chernistry/bernstein

v1.8.0

Eight feature drops focused on memory, observability, and safety.

Features

• Prompt caching for system prompts and role templates. Repeated role-prompt runs now read from the Anthropic prompt cache, cutting input-token cost on warm paths.
• Structured memory layer. Episodic (per-session event trace) and semantic (long-lived facts) stores, queryable from any adapter.
• Shared memory with actor-aware tagging. Cross-agent writes are tagged with the writing agent so reads can filter/scope without manual bookkeeping.
• Pluggable guardrail pipeline. Policy checks run as ordered pluggy hooks, so custom guardrails drop in without forking the orchestrator.
• Agent-specific operational metrics. Per-adapter dashboards for latency, error rate, and tool-use distribution.
• Compound error rate tracking. Separates "agent errored" from "tool errored within agent" so retries target the right layer.
• Structured SSE event types. TUI/web stream consumers get typed events instead of a JSON grab-bag.
• OpenTelemetry semantic conventions for agent spans. Spans now match the OTel GenAI conventions — plug into existing APM without custom parsing.
• Claude Code Routines integration for scripted agent sessions.

Fixes

• Strengthened path-traversal mitigation in the changelog command (Sonar S2083, BLOCKER).
• Resolved 40+ SonarCloud issues across the codebase.
• .sdd/ state directory is no longer tracked in git.
• Dropped unused imports flagged by ruff.

Dependencies

• Bumped python-multipart in the uv group.

Full changelog: https://github.com/chernistry/bernstein/compare/v1.7.4...v1.8.0