Release 0.11.0

v0.11.0 — Context, learning & observability

Released September 29, 2026. The unshipped 0.10.0 milestone is included here; the previous published version is 0.9.0. Existing release tags remain unchanged.

Highlights

• Selectable Hermes ContextEngine: full available history, exact pinned sources, bounded JEV attention, rediscovery and recovery, with no generative history summarizer or Hermes core patch. Middleware mode remains.
• Purpose-first scope safety: embeddings, reranking, helpers/classifiers, tokenizer work and internal JEV bypass reduction; retries stay isolated.
• Learned omission-risk policy: System-2 question revision, batched JEV features and real CatBoost fitting outside JEV's weights. Off by default; operator-approved numeric JSON may veto omissions, never relax guards.
• Observatory: localhost:7474 profile/bot dashboard, separate input/output accounting, configured API-equivalent values and a native Desktop bottom status-bar counter/upward popover.
• Updated architecture and learning-loop schematics, release installation, migration, security and documentation. The original Terminator artwork is unchanged.

Python and Rust package versions are 0.11.0. Rust remains the interoperability companion, not a port of the Python runtime or learner. Installation does not select a ContextEngine, start a server, train a model or enable a policy.

Complete release notes · GitHub release · Quick Start

Evaluation boundaries

The implementation baseline passed 611 tests with two optional integration skips; final release-commit checks are visible in Actions. Synthetic training demonstrates fitting, revision and exact-evidence retention, not live model quality or monetary savings. No production policy is shipped. Output savings and subscription discounts are not inferred. Protected-history overflow still needs host/provider enforcement.