Show HN: OpenClaw Plugin for Claude Code and Codex Orchestration

Posted by goldmar on Tuesday March 10, 2026 @11:02PM
from the show-hn dept.

https://github.com/goldmar/openclaw-code-agent

Doing serious development with vanilla OpenClaw was hit-or-miss. It kept making silly mistakes I had to babysit.

The built-in options didn't work well for me: the CLI is one-shot and ACP lacks plan approval and session persistence. So I built this plugin.

It has three modes: ask (agent proposes a plan, waits for approval before touching anything), delegate (auto-approves safe plans, escalates the risky ones), autonomous (fully hands-off). Paused, completed, and killed sessions can be resumed at any point in time including after a gateway restart. Notifications arrive via chat when it's done or needs input. Supports Claude Code and Codex.

Personally, I really like delegation mode.

Check out the GIFs in the README to see it in action.

( 2 comments | on HN )

Re:Show HN: OpenClaw Plugin for Claude Code and Codex Orches... by alexey-pelykh on Wednesday March 11, 2026 @09:19AM (#47335205)

Most agent orchestration projects skip session persistence across restarts. That's the feature that actually matters for human-in-the-loop workflows, because rebuilding context kills the iteration speed. The ask/delegate/autonomous mode spectrum maps well to how trust builds with these systems.

Genuine question: in delegate mode, when the agent encounters something that needs escalation, does it block until you approve or can it context-switch to other work while waiting? The blocking vs. async decision shapes the whole UX of human-in-the-loop systems.

[ Parent | on HN ]

Re:Show HN: OpenClaw Plugin for Claude Code and Codex Orches... by goldmar on Wednesday March 11, 2026 @10:51AM (#47336376)

For me plan mode and answering questions didn’t work great when I tried using Claude Code with OpenClaw out of the box. So I think that’s the main problem that this plugin solves. And getting proper notifications when something is done or needs a response.

Regarding delegate mode - if something requires attention/clarification, the agent will “block”. That’s just how the mechanism works in the underlying Claude Code and Codex harnesses. There is generally no “let’s continue working on everything else that can be done without answering the questions” mechanism. That would be great, I agree, but probably a feature that should be in Claude Code / Codex themselves rather than this plugin.

[ Parent | on HN ]