Changelog

v0.8.0 — March 23, 2026 (Latest)

Control Plane SDK & Evidence Readiness

NEW llm() is now the preferred SDK client entry point, with openai() retained as a compatibility alias.
NEW SDK docs now present Prysm as one control plane with two first-class paths: proxy and MCP.
NEW Framework examples and guides now cover LangGraph and Agent Framework explicitly, with CrewAI and LlamaIndex surfaced more clearly.
NEW Live golden-path verification now covers proxy, MCP, unified sessions, LangGraph, Agent Framework, CrewAI, and LlamaIndex.
NEW Session evidence can now be copied as a readable summary or exported as JSON from the dashboard.
NEW Trust and security-review docs now describe the current pilot-ready posture more honestly.

Microsoft Agent Framework Integration & Phase 2 Agent Tracing

NEW PrysmAgentMiddleware — captures agent run execution (start/end, timing, messages, metadata).
NEW PrysmFunctionMiddleware — captures function/tool calls (name, arguments, result, timing).
NEW PrysmChatMiddleware — captures LLM chat completions (messages, options, response, tokens).
NEW agent-framework optional dependency: pip install prysmai[agent-framework].
NEW Unified Timeline — correlated view of LLM traces, tool events, and session events.
NEW Tool Performance Dashboard — success rates, latency distributions, failure analysis.
NEW Agent Decision Explainability — trace backward to find why an agent chose a specific action.
NEW Directed Workflow Graph — SVG-based execution graph with pan/zoom and node inspection.
NEW 5 documentation sections in Agent Tracing group.

Governance Layer & LangGraph Integration

NEW GovernanceSession — context manager for monitoring agent sessions with behavioral detection, code scanning, and policy enforcement via MCP.
NEW Early Stopping detector — catches agents that quit before completing their stated task.
NEW Tool Undertriggering detector — identifies when agents have tools available but fail to use them.
NEW Code Security Scanner — scans agent-generated code for injection flaws, hardcoded secrets, unsafe deserialization, and more.
NEW MCP endpoint at /api/mcp — Streamable HTTP transport with JSON-RPC 2.0 for direct agent integration.
NEW PrysmGraphMonitor — LangGraph integration replacing LangChain, with graph-aware telemetry.
NEW governance=True flag on LangGraph and CrewAI integrations for automatic behavioral detection.
NEW Session Explorer, Governance Dashboard, and Policy Manager views in the web dashboard.
BREAKING Removed PrysmCallbackHandler (LangChain). Use PrysmGraphMonitor (LangGraph) instead.

SDK Bug Fixes & Platform Hardening

Fixed PrysmCallbackHandler.on_chain_start crashing with AttributeError when LangChain passes serialized=None.
Added per-request rate limiting with X-RateLimit-* headers.
Fixed PrysmCrewMonitor crashing on malformed delegation tool arguments.
NEW GET /api/v1/models endpoint.
NEW Response headers for security scan results.
NEW PagerDuty alert channel support.

Security Suite & Framework Integrations

NEW Off-topic detection, ML toxicity scoring, NER-based PII detection.
NEW Output policy engine with keyword blocklist, regex, and LLM-judged rules.
NEW Tiered scanning (Basic, Standard, Deep).
NEW Alert channels (Email, Slack, Discord, Webhook).
NEW Framework integrations (CrewAI, LlamaIndex).
NEW Anthropic translation layer.
NEW Confidence analysis and explainability suite.