Crate forge_guardrails 

Guarded LLM-agent workflows with step and tool enforcement.

Re-exports

pub use clients::base as client;

pub use clients::base as streaming;

pub use clients::sampling;

pub use context::hardware;

pub use context::strategies as compact;

pub use core::steps;

pub use core::workflow;

pub use prompts::nudges;

pub use tools::respond;

pub use clients::apply_sampling_defaults;

pub use clients::format_tool;

pub use clients::get_sampling_defaults;

pub use clients::AnthropicClient;

pub use clients::AnyLlmProxyClient;

pub use clients::AnyLlmRuntimeClient;

pub use clients::ApiFormat;

pub use clients::ChunkStream;

pub use clients::ChunkType;

pub use clients::LLMCallInfo;

pub use clients::LLMClient;

pub use clients::LLMRateLimitInfo;

pub use clients::LLMRequestOptions;

pub use clients::LLMResponse;

pub use clients::LLMResponseEnvelope;

pub use clients::LLMUsageDetails;

pub use clients::LlamafileClient;

pub use clients::OllamaClient;

pub use clients::SamplingParams;

pub use clients::StreamChunk;

pub use clients::TextResponse;

pub use clients::TokenUsage;

pub use clients::ToolCall;

pub use clients::MODEL_SAMPLING_DEFAULTS;

pub use context::default_context_warning;

pub use context::detect_hardware;

pub use context::CompactEvent;

pub use context::CompactStrategy;

pub use context::ContextManager;

pub use context::HardwareProfile;

pub use context::MemoryKind;

pub use context::NoCompact;

pub use context::SlidingWindowCompact;

pub use context::TieredCompact;

pub use core::fold_and_serialize;

pub use core::format_tool_call_id;

pub use core::run_inference;

pub use core::FinalResponseScoreFn;

pub use core::InferenceResult;

pub use core::Message;

pub use core::MessageMeta;

pub use core::MessageRole;

pub use core::MessageType;

pub use core::OnChunkFn;

pub use core::OnMessageFn;

pub use core::ParamModel;

pub use core::PrerequisiteCheck;

pub use core::SlotWorker;

pub use core::StepTracker;

pub use core::ToolCallInfo;

pub use core::ToolCallScoreFn;

pub use core::ToolDef;

pub use core::ToolSpec;

pub use core::Workflow;

pub use core::WorkflowRunner;

pub use error::BackendError;

pub use error::BudgetResolutionError;

pub use error::ContextBudgetExceeded;

pub use error::ContextDiscoveryError;

pub use error::ForgeError;

pub use error::HardwareDetectionError;

pub use error::MaxIterationsError;

pub use error::PrerequisiteError;

pub use error::StepEnforcementError;

pub use error::StreamError;

pub use error::ThinkingNotSupportedError;

pub use error::ToolCallError;

pub use error::ToolExecutionError;

pub use error::ToolResolutionError;

pub use error::UnsupportedModelError;

pub use error::WorkflowCancelledError;

pub use guardrails::final_response_top_k_from_logits;

pub use guardrails::recent_errors_from_messages;

pub use guardrails::score_final_response_async;

pub use guardrails::score_tool_call_async;

pub use guardrails::serialize_final_response_state_v1;

pub use guardrails::serialize_state_v1;

pub use guardrails::serialize_state_v2;

pub use guardrails::serialize_state_v3;

pub use guardrails::tool_call_top_k_from_logits;

pub use guardrails::validate_tool_arguments;

pub use guardrails::validate_tool_call_batch;

pub use guardrails::ArgValidationError;

pub use guardrails::ArgValidationKind;

pub use guardrails::ArtifactManifest;

pub use guardrails::CandidateCallForScoring;

pub use guardrails::CheckResult;

pub use guardrails::ClassifierAction;

pub use guardrails::ClassifierArtifact;

pub use guardrails::ClassifierModelKind;

pub use guardrails::ClassifierTopKEntry;

pub use guardrails::ErrorTracker;

pub use guardrails::FinalResponseClass;

pub use guardrails::FinalResponseClassifierArtifact;

pub use guardrails::FinalResponseContext;

pub use guardrails::FinalResponseScore;

pub use guardrails::FinalResponseScorer;

pub use guardrails::FinalResponseToolResult;

pub use guardrails::GuardAction;

pub use guardrails::GuardrailDecision;

pub use guardrails::GuardrailHistory;

pub use guardrails::GuardrailState;

pub use guardrails::GuardrailViolation;

pub use guardrails::Guardrails;

pub use guardrails::LabelThreshold;

pub use guardrails::LabelsFile;

pub use guardrails::NoopFinalResponseScorer;

pub use guardrails::NoopToolCallScorer;

pub use guardrails::Nudge;

pub use guardrails::RetryNudgeFn;

pub use guardrails::ScorerMode;

pub use guardrails::ScoringContext;

pub use guardrails::ScoringExecutor;

pub use guardrails::ScoringMetadata;

pub use guardrails::ScoringPipeline;

pub use guardrails::StepCheck;

pub use guardrails::StepEnforcer;

pub use guardrails::StepPrerequisite;

pub use guardrails::TerminalTool;

pub use guardrails::Thresholds;

pub use guardrails::ToolCallClass;

pub use guardrails::ToolCallScore;

pub use guardrails::ToolCallScorer;

pub use guardrails::ToolSpecForScoring;

pub use guardrails::ValidationResult;

pub use guardrails::WorkflowStateForScoring;

pub use guardrails::DEFAULT_CLASSIFIER_REPO;

pub use guardrails::DEFAULT_CLASSIFIER_REVISION;

pub use guardrails::DEFAULT_FINAL_RESPONSE_CLASSIFIER_REPO;

pub use guardrails::DEFAULT_FINAL_RESPONSE_CLASSIFIER_REVISION;

pub use guardrails::EXPECTED_LABELS;

pub use guardrails::FINAL_RESPONSE_ARTIFACT_SCHEMA_VERSION;

pub use guardrails::FINAL_RESPONSE_EXPECTED_LABELS;

pub use guardrails::FINAL_RESPONSE_INPUT_SCHEMA_VERSION;

pub use guardrails::FINAL_RESPONSE_SERIALIZER;

pub use guardrails::FINAL_RESPONSE_THRESHOLDS_SCHEMA_VERSION;

pub use guardrails::LEGACY_EXPECTED_LABELS;

pub use guardrails::NEXT_INPUT_SCHEMA_VERSION;

pub use guardrails::NEXT_SERIALIZER;

pub use guardrails::V3_SERIALIZER;

pub use prompts::build_tool_prompt;

pub use prompts::classifier_nudge;

pub use prompts::extract_tool_call;

pub use prompts::prerequisite_nudge;

pub use prompts::rescue_tool_call;

pub use prompts::retry_nudge;

pub use prompts::step_nudge;

pub use prompts::unknown_tool_nudge;

pub use prompts::unsafe_batch_nudge;

pub use proxy::extract_passthrough;

pub use proxy::extract_sampling;

pub use proxy::handle_anthropic_messages;

pub use proxy::handle_anthropic_messages_with_scorer;

pub use proxy::handle_anthropic_messages_with_scorers;

pub use proxy::handle_anthropic_messages_with_scorers_and_tool_controls;

pub use proxy::handle_anthropic_messages_with_scorers_and_tool_output_compression;

pub use proxy::handle_anthropic_messages_with_scorers_tool_controls_and_headers;

pub use proxy::handle_chat_completions;

pub use proxy::handle_chat_completions_with_scorer;

pub use proxy::handle_chat_completions_with_scorers;

pub use proxy::handle_chat_completions_with_scorers_and_tool_controls;

pub use proxy::handle_chat_completions_with_scorers_and_tool_output_compression;

pub use proxy::has_respond_tool;

pub use proxy::init_proxy_classifier_log_sink_from_env;

pub use proxy::init_proxy_tool_output_compression_log_sink_from_env;

pub use proxy::init_proxy_training_capture_sink_from_env;

pub use proxy::openai_to_messages;

pub use proxy::respond_tool_openai;

pub use proxy::shutdown_proxy_classifier_log_sink;

pub use proxy::shutdown_proxy_tool_output_compression_log_sink;

pub use proxy::shutdown_proxy_training_capture_sink;

pub use proxy::strip_respond_calls;

pub use proxy::text_response_to_openai;

pub use proxy::text_to_sse_events;

pub use proxy::tool_calls_to_openai;

pub use proxy::tool_calls_to_sse_events;

pub use proxy::AnthropicEventStream;

pub use proxy::AnthropicHandlerError;

pub use proxy::AnthropicHandlerResult;

pub use proxy::HTTPServer;

pub use proxy::HandlerError;

pub use proxy::HandlerResult;

pub use proxy::OpenAiEventStream;

pub use proxy::OpenAiMessageError;

pub use schema_compression::compress_tool_schemas;

pub use schema_compression::patch_anthropic_tool_schemas;

pub use schema_compression::SchemaCompressionMode;

pub use schema_compression::SchemaCompressionStats;

pub use secret_redaction::redact_proxy_request_inputs;

pub use secret_redaction::SecretRedactionError;

pub use secret_redaction::SecretRedactionSummary;

pub use secret_redaction::SECRET_REDACTION_MARKER;

pub use server::setup_backend;

pub use server::BudgetMode;

pub use server::ServerManager;

pub use tool_output::canonical_tool_name;

pub use tool_output::compress_tool_output;

pub use tool_output::detect_family;

pub use tool_output::estimate_tokens;

pub use tool_output::ToolOutputCompressionConfig;

pub use tool_output::ToolOutputCompressionMethod;

pub use tool_output::ToolOutputCompressionMode;

pub use tool_output::ToolOutputCompressionResult;

pub use tool_output::ToolOutputCompressionState;

pub use tool_output::DEFAULT_MAX_DEDUP_ENTRIES_PER_SESSION;

pub use tool_output::DEFAULT_MAX_DEDUP_SESSIONS;

pub use tool_output::DEFAULT_MAX_OUTPUT_BYTES;

pub use tool_policy::evaluate_tool_call_policy;

pub use tool_policy::ToolCallPolicyConfig;

pub use tool_policy::ToolCallPolicyMode;

pub use tool_policy::ToolCallPolicyNudge;

pub use tool_policy::ToolCallPolicyRequestState;

pub use tool_policy::DEFAULT_MAX_WRITE_PAYLOAD_BYTES;

pub use tools::respond_spec;

pub use tools::respond_tool;

pub use tools::RESPOND_TOOL_NAME;

Modules

clients

LLM backend client adapters and traits. Adapters and clients for connecting to various LLM provider backends.

context

Context budget tracking and token compaction strategies. Context tracking and compaction strategies.

core

Core execution core components including message types and tool specifications. Core orchestration engine, workflow definitions, and parsing types.

error

Custom error types for the framework.

guardrails

Response validation and step enforcement guardrails. Guardrails for response validation, error tracking, and step/tool enforcement.

prompts

System and rescue prompt templates. Prompt template system for tool descriptions and tool call extraction.

proxy

HTTP and OpenAI-compatible proxy interface. HTTP proxy interfaces, request handlers, and servers.

schema_compression

Opt-in tool schema description compression for proxy-intercepted requests. Tool schema description compression for proxy-intercepted requests.

secret_redaction

Default-on secret redaction for proxy-bound request inputs. Secret redaction for proxy-bound request inputs.

server

In-process server backend manager. Backend lifecycle manager: process spawning, budget resolution, VRAM tiers.

tool_output

Tool-output compression for proxy-forwarded tool result messages. Tool-output compression for proxy-forwarded tool result messages.

tool_policy

Opt-in tool-call policy nudges for proxy-intercepted tool calls. Opt-in tool-call policy nudges for proxy-intercepted tool calls.

tools

Built-in tools like respond. Built-in tools and executors for the agent runtime.