[sie-release-automation]

🤖 I have created a release beep boop

0.11.0 (2026-11-03)

⚠ BREAKING CHANGES

• rust: adopt maintained Candle ZIP 9 backport (#765)

Features

• adapters: add explicit BERT CrossEncoder runtime LoRA (#775) (b35cfc6)
• adapters: add explicit GLiNER2 entity runtime LoRA (#789) (dca37b7)
• adapters: extend explicit runtime LoRA to XLM-R rerankers (#787) (0fd7c9a)
• adapters: make CrossEncoder runtime LoRA explicitly opt in (fc7eeac)
• adapters: support Perplexity decision checkpoints (#791) (0521464)
• api: report locally ready LoRA adapters (c432d4c)
• api: serve declared runtime LoRAs on direct generation routes (a1f82eb)
• api: support System One and OpenAI Decisions clients (9f22374)
• cross-encoder: add declared runtime LoRA lifecycle (3bc2c41)
• examples: repeat the published benchmark numbers from one client (#880) (12376be)
• examples: replay the published fixed-rate performance records from one client (#881) (31ecf76)
• examples: run the OpenCode example on one GPU with Qwen3.8-27B-FP8 (#756) (f886c03)
• examples: run the OpenCode example on self-hosted SIE with an OpenRouter upstream (#735) (1760536)
• examples: train cost-aware routers on BBEH outcomes (#795) (b44c475)
• examples: translate one product string with Hy-MT2 1.8B (#862) (4689199)
• examples: try a local System One model with the TypeSafe SDK (1b11f4c)
• examples: validate the published-benchmarks example against two records (#885) (142a6e6)
• gateway: add OpenAI-compatible decisions endpoint (188333a)
• gateway: count /v1/systemone under the extract metrics label (cea23d9)
• gateway: serve the Anthropic Messages API on /v1/messages (#753) (1f9e0ff)
• gateway: serve the System One-compatible /v1/systemone (562b8e6)
• gliformer: add opt-in exact encoder CUDA graphs (#796) (a206d16)
• helm: add extraInitContainers and extraVolumes to worker pods (#809) (b94c193)
• helm: generate gateway and worker tokens behind default-off values (#750) (2e6350e)
• models: add ibm-granite/granite-speech-5.0-2b for speech-to-text extraction (3724719)
• models: add perplexity-ai/pplx-decider-flash-v1.2 and serve format-2 joint-protocol checkpoints (93efb56)
• models: add VAGOsolutions/SauerkrautLM-MoEdernColBERT-115M (d278019)
• models: add VAGOsolutions/SauerkrautLM-MoEdernColBERT-115M (b5f9c78)
• models: onboard xiaomi-research/Xiaomi-OCR-0 (#875) (96a905c)
• models: report description and release_date in /v1/models (12276f5)
• sdk: add system_one to the Python and TypeScript clients (85b72f8)
• sdk: report credits with up to six decimals in optional string fields (#820) (f27d528)
• server: add a System One-compatible POST /v1/systemone (3fad998)
• server: add a text adapter and catalog entry for embeddinggemma-2 (#874) (78c5314)
• server: add LightOnOCR-3 Markdown output profile (0e3bdea)
• server: add OpenAI Decisions compatibility for text models (5cebd87)
• server: add opt-in bounded batch progress (#806) (63d0a4b)
• server: add opt-in runtime LoRA for native decision models (81ab717)
• server: add opt-in token auth and warn about exposed servers (#751) (2d56a66)
• server: add SIE_TELEMETRY_ENABLED and announce the 0.13.0 opt-in default (#747) (09b62a3)
• server: forward images and model chat-template kwargs to OpenAI-compatible upstream chat (#866) (5383e6c)
• server: let the fake engine answer chat with scripted tool calls, and run the agent-loop check in CI (#739) (055c5bd)
• server: load declared SGLang LoRAs on demand (d5e024a)
• server: release SGLang embeddings and unify Qwen3 defaults (#845) (81dec29)
• server: serve the Anthropic Messages API on sie-server (#815) (b80a029)
• server: support runtime LoRA for native decision models (7f0703b)
• sglang: load declared LoRA adapters on demand (6ae6f6c)

Bug Fixes

• address the follow-up findings from the #811 auth review (#818) (e566b05)
• api: check tool names on every chat route and report an upstream refusal as a 400 (#737) (dadd7a0)
• api: normalize LoRA request selectors across clients and gateways (23204d8)
• api: normalize LoRA request selectors across clients and gateways (5ef20bf)
• api: preserve LoRA selector semantics across wire encodings (81faec6)
• api: refresh model metadata and scope usage checks (24d4d0f)
• api: tell the caller why a remote upstream refused, without naming it (#744) (f52781a)
• benchmarks: account for detached resident LoRA tensors (f06f414)
• benchmarks: bind worker waits to qualification limits (a2932d8)
• benchmarks: bound cross-encoder evidence size (445c5b3)
• benchmarks: fail closed on final cleanup reporting errors (56bd852)
• benchmarks: observe direct cross-encoder forwards (7919cfd)
• benchmarks: preserve shutdown after process inspection errors (c3017d5)
• benchmarks: retain cleanup evidence on inspection errors (9fc2ce5)
• benchmarks: retain lightweight response evidence safely (84e2b18)
• bench: preserve empty LoRA slot evidence (86d0442)
• bench: validate runtime LoRA profile setup (f97830f)
• bundles: bind saved bundle locks to third-party root versions (#886) (fb2fac2)
• bundles: require transformers 5.19 in the transformers5 bundle (6a0d7b6)
• cache: confine cluster downloads to their destination (0c34750)
• cache: confine cluster downloads to their destination (0c34750)
• cache: confine cluster downloads to their destination (47133d2)
• cli: honor the configured Hub fallback policy (17a88bb)
• cli: keep cluster cache URLs out of startup output (10ea83b)
• cli: preserve configured cache and Hub fallback policy (#828) (a52579f)
• cli: preserve environment cache settings (a52579f)
• cli: preserve environment cache settings (e7f0c29)
• count each temporary token link once and correct the repeated-Authorization warning (#811) (ad4af56)
• cross-encoder: detach resident adapters from base execution (ac29fa1)
• ctranslate2: rejoin French elision only for a French target (4488cc0)
• deps: patch ctutils and Tokio security advisories (3415133)
• deps: update Rust security locks (#846) (38ef58d)
• deps: update rustls for October security advisories (#734) (bb8bdfe)
• deps: update worker TLS locks for emSign advisories (d84f169)
• deps: update workspace TLS locks for emSign advisories (e534137)
• examples: itemize router costs and validate Cloud fee labels (a8d14bb)
• examples: let the agent-loop check wait out a busy endpoint (#743) (461c5a5)
• gateway: keep grammar-constrained requests on a named remote profile (#867) (335021a)
• gateway: keep typed decision not-found errors indistinguishable (f124a99)
• gateway: report invalid LoRA selector fields (b4f642e)
• gateway: resolve configured aliases for typed decisions (27eecf0)
• gateway: resolve configured aliases for typed decisions (714d85b)
• gliner2: count compiled inference prompts (#794) (f3b34e2)
• gliner2: support span extraction and entity LoRA on 2.0 (#799) (98ece06)
• helm: reject mismatched trust-manager TLS namespaces (#783) (25bd316)
• helm: require worker auth capability before startup (bcd15f4)
• helm: require worker auth capability before startup (1e8f561)
• helm: separate bundled dependency credential keys (#782) (ac600a5)
• honor offline cache policy for decision checkpoints (4e8d7ba)
• keep chat streams alive through long tool calls (#746) (00d7d73)
• lora: integrate profile-scoped selector corrections (37cba87)
• lora: preserve profile defaults and report selector errors (a08f9af)
• lora: resolve adapters through configured artifact caches (#833) (e251cac)
• lora: resolve adapters through configured caches (e251cac)
• lora: resolve adapters through configured caches (3dde949)
• models: keep list metadata on the System One models only (f19048f)
• models: restore SGLang KV capacity for Mistral embedders and GLM-OCR (#853) (be5501a)
• models: undo Moses entity escaping in MADLAD-400 output (89f5cc5)
• models: undo Moses entity escaping in MADLAD-400 output (bee4ec0)
• narrow nullable CrossEncoder models at load time (6466ffd)
• refresh LoRA recency after worker selection (fc3f4f4)
• rust: adopt maintained Candle ZIP 9 backport (#765) (175d226)
• score: honor profile LoRA selections (4a92dc1)
• score: honor profile LoRA selections (7423f5f)
• sdk: reject incomplete generation streams (118ba91)
• sdk: reject incomplete Python generation streams (8ff968c)
• sdk: reject truncated TypeScript generation streams (f1bdaf9)
• sdk: retain encode timing and call-local trace context (#733) (7eb2f08)
• security: enforce explicit trust for local ST modules (33e137e)
• security: enforce explicit trust for local ST modules (#803) (71b0c54)
• server: address the run-2 findings of the #818 auth review (#821) (9102b11)
• server: admit explicit typed-decision LoRA adapters (788faf7)
• server: admit serving artifacts from the Hugging Face shared blob store (#812) (53ba100)
• server: align BGE Flash profiles with FP16 default (975d530)
• server: align BGE Flash profiles with FP16 default (54f308b)
• server: align mapped BERT weights before Apple CPU inference (#792) (eebce1d)
• server: allow opting out of TF32 at startup (#825) (c296b96)
• server: constrain bundle image installs to uv.lock (#653) (e920e64)
• server: count GLiNER long-word tokens once across windows (8f36b97)
• server: declare the H100 class NV-Embed-v2 needs (#817) (271a5a9)
• server: fit the Qwen3.6-27B h100 NEXTN profile on one H100 (#816) (2af9d8c)
• server: forward independent GTE remote-code revisions (755058c)
• server: forward independent GTE remote-code revisions (8dd574d)
• server: guard OpenCLIP before model construction (0062790)
• server: guard SGLang parent preprocessing (166f506)
• server: guard Transformers 4 model loading (ae3796b)
• server: guard Transformers 4 model loading (ae3796b)
• server: guard Transformers 4 model loading (1c452fd)
• server: keep the visible answer of a thinking model on the MLX chat path (#736) (a5d1076)
• server: load only the LoRA adapters a model declares (#752) (d183404)
• server: load snowflake-arctic-embed-m-v2.0 without xformers (#813) (c03b06c)
• server: log why a remote chat call failed, and warn when an upstream key is unset (#738) (cd21333)
• server: match BGE Flash native embedding addition order (fab8afd)
• server: match BGE Flash native embedding addition order (59bfe78)
• server: meter compiled GLiNER2 document windows (#797) (b68fe36)
• server: normalize partial QKV LoRA adapters for SGLang (922e62a)
• server: parse LoRA request aliases consistently (b103712)
• server: pin OpenSearch GTE model code independently (b350a98)
• server: pin OpenSearch GTE model code independently (af685ce)
• server: pin supported sentence transformers loader (11d3802)
• server: preserve decision error envelopes for malformed outputs (5d8f4bc)
• server: preserve GLiNER word pieces across windows (acbd679)
• server: recheck token links across the scan and warn only for authenticated repeats (#760) (81922aa)
• server: refuse what a chat backend cannot serve in the gateway's words (#823) (fb06db0)
• server: regenerate runtime LoRA model contract (c1af5bf)
• server: register runtime LoRA modules in their bundles (9459c77)
• server: reject unknown numeric LoRA configuration fields (318ef81)
• server: render chat tools as sent on the single-server route (#814) (fabf094)
• server: resolve declared LoRA profile defaults (21ec205)
• server: retain picture text in Docling exports (#754) (39184b5)
• server: roll back a failed LoRA attach and keep adapter sources out of errors (#758) (6d4371c)
• server: send a queued native prompt to a chat-only upstream as one user turn (#871) (b8c557b)
• server: skip the unknown-token id in the ColBERT punctuation skiplist (b078a94)
• server: stop sending the OS user, cwd and pid on /ws/status (#745) (1570745)
• server: support partial QKV LoRA adapters in SGLang (1f686b4)
• server: use cluster cache for decision and entity LoRA (dd20d51)
• server: use full precision for GLiNER PII default profile (b481950)
• server: use full precision for GLiNER PII default profile (33ff77c)
• server: validate GLiNER2 prompts before document work (#793) (2281ba8)
• server: wake idle workers on admitted work (32c82e5)
• server: wake idle workers on admitted work (6dfdaaf)
• sglang: preserve startup-only LoRA capacity (ae6064c)
• sglang: release LoRA GPU slots on pinned engine unload (ebe6fcc)
• sglang: size GPU memory reservations per model (104ebf1)
• sglang: size memory reservations per model (b342dc5)
• sidecar: expose model-group phase durations (53b2943)
• sidecar: expose model-group phase durations (39b6e68)
• sidecar: serve long-form audio in its own lane, with per-model batch cost caps (#628) (422b99f)
• systemone: keep refusal codes and surface unreadable results (b1bc9a5)
• systemone: type-check the score ordering and refusal span code (b9f4f76)
• telemetry: preserve model-loading Prometheus metric name (#784) (04f8080)
• whisper: restore SDPA attention after word-timestamp requests (#711) (44b51dc)

Performance Improvements

• examples: speed up Laya router training (b77cb3e)
• models: raise the Hy-MT2-1.8B KV budget to 32768 tokens (a573616)
• models: raise the Hy-MT2-1.8B KV budget to 65536 tokens (f625df2)
• models: set the Hy-MT2-1.8B KV budget to 32768 tokens (e8fc8c9)
• server: build GLiNER encoders on the target device (#863) (4ddaaba)

---

This PR was generated with Release Please. See documentation.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: bd7a7db1-29bf-4c9b-97d9-c163c8383106

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Tip

🎯 Customize your risk assessments

Guide how CodeRabbit assesses merge risk and security architecture risk. Highlight sensitive changes and explain your security boundaries using Merge Risk Instructions and Security Risk Instructions in your review settings. Security architecture risk assessments require the Advanced plan or above.

✅ Automatic approvals, with your rules

Let CodeRabbit approve pull requests once required threads are resolved, pre-merge checks pass, and your risk limits are met. Enable Automatic approvals in review settings, with overrides for teams or file paths.

CodeRabbit Verify

• Run CodeRabbit Verify to test the changes in this pull request.

[blacksmith-sh]

Found 1 test failure on Blacksmith runners:

Failure

Test	View Logs
test_repository_locks_bind_current_public_inputs	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}