Codec changelog
Customer-facing what's-new feed, newest first
Wire-additive over v0.4 (v0.4 โ v0.5 happy-path bytes identical). Four new opt-in surfaces (delta-varint stream encoding, discoverable Zstandard dictionaries, GPU-side latent quantize, bolt-on tool dispatcher). 11 client artifacts bumped to 0.5.0 across npm, PyPI, NuGet, crates.io, Maven Central. Engine cohort cut to sglang + vLLM + llama.cpp + ComfyUI + diffusers (TGI dropped). 72/72 wire + 72/72 decode unanimous on the cross-stack matrix; numbers byte-identical to v0.4.1, confirming the wire-additive invariant. Upstream PRs filed at sgl-project/sglang#25544 and vllm-project/vllm#42896, both DCO-signed and through bot review.
A patch release that closes three correctness gaps the v0.4.0 cross-stack matrix had been silently papering over. The ยง1 headline conflated protocol efficiency with model-output behaviour, only Python actually decoded dict-zstd, and the bench's unanimity check only inspected wire bytes (not whether anything decoded). All fixed AND defended by regression tests. New synthetic-stream bench is the protocol-only headline; 24/24 wire AND 24/24 decode unanimous across all 6 clients ร 3 engines for the first time.
Codec gains a sixth negotiation axis on the HELLO/READY handshake, a sanitized, hash-anchored `safety_policy` descriptor that lets servers advertise enforcement (categories, actions, classifier family) without leaking operator-internal banned-id lists or thresholds. Adds an optional `@codecai/web-safety` client package (prefilter + classifier registry), full operator-side enforcement in codec-supervisor (logits processor, multi-token matcher, classifier registry with three v1 implementations), and per-language tokenize/detok benchmarks across all six client libs. Wire numbers unchanged from v0.3.x. v0.4 is wire-additive.
Re-ran the full cross-stack matrix after patching two bench-driver bugs (C/TS token-decode fallback, vllm REPS=1 noise). All three engines ร six client languages now produce byte-identical Codec frames per cell, including vllm, which previously read as 0/24 unanimous in the post-mortem.
First end-to-end run with codec-time-leaf in a metamcp namespace surfaced (and we fixed) a CallToolResult validator bug that was rejecting all leaf-mode results. Codec-aware tool calls now compress 4.2ร through the gateway.
The Codec leaf-mode bypass (the architectural target the entire v0.3 contract was designed for) fires end-to-end. `[Codec][leaf]` log line confirms the gateway is a transparent ID pipe; the tokenizer sits at the leaf where it belongs.
First end-to-end run of the v0.3 stack against codec-metamcp:v0.3.0 on a real lab box. tools/list collapses to 3.6ร over JSON-RPC; text streams hit 18ร over JSON-SSE on protobuf framing.
First end-to-end latent run against codec-diffusers with real SD-1.5 latents on the wire. The seven-pipeline registry collapses bytes exactly as the spec promises (int4 packs 3.9ร over raw, ~5-10ร smaller than JPEG).
Image and video diffusion models now stream VAE latents instead of decoded pixels. 48ร smaller wire weight, decode at the leaf.
Codec-aware MCP gateway 2026-05-09
Tool authors can now ship pre-tokenized results that bypass the gateway's back-compat shim. ~4.7ร wire-byte reduction on real MCP traffic.
tool_calling block in tokenizer maps 2026-05-09
Tokenizer maps now carry the model's tool-calling convention. Auto-derived from the chat template, no per-deployment config.
Six client libraries (TypeScript, Python, Java, Rust, .NET, C) are now byte-identical across the cross-stack benchmark matrix. 36 cells ร 3 sizes, all green.
Servers advertise the active zstd dict on the wire; clients fetch it once and decompress every frame against it. Identification by sha256, not URL.