aphrodite 💋 Core Engine

CCR compression proxy + cdylib - classify, compress, store, preview. Sub-ms, 26 content types, 12,800× max ratio.

The core compression engine. Produces both the aphrodite binary (HTTP proxy on :9797/:9798) and libaphrodite.dylib (best-effort cdylib for external embedding; the Hermes plugin loads the bridge crate's libaphrodite_hermes.dylib instead - see crates/aphrodite-hermes/README.md).

crates.io · docs

Install

# From source (monorepo)
cargo build --release -p aphrodite
# then bootstrap the runtime home (~/.hermes/aphrodite/: binaries/,
# directives/, hotreload/, ccr.db)
aphrodite setup

# From crates.io
cargo install aphrodite

What it does

tool output → classify → template → store → <<<CCR:hash|type|size>>>
                                              │
                                              ▼
                                         Agent (honest preview, not raw output)

Four pipeline stages, all under 1ms:

1. Classify - 26-type regex classifier (diff, build_output, code_rust, …)
2. Template - TOML-driven preview templates per content type
3. Store - blake3 hash → SQLite or in-memory → CCR marker
4. Preview - Structured [type:metadata] the LLM reads instantly

Architecture

src/
├── lib.rs               ← 25 C ABI exports for dylib loading
├── proxy.rs             ← HTTP proxy (:9797 cache, :9798 token)
├── hooks.rs             ← transform_tool_result, terminal, pre/post LLM
├── session.rs           ← Turn lifecycle, conversation index, catalog
├── state.rs             ← AphroditeState, inline store, LRU, markers
├── marker.rs            ← CCR marker formatting + parsing (<<<CCR:…>>>)
├── catalog.rs           ← Full/compact/TOC catalog display
├── chain_split.rs       ← Fine-grained chain splitting (SEG_MARKER segments)
├── resolve.rs           ← Recursive CCR expansion (5 levels deep)
├── stage2.rs            ← Semantic reduction for JSON, build, diff, code
├── struct_extract.rs    ← Code structure maps (Rust, Python, Go, JS/TS)
├── config.rs            ← CLI args + TOML multi-config
├── config_loader.rs     ← TOML + env var loading
├── directives.rs        ← Conversational directives: load, inject, list/swap/add/remove/reset
├── flow.rs              ← Per-turn injected-context assembler (directives + nudges + recall catalog)
├── setup.rs             ← `aphrodite setup` one-shot bootstrap (Gatekeeper-safe artifact install)
├── prefetch.rs          ← Background file read + compress threads
├── poll_worker.rs       ← Auto-backgrounding of slow tool calls
├── preview.rs           ← detect_type, build_preview (shared across proxy + dylib)
├── main.rs              ← Binary entry point
└── retrieve.rs          ← POST /retrieve - hash → content, filter, paginate, truncated flag

C ABI (25 functions)

The 25 #[no_mangle] extern "C" exports in lib.rs (22 written out plus 3 macro-generated: aphrodite_compress, aphrodite_transform, aphrodite_terminal) form the core C ABI. The aphrodite-hermes bridge crate links this crate as an rlib and exposes its own higher-level ABI (libaphrodite_hermes.dylib - see crates/aphrodite-hermes/README.md); the Python plugin's ctypes bindings (_bindings.py) are generated from the bridge header, not from these exports.

Dependencies

• headroom-core - Content transforms + classifier + SQLite backend (vendored fork at vendor/headroom/)
• axum / tokio / tower-http - HTTP proxy (optional, gated behind proxy feature)
• serde / serde_json - Serialization
• blake3 - Content-addressed hashing

See Also

• Installing Aphrodite - which artifact you need, per-platform install guides, troubleshooting
• aphrodite.toml Configuration - full TOML schema this crate's config.rs deserializes
• Hermes Integration - how this binary's sibling dylib crate (aphrodite-hermes) plugs into Hermes Agent

License

CC0-1.0 - public domain.