chernistry/bernstein

Highlights

Spawn error classification — The spawner now categorizes failures (rate limit, missing adapter, permission denied, resource exhausted) and uses the category to decide retry strategy: fail-fast for permanent errors, fallback for transient ones. (#594)

EU AI Act compliance engine — New compliance module with risk classification, conformity assessment templates, and evidence export for regulated environments.

WebSocket frontend + API versioning — Live WebSocket updates for dashboards, plus X-Bernstein-API-Version header for forward-compatible API evolution. (#593)

TUI improvements — New MERGING/IDLE agent states with animated spinners, cost sparkline widget, accessibility mode, graceful fallback for terminals without Textual support, configurable keybindings. (#584, #588)

Features

• Agent context inheritance + OS-level resource limits (cgroups/rlimit) for sandboxed agents
• Consensus verifier: N-model majority vote for high-stakes changes
• Semantic diff analysis for detecting behavioral changes beyond textual diffs
• Prompt size pre-check with automatic model fallback chain (#589)
• Per-endpoint rate limiting + SSE disconnect detection
• Sigstore attestation wired into task completion
• Output fingerprinting module for reproducibility tracking
• Adapter health monitoring + output normalization
• Zero-trust agent JWT scope enforcement + automated SBOM generation
• Cluster JWT auth and tenant isolation (#585)
• Batch size tuning and idle backoff multiplier (#592)

Docs & Guides

• Migration guides: CrewAI → Bernstein, LangGraph → Bernstein
• Adapter selection matrix + Redoc API reference (#590)
• Security hardening guide and interactive quickstart tutorial
• Performance tuning and deployment guides
• Architecture Decision Records (ADR-004 through ADR-008)

Fixes

• Missing super().__init__() in 5 adapter subclasses causing AttributeError on spawn
• SonarCloud quality gate: resolved all bugs, vulnerabilities, and CodeQL alerts
• Rate-limit errors now classified as RETRY_FALLBACK instead of hard failure
• Guard missing SeedConfig attrs (model_fallback, agent_resource_limits)
• TUI crash fix (BigStats.alive → agents)
• Spell checker: allow "ehr" (Electronic Health Record) in HIPAA module

CI/Infra

• GitHub App token for auto-release bot identity (proper avatar)
• JUnit test result upload in CI
• Pentest workflow shell redirect grouping

Full changelog: https://github.com/chernistry/bernstein/compare/v1.5.3...v1.5.4