OpenCockpit gives Pi-backed agents a team—and stops children that repeat themselves

OpenCockpit’s Pi-backed sessions can now hand work to child agents instead of doing everything in one thread. Version 1.1.1, released November 3 at 03:48 UTC, adds a subagent tool for sessions using Ollama, DeepSeek, Kimi or GLM. The release describes single tasks, parallel batches of up to eight tasks with four running at a time, and chains that pass one result into the next.

Each child gets its own model, tool list and instructions, without predefined roles. It runs as a real Pi CLI process with its own session file, leaving an inspectable history even when output is stopped or truncated. Only its result returns to the parent. Children cannot delegate again, and cancelling the parent cancels its children.

A narrow brake on repetition

The update also addresses one recognizable failure: a child repeatedly calling the same tools with the same arguments and receiving the same results. Three identical tool batches—or three repetitions of a cycle up to eight batches long—add an execution notice and corrective feedback. If repetition continues, that child stops with no_progress recorded in its history; parallel siblings keep running.

The project explicitly calls this exact-repeat detection, rather than a semantic judge. It does not promise to catch every unproductive loop or make delegated tool use generally safe. The parent keeps its native execution path.

A separate Codex change lets messages sent during an active turn steer that work instead of waiting behind it. Images attached to a steer remain available until the turn exits, according to the release.

OpenCockpit is an MIT-licensed local web interface around existing engines, not a new model. Install with npm install -g @surething/cockpit, then run cockpit; the documented local address is http://localhost:3457. It requires Node.js 22.19 or newer and supports macOS and Linux, with Windows marked experimental. Git enables repository features, and optional Chrome tooling provides browser control.

Users bring the engine’s credentials: configured Claude access, a Codex login, provider API keys or a locally installed Ollama model. The interface’s local operation does not turn hosted-provider model requests into offline inference. AITrending reviewed the English release and tagged documentation without independently reproducing the delegation or loop-stop behavior.