caipe-io/ai-platform-engineering

Description While actively editing a skill or agent file in the UI, the session can unexpectedly refresh — even during active typing — causing all unsaved changes to be lost. The refresh does not appear to be tied to a specific idle timeout; it can happen during active use, not necessarily after a long period of inactivity. Steps to Reproduce Open the UI and start editing a skill or agent file Type actively without saving Observe that the UI may unexpectedly refresh Unsaved changes are lost Expected Behavior The UI should not auto-refresh during active editing sessions, or should at minimum warn the user / prompt to save before discarding unsaved changes. Actual Behavior The UI silently refreshes and all unsaved changes are lost. This can happen during active use, not just after long idle periods. Impact Users are losing work unexpectedly. A workaround of saving very frequently is required, which is not ideal UX. Additional Context Reported via internal service desk (OPENSD-2417

Background PR #1728 removed the legacy standalone A2A agent architecture (supervisor model, BaseLangGraphAgent, BaseLangGraphAgentExecutor, per-agent A2A servers). The netutils agent was one of those standalone agents. This issue tracks the follow-on work: restore netutils as a thin A2A server and add a create_remote_agent_tool factory to dynamic agents so remote A2A agents can be invoked as LangChain tool calls — without reinstating the monolithic base classes. Architecture User → BFF (Next.js) → dynamic-agents (LangGraph + AG-UI/SSE) │ └─ tool call: RemoteAgentTool │ └─ POST / (A2A tasks/send, JSON-RPC) │ netutils-agent (thin FastAPI + A2A SDK)

Summary Today MCP upstream credentials are configured only at the MCP server (mcp_servers.credential_sources in Mongo). Dynamic Agents, the UI test-tool, and the AgentGateway config bridge all consume that single server-level policy. Agents can select which MCP servers/tools they use (allowed_tools), but they cannot supply or override credentials per agent. For autonomous / service-style agents (background runs, scheduled workflows, delegated execution without a human caller JWT), we need a first-class agent credential model so agents can authenticate to MCP upstreams independently of interactive user OAuth connections. Current behavior Layer Credential source MCP server credential_sources: [] Upstream MCP uses its own config (env/PAT); gateway FGA still uses caller JWT MCP server non-empty credential_sources BFF / Dynamic Agents resolve secret_ref, provider_connection, or caller_token and forward via AgentGateway transforms Agent config No credential_sources; cannot overrid

Summary Extend the platform to support multiple agent runtimes beyond the current LangGraph-based implementation. Users should be able to configure and run agents powered by alternative runtimes — such as the Anthropic Claude SDK, AWS Strands, Hermes, or any community agent framework — through a common interface. Motivation The platform currently assumes a single agent runtime (LangGraph + LangMem). As the agent ecosystem grows, teams want to: Bring their own agent framework without forking the platform Compare runtime behavior for the same agent definition Adopt newer or domain-specific runtimes as they mature (Strands for AWS-native workloads, Hermes for open-weight models, Claude SDK for direct Anthropic integration, etc.) Avoid vendor lock-in at the runtime layer Proposed Design Runtime Abstraction Layer Introduce a AgentRuntime interface (Python) that all runtimes implement: class AgentRuntime(Protocol): async def run(self, messages: list[Message], config: RuntimeConfig)

Summary Add a persistent agent workspace UI that allows users to upload, view, and edit files — including Memory.md — directly within a dynamic agent's configuration. This gives agents persistent context that survives across conversations, complementing the existing ephemeral file system used during conversation runs. Motivation Currently the platform supports: ✅ Ephemeral files created during a conversation (auto-deleted after run) ✅ Multi-file editing in the Skill Workspace (SKILL.md + ancillary files) ✅ Backend fact extraction via LangMem (not surfaced in UI) What's missing is a way for users to pre-load persistent files into an agent's workspace — most importantly a Memory.md that the agent can read as long-term context, but also arbitrary reference files, runbooks, or configuration snippets. Proposed Changes 1. Agent Workspace Files Tab (DynamicAgentEditor.tsx) Add a new "Workspace Files" step/tab to the existing agent editor wizard: File tree listing persisted work