+ * side or the other. The first three come back from every serving endpoint; the last three are
+ * Foundation Model API extensions and are absent on the external-model route, where the body is
+ * whatever the upstream provider's counters normalise to.
+ */
+export type DatabricksLikeUsage = {
+    prompt_tokens?: number | null;
+    completion_tokens?: number | null;
+    /** Databricks states its own total, so it is read rather than reassembled from the parts. */
+    total_tokens?: number | null;
+    /** Reasoning models only. */
+    reasoning_tokens?: number | null;
+    cache_read_input_tokens?: number | null;
+    cache_creation_input_tokens?: number | null;
+};
+/**
+ * The same block as above, as `@databricks/sdk-modelservingquery` hands it over.
+ *
+ * Databricks' own `ExternalModelUsageElement`. The **wire** is the snake_case one above - the
+ * generated client's zod schema parses `prompt_tokens` and transforms it to `promptTokens` on the
+ * way in, exactly as Mistral's generated client does to `UsageInfo`. So this is a second spelling
+ * of one vendor's counters rather than a second set of counters, and it is read through
+ * `DATABRICKS_QUERY_METRIC_NAME_MAP` so both spellings report under the wire's names.
+ *
+ * Three, not six: the cache pair and `reasoning_tokens` have no field on this type at all, so a
+ * call made through the first-party client reports the three it can and nothing more. Absent, not
+ * zero - and the reason the native surface's metric list is a subset of the recipe's.
+ *
+ * Written out rather than aliased to {@link MistralLikeUsage}, which it currently matches
+ * field-for-field: these are two unrelated vendor contracts that happen to agree today, and
+ * collapsing them would make a change on either side look like a change to both.
+ */
+export type DatabricksQueryLikeUsage = {
+    promptTokens?: number | null;
+    completionTokens?: number | null;
+    /** Databricks states its own total here too, so it is read rather than reassembled. */
+    totalTokens?: number | null;
+};
 /** OpenAI's Responses API `usage` - the same counts as above under different names. */
 export type ResponsesLikeUsage = {
     input_tokens?: number | null;

package/dist/index.d.ts

@@ -46,9 +46,15 @@ export { mistralRecipe } from './bindings/client/providers/mistral/mistral-recip
 export type { MistralMetric, MistralMetrics } from './bindings/client/providers/mistral/mistral-recipe';
 export { ollamaRecipe } from './bindings/client/providers/ollama/ollama-recipe';
 export type { OllamaMetric, OllamaMetrics } from './bindings/client/providers/ollama/ollama-recipe';
+export { databricksRecipe, databricksExternalRecipe } from './bindings/client/gateways/databricks/databricks-recipe';
+export type { DatabricksMetricName, DatabricksMetrics } from './bindings/client/gateways/databricks/databricks-recipe';
+export type { DatabricksRecipeOptions } from './bindings/client/gateways/databricks/types';
 export { bedrockRecipe } from './bindings/client/gateways/bedrock/bedrock-recipe';
 export type { BedrockMetricName, BedrockMetrics } from './bindings/client/gateways/bedrock/bedrock-recipe';
 export type { BedrockRecipeOptions } from './bindings/client/gateways/bedrock/types';
+export { cloudflareRecipe } from './bindings/client/gateways/cloudflare/cloudflare-recipe';
+export type { CloudflareMetricName, CloudflareMetrics } from './bindings/client/gateways/cloudflare/cloudflare-recipe';
+export type { CloudflareRecipeOptions } from './bindings/client/gateways/cloudflare/types';
 export { typesafeRecipe, estimateFromState } from './bindings/client/providers/typesafe/typesafe-recipe';
 export type { TypeSafeMetric, TypeSafeMetrics } from './bindings/client/providers/typesafe/typesafe-recipe';
 export { weightedMetrics } from './bindings/shared/weighted-metrics';

package/dist/index.js

@@ -1,6 +1,6 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.normalizeUsage = exports.normalizedMetrics = exports.weightedMetrics = exports.estimateFromState = exports.typesafeRecipe = exports.bedrockRecipe = exports.ollamaRecipe = exports.mistralRecipe = exports.groqRecipe = exports.litellmRecipe = exports.openrouterRecipe = exports.estimateFromMaxOutputTokens = exports.geminiChatRecipe = exports.geminiRecipe = exports.openAIRecipe = exports.estimateFromMaxTokens = exports.anthropicRecipe = exports.hasMethod = exports.createClientRegistry = exports.wrapClient = exports.toMeasurement = exports.entitlementRef = exports.entitlementKind = exports.entitlementId = exports.emptyTokenUsage = exports.vendorMetrics = exports.unitsPerTokens = exports.looksLikeGovernanceNotEnabled = exports.isGovernanceNotEnabled = exports.isAccessDenied = exports.isEngineError = exports.GovernanceNotEnabledError = exports.AccessDeniedError = exports.EngineError = exports.noopLogger = exports.consoleLogger = exports.parseGovernanceHeaders = exports.GOVERNANCE_HEADERS = exports.remedyMessage = exports.describeDecision = exports.flattenRaw = exports.RAW_PREFIX = exports.StiggAgentSDK = void 0;
+exports.normalizeUsage = exports.normalizedMetrics = exports.weightedMetrics = exports.estimateFromState = exports.typesafeRecipe = exports.cloudflareRecipe = exports.bedrockRecipe = exports.databricksExternalRecipe = exports.databricksRecipe = exports.ollamaRecipe = exports.mistralRecipe = exports.groqRecipe = exports.litellmRecipe = exports.openrouterRecipe = exports.estimateFromMaxOutputTokens = exports.geminiChatRecipe = exports.geminiRecipe = exports.openAIRecipe = exports.estimateFromMaxTokens = exports.anthropicRecipe = exports.hasMethod = exports.createClientRegistry = exports.wrapClient = exports.toMeasurement = exports.entitlementRef = exports.entitlementKind = exports.entitlementId = exports.emptyTokenUsage = exports.vendorMetrics = exports.unitsPerTokens = exports.looksLikeGovernanceNotEnabled = exports.isGovernanceNotEnabled = exports.isAccessDenied = exports.isEngineError = exports.GovernanceNotEnabledError = exports.AccessDeniedError = exports.EngineError = exports.noopLogger = exports.consoleLogger = exports.parseGovernanceHeaders = exports.GOVERNANCE_HEADERS = exports.remedyMessage = exports.describeDecision = exports.flattenRaw = exports.RAW_PREFIX = exports.StiggAgentSDK = void 0;
 /**
  * Public surface.
  *
@@ -92,10 +92,28 @@ Object.defineProperty(exports, "ollamaRecipe", { enumerable: true, get: function
 // than a method tree, so the recipe gates `send` and reads the command to tell inference from the
 // rest of the API; it reports Bedrock's own token counters and no `cost`, because there is none on
 // the response to report. See the recipe for what that means for a run that wanted to bill spend.
+// Databricks model serving, as two recipes for two payers over one wire. `databricksRecipe` is a
+// default, matched on the workspace URL: Foundation Model API traffic, billed to your Databricks
+// account in DBUs. `databricksExternalRecipe` is the same wire reaching an endpoint configured
+// with your own upstream key, where the upstream provider invoices - it is opt-in because the two
+// are indistinguishable from the client and naming it is an assertion about your own workspace,
+// not something the SDK detected. Neither reports a `cost`: Databricks states no price on any
+// response, on either route. See the recipe header for what a wrong assertion costs.
+var databricks_recipe_1 = require("./bindings/client/gateways/databricks/databricks-recipe");
+Object.defineProperty(exports, "databricksRecipe", { enumerable: true, get: function () { return databricks_recipe_1.databricksRecipe; } });
+Object.defineProperty(exports, "databricksExternalRecipe", { enumerable: true, get: function () { return databricks_recipe_1.databricksExternalRecipe; } });
 var bedrock_recipe_1 = require("./bindings/client/gateways/bedrock/bedrock-recipe");
 Object.defineProperty(exports, "bedrockRecipe", { enumerable: true, get: function () { return bedrock_recipe_1.bedrockRecipe; } });
 // Built in as well. One method, no stream, and two counters - see the recipe for why a
 // question-answering API has neither a cache counter nor a cost to report.
+// Cloudflare Workers AI - a provider, not a gateway: Cloudflare runs these models on its own GPUs
+// and invoices them, and `@cf/<author>/…` names who made the weights rather than who bills. A
+// default, matched on `api.cloudflare.com`, covering both the OpenAI-compatible route and the
+// `cloudflare` package's own `ai.run`. It reports tokens and no `cost` - Cloudflare bills in
+// Neurons and publishes none of them on any response. See the recipe for the streamed `ai.run`
+// limitation.
+var cloudflare_recipe_1 = require("./bindings/client/gateways/cloudflare/cloudflare-recipe");
+Object.defineProperty(exports, "cloudflareRecipe", { enumerable: true, get: function () { return cloudflare_recipe_1.cloudflareRecipe; } });
 var typesafe_recipe_1 = require("./bindings/client/providers/typesafe/typesafe-recipe");
 Object.defineProperty(exports, "typesafeRecipe", { enumerable: true, get: function () { return typesafe_recipe_1.typesafeRecipe; } });
 Object.defineProperty(exports, "estimateFromState", { enumerable: true, get: function () { return typesafe_recipe_1.estimateFromState; } });

package/dist/internal.d.ts

@@ -26,7 +26,7 @@ export { noCustomer, overrideContext, resolveContext, isEnforced } from './core/
 export type { ResolvedContext } from './core/context';
 export { GovernanceSession } from './core/session';
 export type { SessionDeps } from './core/session';
-export { ANTHROPIC_METRICS, GEMINI_METRICS, MISTRAL_METRICS, MISTRAL_METRIC_NAME_MAP, OLLAMA_METRICS, TYPESAFE_METRICS, GATEWAY_METRICS, OPENAI_CHAT_METRICS, OPENAI_METRICS, OPENAI_RESPONSES_METRICS, METRIC_PATHS, metricName, metricNamesMatchPaths, pickMetrics, pickRenamedMetrics, addMetrics, addTokenUsage, fromFlatTokenUsage, fromGeminiUsage, fromLangChainUsage, fromMistralUsage, fromOllamaUsage, fromOpenAIUsage, fromResponsesUsage, observedMeasurement, } from './bindings/shared/token-usage';
+export { ANTHROPIC_METRICS, GEMINI_METRICS, MISTRAL_METRICS, MISTRAL_METRIC_NAME_MAP, OLLAMA_METRICS, TYPESAFE_METRICS, GATEWAY_METRICS, DATABRICKS_METRICS, DATABRICKS_QUERY_METRIC_NAME_MAP, WORKERS_AI_METRICS, OPENAI_CHAT_METRICS, OPENAI_METRICS, OPENAI_RESPONSES_METRICS, METRIC_PATHS, metricName, metricNamesMatchPaths, pickMetrics, pickRenamedMetrics, addMetrics, addTokenUsage, fromFlatTokenUsage, fromGeminiUsage, fromLangChainUsage, fromMistralUsage, fromOllamaUsage, fromOpenAIUsage, fromResponsesUsage, observedMeasurement, } from './bindings/shared/token-usage';
 export type { ObservedVendor } from './bindings/shared/token-usage';
 export { DEFAULT_RECIPES } from './stigg-agent-sdk';
 export { meterStream } from './bindings/shared/meter-stream';

package/dist/internal.js

@@ -11,7 +11,8 @@
  * can do from the public barrel, do from there instead.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.eventNameFor = exports.meterStream = exports.DEFAULT_RECIPES = exports.observedMeasurement = exports.fromResponsesUsage = exports.fromOpenAIUsage = exports.fromOllamaUsage = exports.fromMistralUsage = exports.fromLangChainUsage = exports.fromGeminiUsage = exports.fromFlatTokenUsage = exports.addTokenUsage = exports.addMetrics = exports.pickRenamedMetrics = exports.pickMetrics = exports.metricNamesMatchPaths = exports.metricName = exports.METRIC_PATHS = exports.OPENAI_RESPONSES_METRICS = exports.OPENAI_METRICS = exports.OPENAI_CHAT_METRICS = exports.GATEWAY_METRICS = exports.TYPESAFE_METRICS = exports.OLLAMA_METRICS = exports.MISTRAL_METRIC_NAME_MAP = exports.MISTRAL_METRICS = exports.GEMINI_METRICS = exports.ANTHROPIC_METRICS = exports.GovernanceSession = exports.isEnforced = exports.resolveContext = exports.overrideContext = exports.noCustomer = exports.prefixed = exports.asEngineError = exports.isMeterMismatchError = exports.isTransportFailure = exports.resolveConfig = exports.unenforcedDecision = exports.toDecision = exports.remedyFor = exports.policyDecision = exports.withDecisionText = exports.describeResolution = exports.RestGovernanceEngine = exports.NodeSdkEngine = exports.CompositeEngine = exports.createGovernanceEngine = void 0;
+exports.meterStream = exports.DEFAULT_RECIPES = exports.observedMeasurement = exports.fromResponsesUsage = exports.fromOpenAIUsage = exports.fromOllamaUsage = exports.fromMistralUsage = exports.fromLangChainUsage = exports.fromGeminiUsage = exports.fromFlatTokenUsage = exports.addTokenUsage = exports.addMetrics = exports.pickRenamedMetrics = exports.pickMetrics = exports.metricNamesMatchPaths = exports.metricName = exports.METRIC_PATHS = exports.OPENAI_RESPONSES_METRICS = exports.OPENAI_METRICS = exports.OPENAI_CHAT_METRICS = exports.WORKERS_AI_METRICS = exports.DATABRICKS_QUERY_METRIC_NAME_MAP = exports.DATABRICKS_METRICS = exports.GATEWAY_METRICS = exports.TYPESAFE_METRICS = exports.OLLAMA_METRICS = exports.MISTRAL_METRIC_NAME_MAP = exports.MISTRAL_METRICS = exports.GEMINI_METRICS = exports.ANTHROPIC_METRICS = exports.GovernanceSession = exports.isEnforced = exports.resolveContext = exports.overrideContext = exports.noCustomer = exports.prefixed = exports.asEngineError = exports.isMeterMismatchError = exports.isTransportFailure = exports.resolveConfig = exports.unenforcedDecision = exports.toDecision = exports.remedyFor = exports.policyDecision = exports.withDecisionText = exports.describeResolution = exports.RestGovernanceEngine = exports.NodeSdkEngine = exports.CompositeEngine = exports.createGovernanceEngine = void 0;
+exports.eventNameFor = void 0;
 // The engine seam's implementations. The interface itself is public - see `index.ts`.
 var create_engine_1 = require("./engine/create-engine");
 Object.defineProperty(exports, "createGovernanceEngine", { enumerable: true, get: function () { return create_engine_1.createGovernanceEngine; } });
@@ -58,6 +59,9 @@ Object.defineProperty(exports, "MISTRAL_METRIC_NAME_MAP", { enumerable: true, ge
 Object.defineProperty(exports, "OLLAMA_METRICS", { enumerable: true, get: function () { return token_usage_1.OLLAMA_METRICS; } });
 Object.defineProperty(exports, "TYPESAFE_METRICS", { enumerable: true, get: function () { return token_usage_1.TYPESAFE_METRICS; } });
 Object.defineProperty(exports, "GATEWAY_METRICS", { enumerable: true, get: function () { return token_usage_1.GATEWAY_METRICS; } });
+Object.defineProperty(exports, "DATABRICKS_METRICS", { enumerable: true, get: function () { return token_usage_1.DATABRICKS_METRICS; } });
+Object.defineProperty(exports, "DATABRICKS_QUERY_METRIC_NAME_MAP", { enumerable: true, get: function () { return token_usage_1.DATABRICKS_QUERY_METRIC_NAME_MAP; } });
+Object.defineProperty(exports, "WORKERS_AI_METRICS", { enumerable: true, get: function () { return token_usage_1.WORKERS_AI_METRICS; } });
 Object.defineProperty(exports, "OPENAI_CHAT_METRICS", { enumerable: true, get: function () { return token_usage_1.OPENAI_CHAT_METRICS; } });
 Object.defineProperty(exports, "OPENAI_METRICS", { enumerable: true, get: function () { return token_usage_1.OPENAI_METRICS; } });
 Object.defineProperty(exports, "OPENAI_RESPONSES_METRICS", { enumerable: true, get: function () { return token_usage_1.OPENAI_RESPONSES_METRICS; } });

package/dist/stigg-agent-sdk.d.ts

@@ -77,7 +77,7 @@ export declare const DEFAULT_RECIPES: ({
         } | undefined;
         attachStream: (stream: any, ingest: (measurement: import("./types").Measurement) => void) => void;
     }[];
-} | import("./bindings/client/core/types").ClientRecipe<"cached_tokens" | "prompt_tokens" | "completion_tokens" | "input_tokens" | "output_tokens" | "total_tokens" | "cost"> | {
+} | import("./bindings/client/core/types").ClientRecipe<"cached_tokens" | "prompt_tokens" | "completion_tokens" | "input_tokens" | "output_tokens" | "total_tokens" | "cost"> | import("./bindings/client/core/types").ClientRecipe<"prompt_tokens" | "completion_tokens" | "input_tokens" | "output_tokens" | "cache_read_input_tokens" | "cache_creation_input_tokens" | "total_tokens" | "reasoning_tokens"> | import("./bindings/client/core/types").ClientRecipe<"cached_tokens" | "prompt_tokens" | "completion_tokens" | "input_tokens" | "output_tokens" | "cache_read_input_tokens" | "cache_creation_input_tokens" | "total_tokens"> | {
     vendor: string;
     metrics: readonly ["prompt_tokens", "completion_tokens", "cached_tokens", "total_tokens"];
     totalTokensMetric: "total_tokens";

package/dist/stigg-agent-sdk.js

@@ -25,6 +25,8 @@ const dimensions_1 = require("./core/dimensions");
 const anthropic_recipe_1 = require("./bindings/client/providers/anthropic/anthropic-recipe");
 const gemini_recipe_1 = require("./bindings/client/providers/gemini/gemini-recipe");
 const litellm_recipe_1 = require("./bindings/client/gateways/litellm/litellm-recipe");
+const databricks_recipe_1 = require("./bindings/client/gateways/databricks/databricks-recipe");
+const cloudflare_recipe_1 = require("./bindings/client/gateways/cloudflare/cloudflare-recipe");
 const groq_recipe_1 = require("./bindings/client/providers/groq/groq-recipe");
 const mistral_recipe_1 = require("./bindings/client/providers/mistral/mistral-recipe");
 const ollama_recipe_1 = require("./bindings/client/providers/ollama/ollama-recipe");
@@ -55,9 +57,20 @@ const types_1 = require("./types");
  * tokens is that mapper's business, and stays the caller's to decide.
  */
 exports.DEFAULT_RECIPES = [
+    // First, and that position is load-bearing. Cloudflare AI Gateway forwards the upstream body
+    // untouched, so an `Anthropic` client pointed at it is an Anthropic client by shape and only the
+    // hostname says otherwise - and `anthropicRecipe` has no base-URL guard, so asked first it would
+    // claim that client and meter gateway traffic as direct Anthropic traffic. The same argument puts
+    // it ahead of `openAIRecipe` for its OpenAI-shaped routes and for Workers AI's compatible one.
+    (0, cloudflare_recipe_1.cloudflareRecipe)(),
     anthropic_recipe_1.anthropicRecipe,
     (0, openrouter_recipe_1.openrouterRecipe)(),
     (0, litellm_recipe_1.litellmRecipe)(),
+    // The Databricks workspace route, matched on host *and* serving path - the host families alone
+    // are every Databricks API there is, most of which are not inference. Only the hosted recipe is
+    // here: `databricksExternalRecipe` covers the same wire reaching an endpoint billed by somebody
+    // else, which no matcher can tell apart, so it is named at `wrap` rather than guessed at.
+    (0, databricks_recipe_1.databricksRecipe)(),
     // Ahead of `openAIRecipe`, like the gateways above and for the same reason: shape cannot tell a
     // Groq client from any other OpenAI-compatible one, so the hostname decides, and the recipe that
     // reads the hostname has to be asked first or the generic one claims the client.

package/dist/types.d.ts

@@ -454,6 +454,14 @@ export type ChatGatedPath = ['sendMessage'] | ['sendMessageStream'];
  * `ollamaRecipe.matches` requires.
  */
 export type OllamaGatedPath = ['chat'] | ['generate'];
+/**
+ * Databricks' first-party client, whose whole surface is `query`. Out of {@link GatedPath} for the
+ * same reason as the two above, and more urgently: `query` is not merely a common method name, it
+ * is the one every database client in the ecosystem already owns. In the fallback list it would
+ * type `stiggContext` onto a `pg`, `mysql2`, Snowflake or BigQuery client - objects no recipe
+ * matches - and that context would be forwarded to the driver as an unknown field.
+ */
+export type DatabricksGatedPath = ['query'];
 /**
  * The paths to overlay when no recipe was named, given what the client actually looks like.
  *
@@ -509,9 +517,35 @@ type GatedVendorShape = {
  * the vendor as an unknown field. So anything the registry would claim first is excluded here,
  * mirroring that precedence.
  */
-export type ChatOf<T> = T extends GatedVendorShape ? GatedPath : T extends OllamaShape ? GatedPath | OllamaGatedPath : T extends {
+export type ChatOf<T> = T extends GatedVendorShape ? GatedPath : T extends OllamaShape ? GatedPath | OllamaGatedPath : T extends DatabricksQueryShape ? GatedPath | DatabricksGatedPath : T extends {
     getHistory: (...args: any[]) => any;
 } ? GatedPath | ChatGatedPath : GatedPath;
+/**
+ * What `databricksRecipe.matches` looks for, as near as a type can get - and the gap is the point.
+ *
+ * The runtime guard is `query` **plus** a `userAgent` naming the Databricks SDK. That second
+ * marker cannot be written here: `userAgent` is declared `private` in the client's `.d.ts`, and a
+ * `private` member makes a class non-assignable to any structural type naming it. So unlike
+ * {@link OllamaShape}, which mirrors its matcher exactly, this keys on the one thing the client
+ * does publish - the shape of what `query` resolves to.
+ *
+ * That still discriminates, which was checked rather than assumed: `QueryEndpointResponse` has no
+ * required property, so a target of all-optional fields would normally be a weak type that
+ * anything satisfies - but a `pg` or Snowflake result shares *none* of these names, and TypeScript's
+ * weak-type rule rejects exactly that case. `created?: bigint` is the field doing most of the work;
+ * nothing else in the ecosystem returns a bigint timestamp beside a `servedModelName`.
+ *
+ * The residual gap is a `query` typed as `Promise<any>`, which matches everything by construction.
+ * That costs a context offered on a method the runtime will decline to gate - the type is wider
+ * than the matcher, never narrower, so nothing is gated that was not typed.
+ */
+type DatabricksQueryShape = {
+    query: (...args: any[]) => Promise<{
+        servedModelName?: string | undefined;
+        predictions?: unknown;
+        created?: bigint | undefined;
+    }>;
+};
 /** What `ollamaRecipe.matches` looks for, as a type. Kept in step with it by hand. */
 type OllamaShape = {
     chat: (...args: any[]) => any;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stigg/agent-sdk-js",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Stigg agent SDK (JS) - run-lifecycle governance for agentic stacks, with examples per layer",
   "main": "dist/index.js",
   "typings": "dist/index.d.ts",
@@ -40,12 +40,15 @@
     "example-client-ollama": "tsx examples/client/ollama.ts",
     "example-client-openrouter": "tsx examples/client/openrouter.ts",
     "example-client-litellm": "tsx examples/client/litellm.ts",
+    "example-client-databricks": "tsx examples/client/databricks.ts",
     "example-client-typesafe": "tsx examples/client/typesafe.ts",
     "example-two-agents": "tsx examples/client/two-agents.ts",
     "example-gateway": "tsx examples/gateway/openrouter.ts",
     "example-gateway-openrouter": "tsx examples/gateway/openrouter.ts",
     "example-gateway-litellm": "tsx examples/gateway/litellm.ts",
-    "example-gateway-bedrock": "tsx examples/gateway/bedrock.ts"
+    "example-gateway-cloudflare": "tsx examples/gateway/cloudflare.ts",
+    "example-gateway-bedrock": "tsx examples/gateway/bedrock.ts",
+    "example-gateway-databricks": "tsx examples/gateway/databricks.ts"
   },
   "nx": {
     "includedScripts": [
@@ -68,10 +71,12 @@
   "devDependencies": {
     "@anthropic-ai/sdk": "^0.124.0",
     "@aws-sdk/client-bedrock-runtime": "^3.1150.0",
+    "@databricks/sdk-modelservingquery": "0.25.0",
     "@google/genai": "^2.28.0",
     "@mistralai/mistralai": "^2.7.0",
     "@openrouter/sdk": "^1.4.25",
     "@typesafe-ai/sdk": "^0.6.0",
+    "cloudflare": "^7.3.0",
     "dotenv": "^17.4.2",
     "groq-sdk": "^1.6.0",
     "ollama": "^0.6.4",