Channel Binary Wire Protocol

1. Overview & Transport Model

The Channel Inference Engine provides a native, low-overhead binary wire protocol designed for high-throughput, low-latency, full-duplex communication over Standard Input / Standard Output (STDIN / STDOUT) when launched in server mode:

./zig-out/bin/channel --serve [OPTIONS]

Core Design Principles:

• Pure Opcode Protocol (Zero Header Flags): Every event, channel stream, and boundary condition is represented by a discrete opcode. Host dispatch is a simple, direct switch (frame.opcode) with zero bitmask checks.
• Native Multimodal Streaming: Supports text, discrete tokens, continuous soft embeddings, raw/encoded images, audio waveforms, and video frames.
• Stream-Injection Memory Model: Memory queries act as in-engine stream injection side-effects; the host receives lightweight telemetry badges rather than raw vector records.
• Transport: Standard POSIX anonymous pipes (STDIN / STDOUT) with lifecycle bound directly to the parent process.
• Byte Ordering: Little-Endian for all multi-byte integers and IEEE 754 floating-point values.
• Fixed 16-Byte Envelope: Every frame begins with a uniform 16-byte header followed by an optional payload.
• Full-Duplex: Host $\to$ Engine (Inbound) and Engine $\to$ Host (Outbound) frames operate concurrently and asynchronously without turn-taking locks.

2. Binary Message Frame Envelope

Every message transmitted in either direction begins with a fixed 16-byte Header:

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                      Magic (0x53554C58)                       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          Version (1)          |           Msg ID              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|            Opcode             |          Reserved (0)         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                         Payload Length                        |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                        Payload Data...                        |
|                            (...)                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

Field Definitions:

3. Protocol Opcode Catalog

Inbound (Host $\to$ Engine) Opcodes: 0x0001 .. 0x00FF

Outbound (Engine $\to$ Host) Opcodes: 0x0101 .. 0x01FF

4. Detailed Payload Specifications

4.1. OP_STREAM_INPUT (0x0001) — Inbound Multimodal Stream

Ingests text, tokens, soft vectors, audio PCM, images, or video frames into Layer 0.

• Payload Layout:
  • data_type: u8:
    • 0x00: Raw UTF-8 text string
    • 0x01: Token IDs ([]u32 little-endian)
    • 0x02: Soft prompt embeddings ([][hidden_dim]f16)
    • 0x03: Audio PCM samples (16kHz 16-bit mono)
    • 0x04: Encoded Image (JPEG/PNG)
    • 0x05: Raw RGB/RGBA patch tensors
  • flags: u8:
    • 0x01: INPUT_FLAG_DIRECT (bypasses outer chat turn wrappers)
    • 0x02: INPUT_FLAG_REASON (explicitly requests deliberate reasoning)
  • data: [payload_len - 2]u8

4.2. OP_PROBE_AUTONOMIC (0x0012) — Autonomic Reflex Probe

Executes a transient constrained logit probe or generative micro-decode directly on the model's KV attention state, followed by $O(1)$ clock rollback.

• Payload Layout:
  • max_decode_tokens: u16: Number of tokens to decode (1 for discrete logit probe; > 1 for micro-decode).
  • num_candidates: u16: Number of candidate token IDs provided in candidates.
  • candidates: [num_candidates]u32: Array of token IDs to mask logits over.
  • prompt_len: u32: Length of prompt text in bytes.
  • prompt: [prompt_len]u8: Raw prompt text.

4.3. OP_AUTONOMIC_RESULT (0x010C) — Autonomic Probe Result

• Payload Layout:
  • selected_index: u16: 0-based index of the winning token in candidates (or 0xFFFF for generative).
  • winning_token: u32: Token ID selected by sampler / argmax.
  • confidence: f32: Softmax probability of winning candidate ($0.0 \dots 1.0$).
  • entropy: f32: Normalized Shannon entropy over the candidate distribution ($0.0 \dots \ln K$).
  • cost_ms: f32: Elapsed evaluation latency in milliseconds on GPU.
  • text_len: u16: Length in bytes of decoded text (if max_decode_tokens > 1).
  • text: [text_len]u8: Decoded UTF-8 string slice.

4.4. OP_RESUME (0x000A) — Elastic Yield Continuation

Resumes autoregressive generation after an elastic soft yield (STOP_ELASTIC_YIELD).

• Payload Layout:
  • Optional action: u8:
    • 0x00: RESUME_ACTION_CONTINUE (resumes token generation along active trajectory)
    • 0x01: RESUME_ACTION_CLOSE_THOUGHT (caps thinking channel, advances <channel|>, and transitions immediately to response generation)

4.5. OP_TOOL_RETURN (0x0005) — Binary Tool Return

Delivers structured tool execution output back into the engine.

• Payload Layout:
  • call_id_len: u16: Length of tool call ID string.
  • call_id: [call_id_len]u8: Identifier of tool call being answered.
  • is_error: u8: 0 for success, 1 for execution failure.
  • result_len: u32: Length of result text.
  • result: [result_len]u8: Text or JSON output of tool execution.

4.6. OP_SET_CONFIG (0x0004) — Runtime Configuration (41 Bytes)

• Payload Layout:
  • temperature: f32 (default: 0.7)
  • top_p: f32 (default: 0.95)
  • min_p: f32 (default: 0.05)
  • top_k: u32 (default: 64)
  • max_tokens: u32 (safety runaway circuit breaker, default: 4096)
  • repeat_penalty: f32 (default: 1.10)
  • repeat_window: u32 (default: 64)
  • quiescence_threshold: f32 (layer diff gating threshold, default: 0.02)
  • thinking_budget: u32 (maximum thinking tokens, default: 512)
  • flags: u8:
    • Bit 0: enable_critique
    • Bit 1: enable_quiescence
    • Bit 2: enable_memory_injection
    • Bit 3: enable_thinking_gate