Ecosystem Roadmap¶

This page tracks the dependency versions Kourai Khryseai relies on, why each version is pinned where it is, and what breaking changes are expected in upcoming major releases. It is the canonical reference for upgrade planning.

📌 Version Pin Summary¶

🤝 A2A Protocol¶

Current State (0.3.x)¶

The Agent-to-Agent protocol is under the Linux Foundation's AAIF governance umbrella (since June 2025, 150+ member organisations). The Python SDK is stable at 0.3.25.

Key patterns we rely on:

• A2AStarletteApplication — mounts the A2A server onto Starlette
• AgentExecutor + TaskUpdater — three-phase task lifecycle (working → artifact → complete)
• SSE via message/stream — real-time status updates to the CLI and VN bridge
• /.well-known/agent.json — agent card auto-handled by the application class
• blocking=True on SendMessage — synchronous specialist calls from Hephaestus

What Changes in 1.0¶

1.0.0a0 was published 2026-03-17. It is alpha — the upper-bound pin <1.0 excludes it automatically. Confirmed breaking changes:

The async with client lifecycle pattern (added in 0.3.23) carries forward unchanged.

Migration Checklist (when 1.0 stabilises)¶

• Remove kind field references from all agent_executor.py files
• Rename callback → push_notification_config in any push notification config
• Audit proto-based type imports — replace dataclass-style construction where needed
• Run test suite against 1.0.0 before removing the upper-bound pin
• Update the pin to >=1.0,<2.0

🔌 MCP SDK¶

Current State (1.26.x)¶

The Model Context Protocol is under the Linux Foundation's AAIF governance umbrella (since December 2025). The Python SDK is stable at 1.26.0.

Key patterns we rely on:

• FastMCP from mcp.server.fastmcp — bundled in the SDK, do not install the separate fastmcp PyPI package (that is Prefect's unrelated fork)
• Stdio transport — recommended for local and subprocess MCP connections
• ClientSession — the main client class; manage one session per server manually (no built-in MultiServerSession)

The MCP spec shifted to Streamable HTTP as the primary transport on 2026-03-26, replacing SSE. SSE is deprecated but still functional throughout 1.x.

What Changes in 2.0¶

v2.0 is pre-alpha. The upper-bound pin <2 excludes all pre-releases. The main branch of the SDK repo is v2 development; v1.x is maintenance-only.

Confirmed scope of breaking changes:

Migration Checklist (when 2.0 stabilises)¶

• Replace all FastMCP imports with McpServer across agent and MCP server files
• Migrate any SSE transport configuration to Streamable HTTP
• Update memory-mcp-server.js sidecar transport if applicable
• Re-verify all MCP server healthchecks in docker-compose.yml
• Update the pin to >=2.0,<3.0

🧠 LiteLLM¶

LiteLLM is not upper-bound pinned — it releases near-daily and rarely introduces breaking changes to the interface we use (litellm.acompletion).

Current State (1.82.x)¶

Notable additions in 2026:

• 6.5× faster SDK initialization
• Dynamic Rate Limiter v3
• Built-in guardrails (zero additional cost)
• MCP proxy support
• New models: Gemini 2.5 Flash/Pro, Grok Code Fast, GPT-5, GPT-5.3-Codex, DeepSeek-v3.1

Provider Selection¶

Controlled by KOURAI_PROVIDER and KOURAI_MODEL_TIER in .env:

🎙️ TTS Ecosystem¶

🏛️ Protocol Governance¶

Both A2A and MCP are under the AI Alliance Infrastructure Foundation (AAIF), a Linux Foundation project. This means:

• Stable, vendor-neutral governance
• Long deprecation windows before breaking changes ship
• Public roadmaps and RFCs
• No single company can unilaterally break the protocol

The joint A2A + MCP interoperability specification is expected Q3 2026 and will define how agents that speak both protocols discover and communicate with each other natively.

📚 References¶