Cortex v1.4.0

1.4.0 - 2026-04-26

Production-Readiness Gate

• Made Docker the authoritative release gate through a pinned Rust 1.95.0 image, so local host toolchain drift no longer defines whether a release is valid.
• Added a single release command that runs warning-suppression scanning, formatting, docs drift checks, package-surface checks, secret/path scanning, strict clippy, the full workspace test suite, and doctests.
• Enforced the zero-warning policy as a release contract: cargo fmt must have no diff, clippy runs with -D warnings -W clippy::pedantic -W clippy::nursery, and Rust warning suppression attributes are rejected instead of tolerated.
• Added release-clean verification support so a tagged release can be validated from a committed tree instead of relying on a dirty workspace.
• Updated operator and testing documentation so the strict Docker gate is the documented release authority, while host and Docker Compose commands remain developer shortcuts.

Retrieval and RAG Evidence

• Introduced a dedicated retrieval evidence plane separate from durable memory. Documents are chunked, indexed, retrieved, reranked, compressed, cited, and promoted as evidence rather than silently becoming recalled memory.
• Added hybrid sparse+dense retrieval with deterministic BM25-style lexical scoring, pluggable dense encoders, configurable paraphrase handling, score normalization, and configurable reranking limits.
• Added extension hooks for learned sparse expansion and late-interaction reranking, preserving scope checks and baseline sparse retrieval even when those hooks are enabled.
• Added actor and access-class filtering to retrieval, so private evidence stays bound to the requesting actor and public evidence is the only cross-actor default.
• Added taint-aware evidence modeling. Retrieved text is treated as untrusted or tainted evidence when appropriate, including retrieved instructions that look like prompt-injection attempts.
• Added citation, source-title, corpus, chunk, span, license, index-version, and score metadata to evidence so the runtime can explain where retrieved material came from.
• Added query transforms, including hypothetical-document style expansion, as query aids only. Transforms are explicitly not evidence and cannot be promoted as source material.
• Added retrieval-quality evaluation metrics and support decisions so low-support results trigger rerank/seek-more behavior instead of being treated as sufficient grounding.
• Added journal event payloads for retrieval decisions, retrieved evidence, and promoted evidence so RAG behavior becomes observable and replayable at the runtime surface.
• Added bilingual retrieval documentation describing the RAG pipeline, evidence safety model, implemented surface, and current limits.

LLM Context Assembly

• Added a dedicated retrieved-evidence prompt layer between situational context and recalled memory.
• Rendered evidence with citation, source, corpus/chunk, span, access class, taint, license, index version, and score metadata instead of flattening it into ordinary conversation history.
• Marked retrieved evidence as inert context for the model: it may support answers, but it is not an instruction source and cannot override system, developer, user, or runtime policy.
• Preserved existing memory semantics by passing retrieval evidence independently from recalled memory.
• Added runtime-facing coverage that verifies retrieved evidence enters the assembled prompt before memory and retains citation/license metadata.

Workspace and Control Contracts

• Added typed workspace frames and items with actor ownership, taint, kind, salience, token budget, and item-count budget validation.
• Added control-decision contracts for continue/wait/delegate/interrupt/complete behavior, including expected value, signal aggregation, conflicts, impasses, and subgoals.
• Added journal event payloads for workspace frame assembly, workspace item promotion, control decisions, and impasse recording.
• Extended public type exports so runtime, turn orchestration, and future plugins can share the same workspace/control/retrieval contract shapes.
• Updated replay and journal type mapping so the new 1.4 runtime events remain visible in payload-type projections.

Documentation and Release Surface

• Bumped the workspace, SDK, plugin compatibility examples, scaffolded plugin templates, installer examples, and public docs to the 1.4.0 release line.
• Updated the README and README.zh architecture descriptions to include the retrieval evidence pipeline, separate evidence context, and expanded event surface.
• Updated executive docs so LLM request assembly explicitly includes retrieved evidence before recalled memory.
• Updated maturity notes to distinguish retrieved evidence from memory and keep the project framed as a local-first runtime with explicit hardening limits.
• Reworked the roadmap into one 1.4.0 production-readiness line, avoiding parallel future-version tracks and keeping remaining work scoped as internal workstreams.
• Added docs/runtime contract checks so retrieval docs, roadmap wording, strict-gate documentation, event counts, plugin version examples, and package metadata stay aligned with the shipped runtime.

Runtime Quality

• Cleaned the persistence command shape to satisfy strict clippy and reduce oversized enum payloads without adding warning suppressions.
• Kept existing actor/session/channel ownership behavior intact while adding retrieval and context-assembly surfaces.
• Preserved installer asset naming and latest-release resolution so install and update continue to fetch the newest published binary by default.

Validation

• Verified the release tree with:
  • ./scripts/gate.sh --docker