$ the-wire · showcase
LangChain core 1.6.3 reconciles gateway tracing, OpenHands scopes agent profiles to MCP servers
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
LangChain shipped langchain-core 1.6.3 with tracing metadata that now reflects the model and provider a gateway actually runs, and OpenHands added per-profile MCP server scoping so an agent can be restricted from servers that make changes.
langchain-core 1.6.3 is out [1]. The release includes a change that lets clientside tracing metadata for model name and provider be overridden from the gateway response [2]. With custom configs, the model and provider the gateway actually runs differ from what the client requested; the new behavior reconciles the trace with what was executed. The same release also refreshes automated model profile data, adding 10 profiles, removing 0, and changing 13 across 4 providers, including `deepseek-flash` at 1,000,000 context and 384,000 output with text+image input, reasoning, and tools [3].
OpenHands shipped v1.18.0 [4] alongside a change that scopes an agent profile to specific MCP servers [5]. Previously the editor never wrote `mcp_server_refs`, so every profile reached every MCP server a user had configured, including servers that can make changes. The profile picker now supports restriction, which the PR ties to a "Code Exploration agent I trust not to edit anything" goal. A separate fix removes the misleading "No budget limit" line from the Token Usage panel, since no way exists for a user to set `max_budget_per_task` [6].
Cline landed two desktop items. The composer provider picker now lists Cline Pass and gains a set-up row [7]; the underlying constraint is that the picker only lists providers with a saved `providers.json` entry, derived from `enabled = Boolean(directSettings)` in `listLocalProviders`. Cloud session foundations also landed, covering authenticated remote Hub connections and recovery of pending session approvals scoped to attached clients [8]. A separate fix corrects OpenCode Go session headers and model API routing: requests now carry a stable conversation ID and use the API protocol declared for the selected model, addressing the reported `Request is missing x-opencode-session and cannot be routed efficiently` failure [9].
LlamaIndex fixed Gemini tool request IDs [10]. `GoogleGenAI.get_tool_calls_from_response()` had discarded the unique per-call ID returned by Gemini and used the tool name as `ToolSelection.tool_id` instead, a behavior present with `llama-index-core==0.14.7`, `llama-index-llms-google-genai==0.8.7`, and `google-genai==1.63.0`. The reporter found it while debugging tool-context updates in a voice AI agent workflow using Gemini LLMs.
Action items
- → Upgrade langchain-core to 1.6.3 if you route through a gateway and rely on trace metadata langchain-ai/langchain [plan]
- → Review OpenHands agent profiles and apply MCP server scoping where agents should not reach mutating servers OpenHands/OpenHands [plan]
- → Upgrade LlamaIndex if you use Gemini tool calling and depend on per-call tool IDs run-llama/llama_index [plan]
- → Update Cline if you use OpenCode Go; session headers and model API routing changed cline/cline [monitor]
References
- [1] langchain-core==1.6.3 ↗ langchain-ai/langchain
- [2] feat(core): Allow model name and provider tracing metadata override based on gateway response ↗ langchain-ai/langchain
- [3] chore(model-profiles): refresh model profile data ↗ langchain-ai/langchain
- [4] v1.18.0 ↗ OpenHands/OpenHands
- [5] feat: scope an agent profile to specific MCP servers ↗ OpenHands/OpenHands
- [6] fix: OHE-3110 : remove misleading "No budget limit" line from Token Usage panel ↗ OpenHands/OpenHands
- [7] feat(desktop): list Cline Pass with Cline and add a set-up row to the composer provider picker (#14058) ↗ cline/cline
- [8] feat(desktop): add cloud session foundations (#13558) ↗ cline/cline
- [9] fix(llms): correct OpenCode Go session headers and model API routing ↗ cline/cline
- [10] fix: gemini tool request ids ↗ run-llama/llama_index