$ the-wire · showcase
Open MCP gateway args fixed, Cline desktop stops losing its backend
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
Two agent-runtime bugs that silently broke working setups got fixed today: LangChain no longer closes off free-form MCP gateway arguments, and Cline's desktop app retries instead of dying when its backend is slow to start.
fix(langchain): preserve open MCP object arguments langchain-ai/langchain
LangChain was converting the free-form object accepted by gateway-style MCP tools (list, describe, call) into an object that accepts no fields, so models had nowhere to put operation-specific keys. The fix marks those direct free-form arguments with additionalProperties: true while preserving constraints supplied by the MCP server; if you run a gateway MCP server without per-operation tools, th...
Desktop v0.0.28 cline/cline
by github-actions[bot]
Cline desktop v0.0.28 addresses the "Desktop backend unavailable" screen: the window used to ask for the backend address exactly once and stop waiting after about 15 seconds, which left slower machines stuck on the error screen and relaunches could lose the same race. Connection attempts now retry and re-resolve the address each time, so a backend that has restarted since is reached.
telemetry: report host-provided core spawn ordinal and reason as metadata (#14104) cline/cline
cline-core now accepts a host-provided spawn ordinal and reason as telemetry metadata, which out-of-process hosts set when they respawn core after a crash, on a rollout fallback, or on a user restart. Previously every spawn reported as user.extension_activated, making a core that crash-restarted three times indistinguishable from a user with three project windows open.
Refine provider model catalog filtering cline/cline
The temporary Cline image-output model restriction is now cross-referenced at its shared helper and both catalog boundaries, so the bundled and runtime filters get removed together when backend support lands. Hub and desktop catalog builders were simplified so each provider's model list is processed once, and the hub webview consumes the server-filtered list directly instead of re-filtering.
docs: retire LlamaHub links run-llama/llama_index
llamahub.ai is being retired and redirected, and llama_index removed every llamahub.ai link from framework docs and the repo README rather than depend on the redirect: generic "browse LlamaHub" links now point at the integrations overview, and package deep links point at the package folder under llama-index-integrations. Quiet day otherwise; the model profile refresh for in-monorepo partner int...