Flow

The human gate is a node, not an interruption.

Agents, CI jobs, approvals and deploys are peers in one graph. Where a person is required, that stop is drawn in the same picture as everything else.

How it answers to the conversation

A run is a thing you can watch, and a thing you can read later.

The graph is the live view. The channel keeps what it did, why it stopped, and who unblocked it.

Nodes are peers
An agent, a test job, an approval and a deploy are the same kind of node.
A molten node is live
One melt edge feeds whatever is running. Nothing else glows.
A gate blocks visibly
The blocked edge is drawn, and the decision card it produced is in the channel with ranked options.
Runs are durable
A run survives a broker restart, a dropped client and a redeploy. It is replayable afterwards.

What it does

Human-in-the-loop as architecture
Not a modal that appears at the end. A gate is declared in the graph, holds the edge, and shows exactly what is waiting on it.
Many agents, one plan
Parallel instances, fan-out and join, and a topology you can read instead of a log you have to reconstruct.
Replay, not archaeology
Every run is recorded with per-step timing and per-command risk classification. "What did it actually do" is a click.

Yours or ours

Crusible provides this one. There is nothing to connect.

The graph is how Crusible runs work — not a view over another tool.

Where it appears

Automation channels are built around it. Worktree and incident channels get one when a run is attached.

Division of labour

Agents run the graph
Provision the environment, run the suite, open the pull request, cut the release candidate. Cruise control, in the literal sense.
Agents route around trouble
A failing node retries under a policy you set, and escalates to a decision card rather than to a silent stop.
A person holds the gates
Production reach, destructive steps and anything outside an agent's permission set. The gate is declared once and enforced every run.

Crusible is opening to teams in private beta.