forge-guardrails 0.1.2

Foundation types for an LLM-agent workflow framework
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141

//! Guarded LLM-agent workflows with step and tool enforcement.

#![warn(missing_docs)]

/// Feature-gated Hugging Face classifier artifact download helpers.
#[cfg(feature = "classifier")]
pub mod classifier_download;
/// LLM backend client adapters and traits.
pub mod clients;
/// Context budget tracking and token compaction strategies.
pub mod context;
/// Core execution core components including message types and tool specifications.
pub mod core;
/// Custom error types for the framework.
pub mod error;
/// Response validation and step enforcement guardrails.
pub mod guardrails;
/// System and rescue prompt templates.
pub mod prompts;
/// HTTP and OpenAI-compatible proxy interface.
pub mod proxy;
/// Opt-in tool schema description compression for proxy-intercepted requests.
pub mod schema_compression;
/// Default-on secret redaction for proxy-bound request inputs.
pub mod secret_redaction;
/// In-process server backend manager.
pub mod server;
/// Tool-output compression for proxy-forwarded tool result messages.
pub mod tool_output;
/// Opt-in tool-call policy nudges for proxy-intercepted tool calls.
pub mod tool_policy;
/// Built-in tools like `respond`.
pub mod tools;

// Legacy module-level re-exports for backwards compatibility and integration tests
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;

#[cfg(feature = "classifier")]
pub use classifier_download::{
    classifier_artifact_needs_download, default_cached_classifier_artifact_dir,
    default_classifier_cache_root, default_tool_call_classifier_artifact_dir,
    download_classifier_artifact_tree, ensure_classifier_artifact_dir, required_files_for_model,
    runtime_required_files, validate_classifier_artifact_dir, ClassifierArtifactKind,
    ClassifierDownloadPlan, ClassifierDownloadReport,
};
pub use clients::{
    apply_sampling_defaults, format_tool, get_sampling_defaults, AnthropicClient,
    AnyLlmProxyClient, AnyLlmRuntimeClient, ApiFormat, ChunkStream, ChunkType, LLMCallInfo,
    LLMClient, LLMRateLimitInfo, LLMRequestOptions, LLMResponse, LLMResponseEnvelope,
    LLMUsageDetails, LlamafileClient, OllamaClient, SamplingParams, StreamChunk, TextResponse,
    TokenUsage, ToolCall, MODEL_SAMPLING_DEFAULTS,
};
pub use context::{
    default_context_warning, detect_hardware, CompactEvent, CompactStrategy, ContextManager,
    HardwareProfile, MemoryKind, NoCompact, SlidingWindowCompact, TieredCompact,
};
pub use core::{
    fold_and_serialize, format_tool_call_id, run_inference, FinalResponseScoreFn, InferenceResult,
    Message, MessageMeta, MessageRole, MessageType, OnChunkFn, OnMessageFn, ParamModel,
    PrerequisiteCheck, SlotWorker, StepTracker, ToolCallInfo, ToolCallScoreFn, ToolDef, ToolSpec,
    Workflow, WorkflowRunner,
};
pub use error::{
    BackendError, BudgetResolutionError, ContextBudgetExceeded, ContextDiscoveryError, ForgeError,
    HardwareDetectionError, MaxIterationsError, PrerequisiteError, StepEnforcementError,
    StreamError, ThinkingNotSupportedError, ToolCallError, ToolExecutionError, ToolResolutionError,
    UnsupportedModelError, WorkflowCancelledError,
};
pub use guardrails::{
    final_response_top_k_from_logits, recent_errors_from_messages, score_final_response_async,
    score_tool_call_async, serialize_final_response_state_v1, serialize_state_v1,
    serialize_state_v2, serialize_state_v3, tool_call_top_k_from_logits, validate_tool_arguments,
    validate_tool_call_batch, ArgValidationError, ArgValidationKind, ArtifactManifest,
    CandidateCallForScoring, CheckResult, ClassifierAction, ClassifierArtifact,
    ClassifierModelKind, ClassifierTopKEntry, ErrorTracker, FinalResponseClass,
    FinalResponseClassifierArtifact, FinalResponseContext, FinalResponseScore, FinalResponseScorer,
    FinalResponseToolResult, GuardAction, GuardrailDecision, GuardrailHistory, GuardrailState,
    GuardrailViolation, Guardrails, LabelThreshold, LabelsFile, NoopFinalResponseScorer,
    NoopToolCallScorer, Nudge, RetryNudgeFn, ScorerMode, ScoringContext, ScoringExecutor,
    ScoringMetadata, ScoringPipeline, StepCheck, StepEnforcer, StepPrerequisite, TerminalTool,
    Thresholds, ToolCallClass, ToolCallScore, ToolCallScorer, ToolSpecForScoring, ValidationResult,
    WorkflowStateForScoring, DEFAULT_CLASSIFIER_REPO, DEFAULT_CLASSIFIER_REVISION,
    DEFAULT_FINAL_RESPONSE_CLASSIFIER_REPO, DEFAULT_FINAL_RESPONSE_CLASSIFIER_REVISION,
    EXPECTED_LABELS, FINAL_RESPONSE_ARTIFACT_SCHEMA_VERSION, FINAL_RESPONSE_EXPECTED_LABELS,
    FINAL_RESPONSE_INPUT_SCHEMA_VERSION, FINAL_RESPONSE_SERIALIZER,
    FINAL_RESPONSE_THRESHOLDS_SCHEMA_VERSION, LEGACY_EXPECTED_LABELS, NEXT_INPUT_SCHEMA_VERSION,
    NEXT_SERIALIZER, V3_SERIALIZER,
};
#[cfg(feature = "classifier")]
pub use guardrails::{
    OnnxFinalResponseScorer, OnnxScorerOptions, OnnxToolCallScorer, MAX_ONNX_SESSION_POOL_SIZE,
};
pub use prompts::{
    build_tool_prompt, classifier_nudge, extract_tool_call, prerequisite_nudge, rescue_tool_call,
    retry_nudge, step_nudge, unknown_tool_nudge, unsafe_batch_nudge,
};
pub use proxy::{
    extract_passthrough, extract_sampling, handle_anthropic_messages,
    handle_anthropic_messages_with_scorer, handle_anthropic_messages_with_scorers,
    handle_anthropic_messages_with_scorers_and_tool_controls,
    handle_anthropic_messages_with_scorers_and_tool_output_compression,
    handle_anthropic_messages_with_scorers_tool_controls_and_headers, handle_chat_completions,
    handle_chat_completions_with_scorer, handle_chat_completions_with_scorers,
    handle_chat_completions_with_scorers_and_tool_controls,
    handle_chat_completions_with_scorers_and_tool_output_compression, has_respond_tool,
    init_proxy_classifier_log_sink_from_env, init_proxy_tool_output_compression_log_sink_from_env,
    init_proxy_training_capture_sink_from_env, openai_to_messages, respond_tool_openai,
    shutdown_proxy_classifier_log_sink, shutdown_proxy_tool_output_compression_log_sink,
    shutdown_proxy_training_capture_sink, strip_respond_calls, text_response_to_openai,
    text_to_sse_events, tool_calls_to_openai, tool_calls_to_sse_events, AnthropicEventStream,
    AnthropicHandlerError, AnthropicHandlerResult, HTTPServer, HandlerError, HandlerResult,
    OpenAiEventStream, OpenAiMessageError,
};
pub use schema_compression::{
    compress_tool_schemas, patch_anthropic_tool_schemas, SchemaCompressionMode,
    SchemaCompressionStats,
};
pub use secret_redaction::{
    redact_proxy_request_inputs, SecretRedactionError, SecretRedactionSummary,
    SECRET_REDACTION_MARKER,
};
pub use server::{setup_backend, BudgetMode, ServerManager};
pub use tool_output::{
    canonical_tool_name, compress_tool_output, detect_family, estimate_tokens,
    ToolOutputCompressionConfig, ToolOutputCompressionMethod, ToolOutputCompressionMode,
    ToolOutputCompressionResult, ToolOutputCompressionState, DEFAULT_MAX_DEDUP_ENTRIES_PER_SESSION,
    DEFAULT_MAX_DEDUP_SESSIONS, DEFAULT_MAX_OUTPUT_BYTES,
};
pub use tool_policy::{
    evaluate_tool_call_policy, ToolCallPolicyConfig, ToolCallPolicyMode, ToolCallPolicyNudge,
    ToolCallPolicyRequestState, DEFAULT_MAX_WRITE_PAYLOAD_BYTES,
};
pub use tools::{respond_spec, respond_tool, RESPOND_TOOL_NAME};