Agent integration

The point of cueloop is the gate between what an agent proposes and what it does. This page covers the local agent loop - the review that happens on your own machine, before any sharing.

The plan gate

For Claude Code and Codex, cueloop installs a plan gate. When the agent produces a plan, the gate opens a review and blocks the agent on your verdict. The agent does not start to build until you respond. You read the plan, annotate the exact lines that need work, and send one of three verdicts:

approve           # the agent proceeds as planned
request changes   # the agent revises against your annotations
reject            # the agent stops

The gate delivers your verdict back to the agent in that agent’s own format, so the agent reads it the same way it reads any other tool result. Your annotations become structured feedback the agent can act on line by line, not a wall of prose it has to parse.

Feedback in the agent’s format

Different agents expect different shapes. cueloop translates your review into the format each agent understands, so a request-changes verdict reads naturally to Claude Code and to Codex alike. You annotate once, in one UI, and cueloop does the shaping.

herdr integration

If you run herdr, cueloop reports agent state to it - whether an agent is waiting on a plan review, building, or idle. herdr can then auto-open a cueloop pane when an agent reaches the gate, so a pending review surfaces the moment it exists instead of waiting for you to check.

The two directions, again

This page is about the direction that reaches the agent: your review, your verdict, back to the agent. The other direction - sharing a plan with a teammate over SSH and merging their annotations into yours - is a separate loop. See Sharing over SSH for that. A collaborator’s notes never reach the agent on their own; only your verdict does.