@axlsdk/axl 0.17.8 → 0.18.0

package/dist/index.d.cts

@@ -1402,6 +1402,100 @@ declare class AxlEventBus implements AsyncIterable<AxlEvent> {
     get isFinished(): boolean;
 }
 
+/**
+ * Opt-in client-side rate governor for provider HTTP calls.
+ *
+ * Wraps the single `fetchWithRetry` chokepoint (see `retry.ts`) so a configured
+ * provider can bound how many requests it has *in flight* at once and, optionally,
+ * how closely spaced its request *initiations* are. This is proactive pacing —
+ * complementary to, not a replacement for, the reactive 429/503/529 backoff that
+ * `fetchWithRetry` already does. The motivating case: eval fan-out
+ * (`concurrency × scorerConcurrency`, up to 25 judge calls by default) on a
+ * shared API key can storm a provider's rate limit; a governor smooths that.
+ *
+ * ## Scope & caveats (v1)
+ * - **Caps request concurrency, not token throughput (TPM).** Streaming `chat()`
+ *   releases the permit at *headers* (the adapter gets the `Response` from
+ *   `fetchWithRetry`, then iterates the body separately), so a slow stream does
+ *   NOT hold a permit for its whole lifetime. Non-streaming `chat()` holds the
+ *   permit through `res.json()`. Neither bounds tokens/min.
+ * - **Per provider instance.** Providers are singletons per (runtime, provider
+ *   type), so one governor governs all calls through that adapter — but NOT
+ *   embedder calls (the embedder is constructed outside the registry) and NOT
+ *   other processes/runtimes sharing the same API key.
+ * - **`minIntervalMs` is global spacing, not a burst bucket.** A single
+ *   last-grant timestamp gates every grant: a permit may be free yet a grant
+ *   still waits out the interval. There is no accumulated burst allowance.
+ * - **Nesting is safe.** A permit is held only across one `fetchWithRetry` call,
+ *   never across a nested `ctx.ask`, so an agent-as-tool chain on the same
+ *   provider under `maxConcurrent: 1` still completes — permits don't stack.
+ *
+ * `observe(res)` is a no-op seam in v1 for a future adaptive (header-driven)
+ * pacing follow-up.
+ */
+/** Configuration for a provider's {@link RateLimiter}. All fields optional. */
+type RateLimitConfig = {
+    /**
+     * Maximum requests in flight at once for this provider. Must be a finite
+     * integer ≥ 1; invalid values disable the concurrency cap (with a warning).
+     * `1` serializes all requests (allowed, with a warning — it's a throughput
+     * floor, not a deadlock: nested same-provider calls still complete because a
+     * permit is never held across a nested ask).
+     */
+    maxConcurrent?: number;
+    /**
+     * Minimum milliseconds between successive request *grants* (global spacing,
+     * not a per-permit rate). A permit can be free yet a grant still waits out
+     * this interval. No burst bucket.
+     */
+    minIntervalMs?: number;
+    /**
+     * If set, `acquire()` rejects (fail loud) when a caller has queued longer than
+     * this many ms, instead of waiting indefinitely — surfaces a misconfigured
+     * cap rather than silently hanging the run.
+     */
+    acquireTimeoutMs?: number;
+};
+/**
+ * Counting semaphore + FIFO waiter queue. Dependency-free. Not shared across
+ * processes — purely an in-process pacing aid.
+ */
+declare class RateLimiter {
+    private readonly maxConcurrent;
+    private readonly minIntervalMs;
+    private readonly acquireTimeoutMs?;
+    private active;
+    private lastGrantAt;
+    private readonly queue;
+    private spacingTimer?;
+    private warnedQueued;
+    constructor(config?: RateLimitConfig);
+    /**
+     * Acquire a permit. Resolves when one is granted (respecting `maxConcurrent`
+     * and `minIntervalMs`). The caller MUST pair every resolved `acquire()` with
+     * exactly one {@link release}.
+     *
+     * Abort handling:
+     * - Pre-aborted signal ⇒ rejects immediately, no permit taken.
+     * - Aborted while queued ⇒ spliced out of the queue, no counter change, rejects.
+     * - (Once granted, the abort listener is removed — the holder owns the permit
+     *   and releases it via `fetchWithRetry`'s `finally`.)
+     *
+     * With `acquireTimeoutMs`, a waiter that sits in the queue too long rejects.
+     */
+    acquire(signal?: AbortSignal): Promise<void>;
+    /** Release a previously acquired permit and wake the next waiter. */
+    release(): void;
+    /** v1 no-op seam for a future adaptive (rate-limit-header-driven) pacing follow-up. */
+    observe(_res: Response): void;
+    /** Drain the queue: grant head waiters while a permit is free and spacing allows. */
+    private pump;
+    private grant;
+    /** Reject a still-queued waiter (abort or timeout) and remove it cleanly. */
+    private settleWaiter;
+    private clearWaiterTimers;
+}
+
 /** A pending human decision awaiting resolution. */
 type PendingDecision = {
     executionId: string;
@@ -1645,6 +1739,15 @@ type MemoryConfig = {
 type ProviderConfig = {
     apiKey?: string;
     baseUrl?: string;
+    /**
+     * Opt-in client-side rate governor for this provider's HTTP calls (see
+     * {@link RateLimitConfig}). Bounds in-flight request concurrency (and,
+     * optionally, request spacing) through the shared `fetchWithRetry` chokepoint.
+     * Omitted ⇒ no governor (behavior unchanged). Caveat: governs **chat** calls
+     * through this provider instance only — NOT memory-embedder calls (constructed
+     * outside the registry) and NOT other processes sharing the same API key.
+     */
+    rateLimit?: RateLimitConfig;
 };
 /** MCP server configuration */
 type McpServerConfig$1 = {
@@ -2786,6 +2889,26 @@ type EvalProgressEventShape = {
     totalItems: number;
     failures: number;
 };
+/**
+ * Structural shape of `@axlsdk/eval`'s `EvalConfig` as seen by the runtime.
+ * `axl` is a peer of `eval` (eval depends on axl, not the reverse), so the core
+ * can't import the real `EvalConfig` type — it forwards `config` to `runEval`
+ * at runtime regardless. This single declaration is the ONE place to keep in
+ * sync with `EvalConfig`'s forwarded fields, so `eval()` and `runRegisteredEval()`
+ * don't drift apart (they previously carried two hand-copied inline shapes).
+ * `dataset`/`scorers` stay `unknown[]`/`unknown` because their real types live in
+ * `@axlsdk/eval`; only the primitive knobs are structurally typed here.
+ */
+type RuntimeEvalConfigShape = {
+    workflow: string;
+    dataset: unknown;
+    scorers: unknown[];
+    concurrency?: number;
+    scorerConcurrency?: number;
+    budget?: string;
+    failOnScorerErrorRate?: number;
+    metadata?: Record<string, unknown>;
+};
 type CreateContextOptions = {
     metadata?: Record<string, unknown>;
     /** Cost budget for the context (e.g., '$0.50'). Enforced via finish_and_stop policy. */
@@ -3160,14 +3283,7 @@ declare class AxlRuntime extends EventEmitter {
      *
      * @see Spec Section 13.5
      */
-    eval(config: {
-        workflow: string;
-        dataset: unknown;
-        scorers: unknown[];
-        concurrency?: number;
-        budget?: string;
-        metadata?: Record<string, unknown>;
-    }, options?: {
+    eval(config: RuntimeEvalConfigShape, options?: {
         onProgress?: (event: EvalProgressEventShape) => void;
         signal?: AbortSignal;
         captureTraces?: boolean;
@@ -3749,9 +3865,11 @@ declare class OpenAIProvider implements Provider {
     readonly name = "openai";
     private baseUrl;
     private apiKey;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3773,9 +3891,11 @@ declare class OpenAIResponsesProvider implements Provider {
     readonly name = "openai-responses";
     private baseUrl;
     private apiKey;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3814,9 +3934,11 @@ declare class AnthropicProvider implements Provider {
     private baseUrl;
     private apiKey;
     private currentModel?;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3876,9 +3998,11 @@ declare class GeminiProvider implements Provider {
     private baseUrl;
     private apiKey;
     private callCounter;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -4361,4 +4485,4 @@ declare class NoopSpanManager implements SpanManager {
  */
 declare function createSpanManager(config?: TelemetryConfig): Promise<SpanManager>;
 
-export { AXL_EVENT_TYPES, type Agent, type AgentCallEndData, type AgentCallInfo, type AgentCallParams, type AgentCallStartData, type AgentConfig, AnthropicProvider, type AnyWorkflow, type AskOptions, type AskScoped, type AwaitHumanData, type AwaitHumanOptions, type AwaitHumanResolvedData, type AxlConfig, AxlError, type AxlEvent, type AxlEventBase, AxlEventBus, type AxlEventOf, type AxlEventType, AxlRuntime, AxlStream, BudgetExceededError, type BudgetOptions, type BudgetResult, CONFIG_CANDIDATES, COST_BEARING_LEAF_TYPES, type CallbackMeta, type ChatMessage, type ChatOptions, type CheckpointEventData, type CoalescedPartialObject, type CreateContextOptions, type DelegateData, type DelegateOptions, type Effort, type EmbedResult, type EmbedUsage, type Embedder, type EvalExecuteWorkflow, type EvalHistoryEntry, type EventStreamOptions, EventStreamOverflowError, type ExecuteOptions, type ExecutionInfo, type ExecutionState, GeminiProvider, type GuardrailBlockHandler, type GuardrailData, GuardrailError, type GuardrailResult, type GuardrailsConfig, type HandoffDescriptor, type HandoffRecord, type HandoffReturnData, type HandoffStartData, type HumanDecision, InMemoryVectorStore, type InputGuardrail, type MapOptions, MaxTurnsError, McpManager, type McpServer, type McpServerConfig, type McpToolDefinition, type McpToolResult, type MemoryConfig, type MemoryEventData, MemoryManager, MemoryStore, NoConsensus, NoopSpanManager, OpenAIEmbedder, OpenAIProvider, OpenAIResponsesProvider, type OutputGuardrail, type OutputValidator, type PendingDecision, type Provider, type ProviderAdapter, ProviderRegistry, type ProviderResponse, QuorumNotMet, REDACTED, REDACTION_RULES, type RaceOptions, type RecallOptions, type RecallResult, RedisStore, type RememberOptions, type RememberResult, type ResolvedThinkingOptions, type Result, type RetryPolicy, SQLiteStore, type SchemaCheckData, Session, type SessionOptions, type SpanHandle, type SpanManager, type SpawnOptions, SqliteVectorStore, type StateStore, type StreamChunk, type StringDeltaData, type StringStreamEvent, type StringStreamFilter, type TelemetryConfig, TimeoutError, type Tool, type ToolApprovalData, type ToolCallData, type ToolCallMessage, type ToolCallStartData, type ToolChoice, type ToolConfig, ToolDenied, type ToolDeniedData, type ToolHooks, type ValidateData, type ValidateResult, ValidationError, type VectorEntry, type VectorResult, type VectorStore, type VerifyData, VerifyError, type VerifyOptions, type VerifyRetry, type VoteOptions, type Workflow, type WorkflowConfig, WorkflowContext, type WorkflowContextInit, type WorkflowEndData, type WorkflowStartData, agent, createSpanManager, defineConfig, eventCostContribution, expandGlob, extractJson, findConfig, importModule, isCostBearingLeaf, isRootLevel, needsTsxLoader, parsePartialJson, pickDefault, pickExport, redactEvent, registerConditions, resolveRuntime, resolveThinkingOptions, stringStreamFromEvents, tool, workflow, zodToJsonSchema };
+export { AXL_EVENT_TYPES, type Agent, type AgentCallEndData, type AgentCallInfo, type AgentCallParams, type AgentCallStartData, type AgentConfig, AnthropicProvider, type AnyWorkflow, type AskOptions, type AskScoped, type AwaitHumanData, type AwaitHumanOptions, type AwaitHumanResolvedData, type AxlConfig, AxlError, type AxlEvent, type AxlEventBase, AxlEventBus, type AxlEventOf, type AxlEventType, AxlRuntime, AxlStream, BudgetExceededError, type BudgetOptions, type BudgetResult, CONFIG_CANDIDATES, COST_BEARING_LEAF_TYPES, type CallbackMeta, type ChatMessage, type ChatOptions, type CheckpointEventData, type CoalescedPartialObject, type CreateContextOptions, type DelegateData, type DelegateOptions, type Effort, type EmbedResult, type EmbedUsage, type Embedder, type EvalExecuteWorkflow, type EvalHistoryEntry, type EventStreamOptions, EventStreamOverflowError, type ExecuteOptions, type ExecutionInfo, type ExecutionState, GeminiProvider, type GuardrailBlockHandler, type GuardrailData, GuardrailError, type GuardrailResult, type GuardrailsConfig, type HandoffDescriptor, type HandoffRecord, type HandoffReturnData, type HandoffStartData, type HumanDecision, InMemoryVectorStore, type InputGuardrail, type MapOptions, MaxTurnsError, McpManager, type McpServer, type McpServerConfig, type McpToolDefinition, type McpToolResult, type MemoryConfig, type MemoryEventData, MemoryManager, MemoryStore, NoConsensus, NoopSpanManager, OpenAIEmbedder, OpenAIProvider, OpenAIResponsesProvider, type OutputGuardrail, type OutputValidator, type PendingDecision, type Provider, type ProviderAdapter, ProviderRegistry, type ProviderResponse, QuorumNotMet, REDACTED, REDACTION_RULES, type RaceOptions, type RateLimitConfig, RateLimiter, type RecallOptions, type RecallResult, RedisStore, type RememberOptions, type RememberResult, type ResolvedThinkingOptions, type Result, type RetryPolicy, SQLiteStore, type SchemaCheckData, Session, type SessionOptions, type SpanHandle, type SpanManager, type SpawnOptions, SqliteVectorStore, type StateStore, type StreamChunk, type StringDeltaData, type StringStreamEvent, type StringStreamFilter, type TelemetryConfig, TimeoutError, type Tool, type ToolApprovalData, type ToolCallData, type ToolCallMessage, type ToolCallStartData, type ToolChoice, type ToolConfig, ToolDenied, type ToolDeniedData, type ToolHooks, type ValidateData, type ValidateResult, ValidationError, type VectorEntry, type VectorResult, type VectorStore, type VerifyData, VerifyError, type VerifyOptions, type VerifyRetry, type VoteOptions, type Workflow, type WorkflowConfig, WorkflowContext, type WorkflowContextInit, type WorkflowEndData, type WorkflowStartData, agent, createSpanManager, defineConfig, eventCostContribution, expandGlob, extractJson, findConfig, importModule, isCostBearingLeaf, isRootLevel, needsTsxLoader, parsePartialJson, pickDefault, pickExport, redactEvent, registerConditions, resolveRuntime, resolveThinkingOptions, stringStreamFromEvents, tool, workflow, zodToJsonSchema };

package/dist/index.d.ts

@@ -1402,6 +1402,100 @@ declare class AxlEventBus implements AsyncIterable<AxlEvent> {
     get isFinished(): boolean;
 }
 
+/**
+ * Opt-in client-side rate governor for provider HTTP calls.
+ *
+ * Wraps the single `fetchWithRetry` chokepoint (see `retry.ts`) so a configured
+ * provider can bound how many requests it has *in flight* at once and, optionally,
+ * how closely spaced its request *initiations* are. This is proactive pacing —
+ * complementary to, not a replacement for, the reactive 429/503/529 backoff that
+ * `fetchWithRetry` already does. The motivating case: eval fan-out
+ * (`concurrency × scorerConcurrency`, up to 25 judge calls by default) on a
+ * shared API key can storm a provider's rate limit; a governor smooths that.
+ *
+ * ## Scope & caveats (v1)
+ * - **Caps request concurrency, not token throughput (TPM).** Streaming `chat()`
+ *   releases the permit at *headers* (the adapter gets the `Response` from
+ *   `fetchWithRetry`, then iterates the body separately), so a slow stream does
+ *   NOT hold a permit for its whole lifetime. Non-streaming `chat()` holds the
+ *   permit through `res.json()`. Neither bounds tokens/min.
+ * - **Per provider instance.** Providers are singletons per (runtime, provider
+ *   type), so one governor governs all calls through that adapter — but NOT
+ *   embedder calls (the embedder is constructed outside the registry) and NOT
+ *   other processes/runtimes sharing the same API key.
+ * - **`minIntervalMs` is global spacing, not a burst bucket.** A single
+ *   last-grant timestamp gates every grant: a permit may be free yet a grant
+ *   still waits out the interval. There is no accumulated burst allowance.
+ * - **Nesting is safe.** A permit is held only across one `fetchWithRetry` call,
+ *   never across a nested `ctx.ask`, so an agent-as-tool chain on the same
+ *   provider under `maxConcurrent: 1` still completes — permits don't stack.
+ *
+ * `observe(res)` is a no-op seam in v1 for a future adaptive (header-driven)
+ * pacing follow-up.
+ */
+/** Configuration for a provider's {@link RateLimiter}. All fields optional. */
+type RateLimitConfig = {
+    /**
+     * Maximum requests in flight at once for this provider. Must be a finite
+     * integer ≥ 1; invalid values disable the concurrency cap (with a warning).
+     * `1` serializes all requests (allowed, with a warning — it's a throughput
+     * floor, not a deadlock: nested same-provider calls still complete because a
+     * permit is never held across a nested ask).
+     */
+    maxConcurrent?: number;
+    /**
+     * Minimum milliseconds between successive request *grants* (global spacing,
+     * not a per-permit rate). A permit can be free yet a grant still waits out
+     * this interval. No burst bucket.
+     */
+    minIntervalMs?: number;
+    /**
+     * If set, `acquire()` rejects (fail loud) when a caller has queued longer than
+     * this many ms, instead of waiting indefinitely — surfaces a misconfigured
+     * cap rather than silently hanging the run.
+     */
+    acquireTimeoutMs?: number;
+};
+/**
+ * Counting semaphore + FIFO waiter queue. Dependency-free. Not shared across
+ * processes — purely an in-process pacing aid.
+ */
+declare class RateLimiter {
+    private readonly maxConcurrent;
+    private readonly minIntervalMs;
+    private readonly acquireTimeoutMs?;
+    private active;
+    private lastGrantAt;
+    private readonly queue;
+    private spacingTimer?;
+    private warnedQueued;
+    constructor(config?: RateLimitConfig);
+    /**
+     * Acquire a permit. Resolves when one is granted (respecting `maxConcurrent`
+     * and `minIntervalMs`). The caller MUST pair every resolved `acquire()` with
+     * exactly one {@link release}.
+     *
+     * Abort handling:
+     * - Pre-aborted signal ⇒ rejects immediately, no permit taken.
+     * - Aborted while queued ⇒ spliced out of the queue, no counter change, rejects.
+     * - (Once granted, the abort listener is removed — the holder owns the permit
+     *   and releases it via `fetchWithRetry`'s `finally`.)
+     *
+     * With `acquireTimeoutMs`, a waiter that sits in the queue too long rejects.
+     */
+    acquire(signal?: AbortSignal): Promise<void>;
+    /** Release a previously acquired permit and wake the next waiter. */
+    release(): void;
+    /** v1 no-op seam for a future adaptive (rate-limit-header-driven) pacing follow-up. */
+    observe(_res: Response): void;
+    /** Drain the queue: grant head waiters while a permit is free and spacing allows. */
+    private pump;
+    private grant;
+    /** Reject a still-queued waiter (abort or timeout) and remove it cleanly. */
+    private settleWaiter;
+    private clearWaiterTimers;
+}
+
 /** A pending human decision awaiting resolution. */
 type PendingDecision = {
     executionId: string;
@@ -1645,6 +1739,15 @@ type MemoryConfig = {
 type ProviderConfig = {
     apiKey?: string;
     baseUrl?: string;
+    /**
+     * Opt-in client-side rate governor for this provider's HTTP calls (see
+     * {@link RateLimitConfig}). Bounds in-flight request concurrency (and,
+     * optionally, request spacing) through the shared `fetchWithRetry` chokepoint.
+     * Omitted ⇒ no governor (behavior unchanged). Caveat: governs **chat** calls
+     * through this provider instance only — NOT memory-embedder calls (constructed
+     * outside the registry) and NOT other processes sharing the same API key.
+     */
+    rateLimit?: RateLimitConfig;
 };
 /** MCP server configuration */
 type McpServerConfig$1 = {
@@ -2786,6 +2889,26 @@ type EvalProgressEventShape = {
     totalItems: number;
     failures: number;
 };
+/**
+ * Structural shape of `@axlsdk/eval`'s `EvalConfig` as seen by the runtime.
+ * `axl` is a peer of `eval` (eval depends on axl, not the reverse), so the core
+ * can't import the real `EvalConfig` type — it forwards `config` to `runEval`
+ * at runtime regardless. This single declaration is the ONE place to keep in
+ * sync with `EvalConfig`'s forwarded fields, so `eval()` and `runRegisteredEval()`
+ * don't drift apart (they previously carried two hand-copied inline shapes).
+ * `dataset`/`scorers` stay `unknown[]`/`unknown` because their real types live in
+ * `@axlsdk/eval`; only the primitive knobs are structurally typed here.
+ */
+type RuntimeEvalConfigShape = {
+    workflow: string;
+    dataset: unknown;
+    scorers: unknown[];
+    concurrency?: number;
+    scorerConcurrency?: number;
+    budget?: string;
+    failOnScorerErrorRate?: number;
+    metadata?: Record<string, unknown>;
+};
 type CreateContextOptions = {
     metadata?: Record<string, unknown>;
     /** Cost budget for the context (e.g., '$0.50'). Enforced via finish_and_stop policy. */
@@ -3160,14 +3283,7 @@ declare class AxlRuntime extends EventEmitter {
      *
      * @see Spec Section 13.5
      */
-    eval(config: {
-        workflow: string;
-        dataset: unknown;
-        scorers: unknown[];
-        concurrency?: number;
-        budget?: string;
-        metadata?: Record<string, unknown>;
-    }, options?: {
+    eval(config: RuntimeEvalConfigShape, options?: {
         onProgress?: (event: EvalProgressEventShape) => void;
         signal?: AbortSignal;
         captureTraces?: boolean;
@@ -3749,9 +3865,11 @@ declare class OpenAIProvider implements Provider {
     readonly name = "openai";
     private baseUrl;
     private apiKey;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3773,9 +3891,11 @@ declare class OpenAIResponsesProvider implements Provider {
     readonly name = "openai-responses";
     private baseUrl;
     private apiKey;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3814,9 +3934,11 @@ declare class AnthropicProvider implements Provider {
     private baseUrl;
     private apiKey;
     private currentModel?;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -3876,9 +3998,11 @@ declare class GeminiProvider implements Provider {
     private baseUrl;
     private apiKey;
     private callCounter;
+    private governor?;
     constructor(options?: {
         apiKey?: string;
         baseUrl?: string;
+        rateLimit?: RateLimitConfig;
     });
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
@@ -4361,4 +4485,4 @@ declare class NoopSpanManager implements SpanManager {
  */
 declare function createSpanManager(config?: TelemetryConfig): Promise<SpanManager>;
 
-export { AXL_EVENT_TYPES, type Agent, type AgentCallEndData, type AgentCallInfo, type AgentCallParams, type AgentCallStartData, type AgentConfig, AnthropicProvider, type AnyWorkflow, type AskOptions, type AskScoped, type AwaitHumanData, type AwaitHumanOptions, type AwaitHumanResolvedData, type AxlConfig, AxlError, type AxlEvent, type AxlEventBase, AxlEventBus, type AxlEventOf, type AxlEventType, AxlRuntime, AxlStream, BudgetExceededError, type BudgetOptions, type BudgetResult, CONFIG_CANDIDATES, COST_BEARING_LEAF_TYPES, type CallbackMeta, type ChatMessage, type ChatOptions, type CheckpointEventData, type CoalescedPartialObject, type CreateContextOptions, type DelegateData, type DelegateOptions, type Effort, type EmbedResult, type EmbedUsage, type Embedder, type EvalExecuteWorkflow, type EvalHistoryEntry, type EventStreamOptions, EventStreamOverflowError, type ExecuteOptions, type ExecutionInfo, type ExecutionState, GeminiProvider, type GuardrailBlockHandler, type GuardrailData, GuardrailError, type GuardrailResult, type GuardrailsConfig, type HandoffDescriptor, type HandoffRecord, type HandoffReturnData, type HandoffStartData, type HumanDecision, InMemoryVectorStore, type InputGuardrail, type MapOptions, MaxTurnsError, McpManager, type McpServer, type McpServerConfig, type McpToolDefinition, type McpToolResult, type MemoryConfig, type MemoryEventData, MemoryManager, MemoryStore, NoConsensus, NoopSpanManager, OpenAIEmbedder, OpenAIProvider, OpenAIResponsesProvider, type OutputGuardrail, type OutputValidator, type PendingDecision, type Provider, type ProviderAdapter, ProviderRegistry, type ProviderResponse, QuorumNotMet, REDACTED, REDACTION_RULES, type RaceOptions, type RecallOptions, type RecallResult, RedisStore, type RememberOptions, type RememberResult, type ResolvedThinkingOptions, type Result, type RetryPolicy, SQLiteStore, type SchemaCheckData, Session, type SessionOptions, type SpanHandle, type SpanManager, type SpawnOptions, SqliteVectorStore, type StateStore, type StreamChunk, type StringDeltaData, type StringStreamEvent, type StringStreamFilter, type TelemetryConfig, TimeoutError, type Tool, type ToolApprovalData, type ToolCallData, type ToolCallMessage, type ToolCallStartData, type ToolChoice, type ToolConfig, ToolDenied, type ToolDeniedData, type ToolHooks, type ValidateData, type ValidateResult, ValidationError, type VectorEntry, type VectorResult, type VectorStore, type VerifyData, VerifyError, type VerifyOptions, type VerifyRetry, type VoteOptions, type Workflow, type WorkflowConfig, WorkflowContext, type WorkflowContextInit, type WorkflowEndData, type WorkflowStartData, agent, createSpanManager, defineConfig, eventCostContribution, expandGlob, extractJson, findConfig, importModule, isCostBearingLeaf, isRootLevel, needsTsxLoader, parsePartialJson, pickDefault, pickExport, redactEvent, registerConditions, resolveRuntime, resolveThinkingOptions, stringStreamFromEvents, tool, workflow, zodToJsonSchema };
+export { AXL_EVENT_TYPES, type Agent, type AgentCallEndData, type AgentCallInfo, type AgentCallParams, type AgentCallStartData, type AgentConfig, AnthropicProvider, type AnyWorkflow, type AskOptions, type AskScoped, type AwaitHumanData, type AwaitHumanOptions, type AwaitHumanResolvedData, type AxlConfig, AxlError, type AxlEvent, type AxlEventBase, AxlEventBus, type AxlEventOf, type AxlEventType, AxlRuntime, AxlStream, BudgetExceededError, type BudgetOptions, type BudgetResult, CONFIG_CANDIDATES, COST_BEARING_LEAF_TYPES, type CallbackMeta, type ChatMessage, type ChatOptions, type CheckpointEventData, type CoalescedPartialObject, type CreateContextOptions, type DelegateData, type DelegateOptions, type Effort, type EmbedResult, type EmbedUsage, type Embedder, type EvalExecuteWorkflow, type EvalHistoryEntry, type EventStreamOptions, EventStreamOverflowError, type ExecuteOptions, type ExecutionInfo, type ExecutionState, GeminiProvider, type GuardrailBlockHandler, type GuardrailData, GuardrailError, type GuardrailResult, type GuardrailsConfig, type HandoffDescriptor, type HandoffRecord, type HandoffReturnData, type HandoffStartData, type HumanDecision, InMemoryVectorStore, type InputGuardrail, type MapOptions, MaxTurnsError, McpManager, type McpServer, type McpServerConfig, type McpToolDefinition, type McpToolResult, type MemoryConfig, type MemoryEventData, MemoryManager, MemoryStore, NoConsensus, NoopSpanManager, OpenAIEmbedder, OpenAIProvider, OpenAIResponsesProvider, type OutputGuardrail, type OutputValidator, type PendingDecision, type Provider, type ProviderAdapter, ProviderRegistry, type ProviderResponse, QuorumNotMet, REDACTED, REDACTION_RULES, type RaceOptions, type RateLimitConfig, RateLimiter, type RecallOptions, type RecallResult, RedisStore, type RememberOptions, type RememberResult, type ResolvedThinkingOptions, type Result, type RetryPolicy, SQLiteStore, type SchemaCheckData, Session, type SessionOptions, type SpanHandle, type SpanManager, type SpawnOptions, SqliteVectorStore, type StateStore, type StreamChunk, type StringDeltaData, type StringStreamEvent, type StringStreamFilter, type TelemetryConfig, TimeoutError, type Tool, type ToolApprovalData, type ToolCallData, type ToolCallMessage, type ToolCallStartData, type ToolChoice, type ToolConfig, ToolDenied, type ToolDeniedData, type ToolHooks, type ValidateData, type ValidateResult, ValidationError, type VectorEntry, type VectorResult, type VectorStore, type VerifyData, VerifyError, type VerifyOptions, type VerifyRetry, type VoteOptions, type Workflow, type WorkflowConfig, WorkflowContext, type WorkflowContextInit, type WorkflowEndData, type WorkflowStartData, agent, createSpanManager, defineConfig, eventCostContribution, expandGlob, extractJson, findConfig, importModule, isCostBearingLeaf, isRootLevel, needsTsxLoader, parsePartialJson, pickDefault, pickExport, redactEvent, registerConditions, resolveRuntime, resolveThinkingOptions, stringStreamFromEvents, tool, workflow, zodToJsonSchema };