LangChain tightens tool gatekeeping, Cline rewind matures, OpenHands surfaces agents

LangChain spent the month closing the gaps that let injected tool arguments and streaming data slip past their guards. The core redaction fix on streaming callback options [1] plugs a credential leak that could surface secrets in callback payloads, while a companion patch respects include_injected=False with filter_args [2]. Tool argument handling locks down subclasses of BaseTool [3], documents reserved argument names [4], and propagates model middleware control flow [5]. Pydantic 2.14 breaks were patched mid-month, and core 1.6.0 shipped with a bundled security warning.

Cline's checkpoint system, the month's quiet centerpiece, now creates reliably and does full workspace rewind on restore, with CLI /undo repaired [6]. Hub draining and replay arrived without dropping work. Desktop went native with background alerts, image generation shipped in v4.1.11, and telemetry bugs surfaced, were fixed, and hook telemetry got patched . But desktop plugins crashed on launch , a regression the team walked back. "render CANCELLED and SKIPPED run statuses" reflects the same reliability push, keeping automations honest [7].

OpenHands released OpenHands 1.10.0, then taught the UI to show live agent activity in chat [8] and routed runtime services through server_info [9]. A git sync UI shipped mid-month, and CI got unblocked. The composer no longer gates on the seeded default profile's LLM ref [10].

Agent infrastructure got safer across the board. Anthropic integration fixed tool blocks, OpenAI 3.0 support arrived in LangChain, and MCP adapter shipped. Cline patched sandbox leaks and multiple CVEs, while LlamaIndex purged unfixable packages. The stack closed out August with agent tool signatures enforced before execution , a fitting cap to a month of gate tightening.