chore: release main

.release-please-manifest.json

@@ -1,6 +1,6 @@
 {
-  "crates/higgs-models": "1.8.0",
-  "crates/higgs-engine": "1.8.0",
-  "crates/higgs": "1.8.0",
-  "crates/higgs-bench": "1.0.2"
+  "crates/higgs-models": "2.0.0",
+  "crates/higgs-engine": "2.0.0",
+  "crates/higgs": "2.0.0",
+  "crates/higgs-bench": "1.0.3"
 }

crates/higgs-bench/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "higgs-bench"
-version = "1.0.2"
+version = "1.0.3"
 edition = { workspace = true }
 license = { workspace = true }
 rust-version = { workspace = true }

crates/higgs-engine/CHANGELOG.md

@@ -1,5 +1,81 @@
 # Changelog
 
+## [2.0.0](https://github.com/dusterbloom/higgs/compare/higgs-engine-v1.8.0...higgs-engine-v2.0.0) (2026-10-19)
+
+
+### ⚠ BREAKING CHANGES
+
+* higgs start is now config/profile-only, attach is a strict daemon dashboard, shellenv/exec fail fast on invalid or unreachable targets, and exact local model matches now take precedence over regex routes.
+
+### Features
+
+* add qwen3.5/qwen3.6 turboquant stack ([4f165ee](https://github.com/dusterbloom/higgs/commit/4f165eeca1166ac1d1b62cbaf523b83d344b9a22))
+* **bench:** context-frontier benchmark with shared cache rollback helper (PR2 / ds4 P8) ([#253](https://github.com/dusterbloom/higgs/issues/253)) ([4510835](https://github.com/dusterbloom/higgs/commit/451083523816e33cf636d796e240c6bb21a0a62d))
+* **bonsai-q1:** packed engine scaffold with upstream MLX guard ([#142](https://github.com/dusterbloom/higgs/issues/142)) ([fe43aab](https://github.com/dusterbloom/higgs/commit/fe43aabe44104ab285e19e8ff73c724c0875cbe0))
+* **bonsai:** run Bonsai-Q1 bits=1 on vanilla MLX via JIT qmv_fast kernels (no mlx-rs fork) ([#182](https://github.com/dusterbloom/higgs/issues/182)) ([b604754](https://github.com/dusterbloom/higgs/commit/b604754cab79f3f91a95d863324db46e8765e14a))
+* **cache:** enforce retained-byte continuation ([2057cc7](https://github.com/dusterbloom/higgs/commit/2057cc74145615953a4756f799be7b87fdb80c98))
+* **chat:** log raw model text when the streaming tool-call parser rejects output ([7e9cf43](https://github.com/dusterbloom/higgs/commit/7e9cf43a463c66b5a4ecd091253781368fbb53df))
+* **chat:** streaming tool-call deltas + Qwen-friendly arg normalisation ([#164](https://github.com/dusterbloom/higgs/issues/164)) ([8f01163](https://github.com/dusterbloom/higgs/commit/8f01163a0a60950c011048f8bd3c6d906b3e8a01))
+* **chat:** warn and count large stateless prefills and disabled session routing ([0203e29](https://github.com/dusterbloom/higgs/commit/0203e29c3e594b9bf92731a31c4e35dcc9fb485e))
+* enable MiniCPM5-1B (explicit head_dim + special-token decode + tojson kwarg + tool parser) ([#176](https://github.com/dusterbloom/higgs/issues/176)) ([05464e2](https://github.com/dusterbloom/higgs/commit/05464e22fd67a713ccf8a42d6547fab260d401cb))
+* **engine:** chunked prefill yield quantum for the batch serving loop (PR8 / ds4 P9) ([#259](https://github.com/dusterbloom/higgs/issues/259)) ([247147e](https://github.com/dusterbloom/higgs/commit/247147eacfa04ad97af6d1b5aca974594e8da736))
+* **engine:** disk-backed KV prefix store for restart-resume (PR6 / ds4 P4) ([#257](https://github.com/dusterbloom/higgs/issues/257)) ([0435d11](https://github.com/dusterbloom/higgs/commit/0435d112ff8abf3de65ed554a95e3e3a9b69d690))
+* harden CLI, dashboard, routing, and MLX runtime ([4dfc930](https://github.com/dusterbloom/higgs/commit/4dfc930365ec1d8eb8143508fe63c41b21001ba1))
+* harden default security posture ([1b41a5c](https://github.com/dusterbloom/higgs/commit/1b41a5c93696a7492ef4bf7e941982341edf50ab))
+* inference engine — chunked prefill, prefix cache, thinking budget, MTP ([4fa0971](https://github.com/dusterbloom/higgs/commit/4fa09711096dc40cff8b2b33b6c5ad1b87e4b128))
+* **models:** add Gemma 3 and Gemma 4 inference support ([4f301fc](https://github.com/dusterbloom/higgs/commit/4f301fcc0d45e2922df4203149f05a24a9f46ae6))
+* **models:** add LiquidAI LFM2-MoE (lfm2_moe) ([3ea3fa1](https://github.com/dusterbloom/higgs/commit/3ea3fa12b30813215a5324723b201c55d947b37e))
+* **models:** load MiniCPM-V 4.7 (minicpmv4_7) via the qwen3_5 loader ([695f220](https://github.com/dusterbloom/higgs/commit/695f22057029e51fb8b11d75132395ee85a7b0c5))
+* **models:** MLA latent KV cache with weight absorption for DeepSeek-V2 (PR3 / ds4 P1) ([#254](https://github.com/dusterbloom/higgs/issues/254)) ([482d2ec](https://github.com/dusterbloom/higgs/commit/482d2eceab6d616f9e6e79c37d6af06624c42410))
+* **mtp:** release speculative decoding optimizations ([0e5e458](https://github.com/dusterbloom/higgs/commit/0e5e458ce2b59ab5f1a6aaaa1fe91af6f298fe39))
+* **mtp:** release speculative decoding optimizations ([1d15599](https://github.com/dusterbloom/higgs/commit/1d15599ea86fccb022d786a56b2904131078b150))
+* Qwen3.5 model architecture, TurboQuant KV cache, and model-level optimizations ([1514737](https://github.com/dusterbloom/higgs/commit/1514737d4e1fb108c68b3150e9104aae5b20f400))
+* **qwen35:** Qwen3.6 MoE MTP speculative decode + checkpoint memory-safety fix ([#183](https://github.com/dusterbloom/higgs/issues/183)) ([4989955](https://github.com/dusterbloom/higgs/commit/49899557c5e4fef561c6d79ab957ad0656f486ab))
+* **serve:** stream prefill progress (llama.cpp-compatible prompt_progress) ([#184](https://github.com/dusterbloom/higgs/issues/184)) ([e7b2c8a](https://github.com/dusterbloom/higgs/commit/e7b2c8a41689c7e8e1f1fc4d657a4fa17bba2a96))
+* thinking mode support — reasoning parser and chat template ([a3cc52e](https://github.com/dusterbloom/higgs/commit/a3cc52e7e1f3ef12118b011156cc7b7bad6c05fb))
+* tune qwen3.6 thinking defaults ([6969834](https://github.com/dusterbloom/higgs/commit/696983431b7c404bf5efea4978398251603ac6ec))
+
+
+### Bug Fixes
+
+* address coderabbit issues ([6f10bc6](https://github.com/dusterbloom/higgs/commit/6f10bc6d733c192d30c5ab3aed76e4a2e2a69818))
+* address follow-up review issues ([5c857ce](https://github.com/dusterbloom/higgs/commit/5c857ce32f5eeab4fe020c9760a2bd5a0e984c05))
+* address remaining draft review issues ([4396e88](https://github.com/dusterbloom/higgs/commit/4396e8840e399af4abb60c84dfc4412ba8b2131f))
+* address review items across models and engine — PR [#74](https://github.com/dusterbloom/higgs/issues/74) ([8863e2d](https://github.com/dusterbloom/higgs/commit/8863e2d11dacbdada6e560556813c051bf265d8b))
+* **api:** leave required-continuation prefix proof to the worker ([74dfb05](https://github.com/dusterbloom/higgs/commit/74dfb0515db9ff7138fc64cf91f9cb98eee4fec2))
+* **cache:** acknowledge required dspark streams ([7c3f782](https://github.com/dusterbloom/higgs/commit/7c3f782d216a64201a1ddda43ca97d42a9e9bc1d))
+* **cache:** bind and reserve exact retention ([347e8b9](https://github.com/dusterbloom/higgs/commit/347e8b975031ea8de9851f464757055c9aaa1853))
+* **cache:** bind retained reservations to owners ([be3598b](https://github.com/dusterbloom/higgs/commit/be3598bbbe6f791d2ebe716cb6e84151c25ec569))
+* **cache:** bound retained DFlash tap tails ([579c501](https://github.com/dusterbloom/higgs/commit/579c501a3f0848dad16bc7d2545bc589614bd48b))
+* **cache:** keep live retention claims owned ([1bbe259](https://github.com/dusterbloom/higgs/commit/1bbe2591babfad8b259f030c0363ace023014d0b))
+* **cache:** make retained cap updates atomic ([004c881](https://github.com/dusterbloom/higgs/commit/004c881a414edb286d7bad34ab0503ac6c3f11be))
+* **cache:** reserve retained seed slots ([d45f1b8](https://github.com/dusterbloom/higgs/commit/d45f1b822c2b012d39004df9a6034b022aeabded))
+* cargo fmt + restore higgs crate build on PR [#74](https://github.com/dusterbloom/higgs/issues/74) ([050af95](https://github.com/dusterbloom/higgs/commit/050af9590968c129dc8f0a5003d4af47f6591da5))
+* **chat:** evict idle retained sessions before seed admission ([b8adfd4](https://github.com/dusterbloom/higgs/commit/b8adfd423d560b7653ebf7725454662e956a3637))
+* **chat:** render prompts like HF for reasoning + tool models ([32c06c8](https://github.com/dusterbloom/higgs/commit/32c06c827f95407f980c4f9db50b9bcfbefb0da0))
+* clear lint and review blockers ([9c8b878](https://github.com/dusterbloom/higgs/commit/9c8b878bb654ba8865f3455c91ba457b83d1161b))
+* **deps:** migrate to sha2 0.11 ([#273](https://github.com/dusterbloom/higgs/issues/273)) ([eb817f9](https://github.com/dusterbloom/higgs/commit/eb817f97f6f5fd27c6f6d1f1b99c358fe53c6724))
+* doc_markdown lints in detok test comments ([deefbe3](https://github.com/dusterbloom/higgs/commit/deefbe39345669e101b9f1fc5caa90de0aaaa973))
+* **engine,models:** cache config propagation, ndim guards, FSM ordering ([37d2ddd](https://github.com/dusterbloom/higgs/commit/37d2ddd3c222614373f9ee3c49b43e5e69e35c62))
+* **engine,models:** clippy clean lib targets on PR [#74](https://github.com/dusterbloom/higgs/issues/74) (51→0 errors) ([ebf3192](https://github.com/dusterbloom/higgs/commit/ebf319238ed9a79c8efc18bba6b0fbfc6b98f3e3))
+* **engine:** clippy clean 6 files on PR [#74](https://github.com/dusterbloom/higgs/issues/74) (85→51 errors) ([7dc7370](https://github.com/dusterbloom/higgs/commit/7dc737052edea2fdf96c4ab7a097051c0ba34562))
+* **engine:** don't cap reasoning on templates that force <think> ([bb22480](https://github.com/dusterbloom/higgs/commit/bb2248020a80d34411d91a7cf02e17fd1dc55c07))
+* **engine:** explain required-continuation misses ([8c3fea7](https://github.com/dusterbloom/higgs/commit/8c3fea705955c03f76ddd4b1d678c4a6fae195bb))
+* **engine:** keep quoted chat markers out of the structure channel ([f587928](https://github.com/dusterbloom/higgs/commit/f5879288254a7f8615ed59889ced53ee5dd236e9))
+* **engine:** prefix cache full-hit, scheduler leak, gather bounds check ([3e35b3f](https://github.com/dusterbloom/higgs/commit/3e35b3f8e3fe83604883418e408bae09e3533c39))
+* make pre-push pass on dust stack ([2695165](https://github.com/dusterbloom/higgs/commit/26951659aa8ff56d0362a6916827dbe9cf6fae88))
+* **mlx:** restore 0.31 runtime compatibility ([e36ba68](https://github.com/dusterbloom/higgs/commit/e36ba68b8b824b62d06ec7e2e9885231ab309c1f))
+* rename first-token detok bindings to avoid shadowing ([8d16696](https://github.com/dusterbloom/higgs/commit/8d16696c2d59750d3d742f203ab7316a4785b2be))
+* route first streamed token through incremental detokenizer ([722218d](https://github.com/dusterbloom/higgs/commit/722218df3cfb10c8671721cf45888ec45e31187c))
+* split long first doc paragraphs for clippy ([c195312](https://github.com/dusterbloom/higgs/commit/c1953125d8399f899236d7917f26b94051fccb4b))
+* stabilize dust stack CI ([8e629e3](https://github.com/dusterbloom/higgs/commit/8e629e35e53bb65b9a6dd718be5d2d50a79e297c))
+
+
+### Performance Improvements
+
+* **escha:** async layer eval off, 3 retained sessions ([fef5775](https://github.com/dusterbloom/higgs/commit/fef5775c75bd2908584ab941377317a90614b3d1))
+* incremental detokenization for streaming generation ([a949e50](https://github.com/dusterbloom/higgs/commit/a949e50fb3a5c2d650c2d7de367eb4afea444138))
+
 ## [1.8.0](https://github.com/panbanda/higgs/compare/higgs-engine-v1.7.0...higgs-engine-v1.8.0) (2026-09-02)
 
 

crates/higgs-engine/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "higgs-engine"
-version = "1.8.0"
+version = "2.0.0"
 description = "Inference engine (tokenization, generation loop, prompt caching) for higgs"
 edition.workspace = true
 license.workspace = true