v1.0.20260622-rc1

ADEPT (Agentic Discovery and Exploration Platform for Tools) v1.0.20260622-rc1

First public release. ADEPT is a production-ready agentic platform that connects
large language models to scientific computing infrastructure through a secure,
three-tier architecture with OpenAI-compatible APIs.

Platform architecture:

• Three-tier security model: Agent Gateway (authentication proxy with Keycloak
  JWT/RBAC), Orchestration Service (LangGraph state machine with PostgreSQL
  persistence), and stateless MCP Tool Servers
• Dual API surface: OpenAI-compatible Response API (primary, streaming) and
  Assistants API (legacy compatibility) — drop-in replacement for existing
  OpenAI client code
• Multi-agent orchestration: Router and Graph execution modes with dynamic
  tool discovery, role personas with purpose-specific LLM routing via YAML
  model catalog, and session-scoped state isolation
• Agent-to-Agent (A2A) mesh: Cross-gateway delegation for federated
  multi-site deployments via Gateway Registry
• Provider-agnostic LLM layer: Azure OpenAI, AWS Bedrock, Ollama, and
  Anthropic backends through LiteLLM with per-purpose model routing

Scientific computing capabilities:

• HPC integration: Agent-driven Slurm job submission with session-scoped
  workspaces, SIF container execution, multi-agent LLM-as-judge validation
• MCP tool ecosystem: 28 built-in tools spanning BLAST, UniProt, PubChem,
  AlphaFold, web search, RAG (ChromaDB + PDF/CSV), code execution sandbox,
  Nextflow pipelines, video transcription
• Secure code execution: nsjail-isolated sandbox with Docker-in-Docker,
  session-scoped file systems
• File management: Upload, registration, presigned URL generation with
  HMAC-SHA256 signing, batch processing with RAG index creation

Developer experience:

• CLI and SDK: Python SDK with device-flow authentication, interactive CLI
  with background process management, A2A peer registration
• IDE integration: VS Code MCP connector, Claude Code subagents,
  29 registered development lifecycle skills
• Deployment automation: 6-phase pipeline, Ansible playbooks, CDK/Pulumi
  IaC, SSM credential injection
• Observability: Langfuse tracing, Fluent Bit log aggregation, structured
  cost tracking per LLM purpose

Security:

• Keycloak RBAC with group-based ACL tool provisioning
• 7-pass pre-release security scanner (secrets, PII, CVE, toxicity)
• Session isolation across all tiers (file, state, execution, network)
• HMAC-SHA256 presigned URLs with constant-time validation

License: BSD-2-Clause