Crate codec_rs 

codec-rs

Rust port of the Codec binary transport protocol — the functional twin of Codec.Net, @codecai/web, and codecai.

Codec carries uint32 token IDs on the wire instead of UTF-8 / JSON, deferring text decoding to the presentation layer. This crate lets a Rust client decode/encode Codec frames, parse tokenizer maps, detokenize/tokenize, watch for tool calls, translate cross-vocab, and load maps over HTTP — all with full sha256 verification.

Quick start

use codec_rs::{TokenizerMap, Detokenizer, DetokenizeOptions};

let map = TokenizerMap::from_json(json).unwrap();
let mut detok = Detokenizer::new(&map);
for frame in frames {
    let opts = DetokenizeOptions { partial: !frame.done, render_special: false };
    let text = detok.render(&frame.ids, opts);
    print!("{text}");
}

See module-level docs and the project README for the full surface.

Re-exports

pub use version_signaling::parse_version_policy_document;

pub use version_signaling::parse_version_required;

pub use version_signaling::well_known_version_policy_url;

pub use version_signaling::CodecVersionPolicyDocument;

pub use version_signaling::CodecVersionRequiredBody;

pub use version_signaling::HttpStatus;

pub use version_signaling::VersionSignalingError;

pub use version_signaling::CODEC_CLIENT_VERSION;

pub use version_signaling::CODEC_CLIENT_VERSION_HEADER;

pub use version_signaling::CODEC_MIN_VERSION_HEADER;

pub use version_signaling::CODEC_REQUIRED_FEATURES_HEADER;

pub use version_signaling::discover_version_policy_blocking;

pub use byte_encoder::decode_byte_level_token;

pub use byte_encoder::encode_byte_level_chars;

pub use byte_encoder::METASPACE;

pub use compression::hash_zstd_dict;

pub use compression::select_zstd_dict_for_response;

pub use compression::CodecZstdDictError;

pub use detokenize::Detokenizer;

pub use detokenize::DetokenizeOptions;

pub use frame::CodecFrame;

pub use frame::IMapCache;

pub use frame::MapCache;

pub use frame::MemoryMapCache;

pub use longest_match::LongestMatchTokenizer;

pub use longest_match::Tokenize;

pub use map::TokenizerMap;

pub use map::TokenizerMapError;

pub use map::ToolCallingArgsFormat;

pub use map::ToolCallingBlock;

pub use map::ToolCallingConvention;

pub use map::ToolCallingMarkers;

pub use map::ToolCallingResultFormat;

pub use map_loader::LoadError;

pub use map_loader::LoadOptions;

pub use map_loader::MapLoader;

pub use map_loader::TokenizerMapHashMismatchError;

pub use safety_policy::Category as SafetyCategory;

pub use safety_policy::CategoryAction;

pub use safety_policy::ClassifierBlock as SafetyClassifierBlock;

pub use safety_policy::ClassifierHost;

pub use safety_policy::ClientHooksBlock as SafetyClientHooksBlock;

pub use safety_policy::EngineFeature;

pub use safety_policy::PublisherBlock as SafetyPublisherBlock;

pub use safety_policy::RulesSummary as SafetyRulesSummary;

pub use safety_policy::SafetyPolicyDescriptor;

pub use safety_policy::SafetyPolicyError;

pub use safety_policy::SafetyPolicyPointer;

pub use safety_policy::POLICY_WELL_KNOWN_BASE;

pub use safety_policy::discover_safety_policy;

pub use safety_policy::load_safety_policy;

pub use stream::decode_msgpack_stream;

pub use stream::decode_protobuf_frame;

pub use stream::decode_protobuf_stream;

pub use stream::MsgpackFrameIter;

pub use stream::ProtobufFrameIter;

pub use stream::StreamError;

pub use pretok_program::run_pretok_program;

pub use pretok_program::PreTokOp;

pub use pretok_program::PreTokProgram;

pub use tokenize::BPETokenizer;

pub use tokenize::ITokenizer;

pub use tool_watcher::ToolWatcher;

pub use tool_watcher::ToolWatcherError;

pub use tool_watcher::WatcherEvent;

pub use tool_watcher::WatcherEventKind;

pub use translator::static_translation_table;

pub use translator::translate_one_shot;

pub use translator::Translator;

Modules

byte_encoder

GPT-2 byte ↔ unicode mapping table and helpers shared by the Detokenizer and BPE encoder.

compression

Client-side helpers for the Codec compression contract.

detokenize

Stateful detokenizer: token IDs → text.

frame

CodecFrame and the pluggable map cache.

longest_match

Vocab-only longest-prefix-match tokenizer.

map

TokenizerMap — the per-model dialect record. Maps are content-addressed (sha256) and immutable.

map_loader

Fetch, verify, and cache tokenizer maps.

pretok_program

Pre-tokenizer program interpreter.

safety_policy

Safety-policy descriptor loading, validation, and discovery.

stream

Stream decoders for the two Codec wire formats.

tokenize

Pure-Rust BPE encoder. Text → token IDs.

tool_watcher

Tool-call / region watcher.

translator

Translator — cross-vocab token-stream pipe.

version_signaling

Codec v0.4 version negotiation — client-side primitives.