Crate agent_sdk_providers 

LLM provider trait, streaming primitives, and first-party provider implementations.

This crate defines the LlmProvider trait that all LLM backends implement, streaming types for incremental response processing, attachment validation, model capability metadata, and seven first-party provider implementations.

Provider Implementations

Provider	Module	Protocol
Anthropic (Messages API)	impls::anthropic	SSE streaming
OpenAI (Chat Completions)	impls::openai	SSE streaming
OpenAI (Responses API)	impls::openai_responses	SSE streaming
OpenAI Codex / ChatGPT	impls::openai_codex_responses	WebSocket
Google Gemini	impls::gemini	SSE streaming
Google Vertex AI	impls::vertex	SSE streaming
Cloudflare AI Gateway	impls::cloudflare_ai_gateway	Proxy wrapper

Re-exports

pub use fallback::FallbackProvider;

pub use provider::LlmProvider;

pub use provider::ModelInfo;

pub use provider::StructuredOutputSupport;

pub use provider::collect_stream;

pub use provider::probe_http_reachability;

pub use record_replay::RecordReplayMode;record-replay

pub use record_replay::RecordReplayProvider;record-replay

pub use refresh::RefreshingProvider;

pub use refresh::is_unauthorized_error;

pub use router::ModelRouter;

pub use router::ModelTier;

pub use router::TaskComplexity;

pub use streaming::StreamAccumulator;

pub use streaming::StreamBox;

pub use streaming::StreamDelta;

pub use structured::StructuredConfig;structured-output

pub use structured::StructuredOutput;structured-output

pub use structured::StructuredOutputError;structured-output

pub use structured::StructuredStream;structured-output

pub use structured::StructuredStreamUpdate;structured-output

pub use structured::run_structured;structured-output

pub use structured::run_structured_stream;structured-output

pub use impls::CloudflareAIGatewayProvider;cloudflare

pub use impls::GeminiProvider;gemini

pub use impls::OpenAICodexResponsesProvider;openai-codex

pub use impls::VertexProvider;vertex

pub use impls::AnthropicProvider;anthropic

pub use impls::is_oauth_token;anthropic

pub use impls::OpenAIAllowedToolsMode;openai

pub use impls::OpenAIApiSurface;openai

pub use impls::OpenAIPromptCacheMode;openai

pub use impls::OpenAIPromptCacheTtl;openai

pub use impls::OpenAIProvider;openai

pub use impls::OpenAIReasoningConfig;openai

pub use impls::OpenAIReasoningContext;openai

pub use impls::OpenAIReasoningEffort;openai

pub use impls::OpenAIReasoningMode;openai

pub use impls::OpenAIReasoningSummary;openai

pub use impls::OpenAIResponsesProvider;openai

pub use impls::OpenAITextVerbosity;openai

pub use impls::OpenAIToolChoice;openai

pub use impls::supports_reasoning_summary;openai

pub use model_capabilities::ModelCapabilities;

pub use model_capabilities::PricePoint;

pub use model_capabilities::Pricing;

pub use model_capabilities::SourceStatus;

pub use model_capabilities::get_model_capabilities;

pub use model_capabilities::supported_model_capabilities;

pub use model_features::ModelApiSurface;

pub use model_features::ModelFeatureSet;

pub use model_features::ModelFeatures;

pub use model_features::ModelInputModality;

pub use model_features::ModelPromptCacheFeatures;

pub use model_features::ModelPromptCacheMode;

pub use model_features::ModelPromptCacheTtl;

pub use model_features::ModelReasoningContext;

pub use model_features::ModelReasoningEffort;

pub use model_features::ModelReasoningFeatures;

pub use model_features::ModelReasoningMode;

pub use model_features::ModelReasoningStateReplay;

pub use model_features::ModelReasoningSummary;

pub use model_features::ModelToolChoice;

pub use model_features::ModelToolFeatures;

pub use model_features::get_model_features;

pub use model_features::supported_model_features;

pub use model_catalog::CatalogEntry;model-discovery

pub use model_catalog::ModelCatalogSource;model-discovery

pub use model_catalog::ModelRegistry;model-discovery

pub use model_catalog::ModelsDevSource;model-discovery

pub use model_catalog::OpenRouterSource;model-discovery

pub use model_catalog::PricingTier;model-discovery

pub use model_catalog::ResolvedModel;model-discovery

pub use model_catalog::ResolvedSource;model-discovery

pub use model_catalog::applicable_pricing;model-discovery

pub use search::BraveSearchProvider;

pub use search::SearchProvider;

pub use search::SearchResponse;

pub use search::SearchResult;

Modules

attachmentsanthropic or cloudflare or gemini or openai-codex or openai or vertex

fallback

Provider failover: try an ordered list of providers until one succeeds.

impls

First-party LLM provider implementations.

model_capabilities

model_catalogmodel-discovery

Dynamic model capability/pricing discovery from third-party feeds.

model_features

Static feature metadata for OpenAI models.

provider

LLM provider trait and streaming helpers.

record_replayrecord-replay

Deterministic record / replay provider wrapper (feature record-replay).

refresh

Provider wrapper that refreshes credentials on 401 and retries once.

router

search

Search provider trait and implementations.

streaming

Streaming types for LLM responses.

structuredstructured-output

Schema-validated structured output.

Structs

CacheConfig

Caller-facing control over provider-side prompt caching.

ChatRequest

ChatResponse

ContentSource

Source data for image and document content blocks.

Message

ResponseFormat

Requests that the model constrain its final answer to a JSON Schema.

SnapcompactFrameDigest

Content digest for one rendered Snapcompact frame artifact.

SnapcompactIntegrity

Content-integrity pins computed at a Snapcompact persist site.

SnapcompactMetadata

Exact-source metadata for a locally rendered Snapcompact checkpoint.

ThinkingConfig

Configuration for extended thinking.

Tool

Usage

Enums

CacheTtl

Time-to-live for a provider-side prompt-cache breakpoint.

ChatOutcome

Content

ContentBlock

A provider-compatible content block.

Effort

Effort level for adaptive thinking via output_config.

ImageDetail

Provider rendering detail requested for an image content block.

Role

ServedSpeed

Which speed tier a provider actually used, as reported back on the response.

SpeedTier

Inference speed tier — the “pay a premium for lower latency” knob.

StopReason

ThinkingDisplay

How thinking content is returned in responses.

ThinkingMode

The mode of extended thinking.

ToolChoice

Controls whether the model must use a tool.

Constants

COMPACTION_SUMMARY_PREFIX

Legacy on-disk marker used for rollback-readable compaction entries.

SNAPCOMPACT_HISTORY_IMAGE_WARNING

Fixed guard that separates rendered Snapcompact frames from active instructions.

USER_CANCELLED_TOOL_RESULT

Default tool_result text used to close a tool_use block the user cancelled (or otherwise abandoned) before it produced a real result.

Functions

balance_tool_results

Close every unanswered tool_use loop in messages.

canonical_snapcompact_checkpoint

Validates and returns the checkpoint metadata for one canonical Snapcompact replacement.

has_unbalanced_tool_use

True when messages contains a tool_use block whose id is not answered by any tool_result block anywhere in the conversation.

is_provider_valid_tool_sequence

Whether messages satisfy provider_tool_sequence_error_index.

orphaned_tool_result_message

Build the raw audit message appended by append-only orphan repair.

parse_retry_after

Parse the value of an HTTP Retry-After header into a Duration.

provider_tool_sequence_error_index

Return the first message index that violates provider tool-call ordering.

render_compaction_summary_for_provider

Render a typed compaction summary as inert historical data for providers.

sha256_hex

Lowercase hex SHA-256 of bytes.

snapcompact_integrity

Computes the integrity pins for a Snapcompact source and its rendered frames, given (frame_artifact_id, png_bytes) pairs in declared order.